[PDF]

Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025) PDF Free Download

1 / 19
1 views•19 pages

Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025) PDF Free Download

Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025) PDF free Download. Think more deeply and widely.

C
e
r
eb
r
a
s
vs
S
a
m
ba
N
ov
a
vs
G
roq
:
AI C
hi
p
C
omp
a
r
i
son
(2025)
B
y
A
d
r
ie
n
L
a
ur
e
nt
,
CEO
a
t
I
ntu
i
t
i
on
L
ab
s
•
10/23/2025
•
40
m
i
n
r
ead
ce
r
eb
r
a
s s
a
m
ba
nov
a g
roq
ai
ha
r
d
w
a
r
e ai
acce
l
e
r
a
tors w
afe
r
-
s
ca
l
e
e
n
gi
n
e
l
a
n
g
u
age
pro
ce
ss
i
n
g
un
i
t
ai
i
n
fe
r
e
n
ce ai
chi
p
c
omp
a
r
i
son
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 1 of 19
E
x
ec
ut
i
v
e
S
umm
a
ry
T
hi
s
r
e
port
prov
ide
s
a
de
t
ai
l
ed
c
omp
a
r
a
t
i
v
e
a
n
a
lys
i
s
o
f
t
h
r
ee
e
m
e
r
gi
n
g
AI
ha
r
d
w
a
r
e
c
omp
a
n
ie
s
–
C
e
r
eb
r
a
s
S
yst
e
ms
,
S
a
m
ba
N
ov
a
S
yst
e
ms
,
a
n
d
G
roq
–
a
s
o
f
O
c
to
be
r
2025.
T
he
s
e
c
omp
a
n
ie
s
each
o
ffe
r
sp
ecia
l
i
z
ed
AI
acce
l
e
r
a
tors
t
a
r
ge
t
i
n
g
high
-
p
e
r
f
orm
a
n
ce
tr
ai
n
i
n
g
a
n
d
i
n
fe
r
e
n
ce
wor
k
lo
ad
s
,
i
n
c
omp
e
t
i
t
i
on
w
i
t
h
i
n
c
um
be
nt
GP
U
solut
i
ons
(
not
ab
ly
N
V
IDIA
ʼ
s
GP
U
s
).
C
e
r
eb
r
a
s
p
i
on
ee
r
ed
w
afe
r
-
s
ca
l
e
pro
ce
ssors
to
m
a
x
i
m
i
z
e
p
a
r
a
ll
e
l
i
sm
,
S
a
m
ba
N
ov
a
de
v
e
lop
ed
a
r
ec
on
fig
ur
ab
l
e
da
t
af
low
a
r
chi
t
ec
tur
e
(
RD
U
s
)
i
n
i
nt
eg
r
a
t
ed
syst
e
ms
,
a
n
d
G
roq
de
s
ig
n
ed
a
nov
e
l
str
ea
m
i
n
g
L
a
n
g
u
age
P
ro
ce
ss
i
n
g
U
n
i
t
(
LP
U
)
f
o
c
us
ed
pr
i
m
a
r
i
ly
on
i
n
fe
r
e
n
ce
.
A
s
o
f
2025,
a
ll
t
h
r
ee
fi
rms
ha
v
e
achie
v
ed
mult
i
-
bi
ll
i
on
-
d
oll
a
r
v
a
lu
a
t
i
ons
a
n
d
s
ec
ur
ed
m
aj
or
f
un
di
n
g
:
C
e
r
eb
r
a
s
r
ai
s
ed
$
1.1
bi
ll
i
on
a
t
a
n
$
8.1
bi
ll
i
on
v
a
lu
a
t
i
on
(
[1]
www
.
r
e
ut
e
rs
.
c
om
),
G
roq
r
ai
s
ed
$
750
m
i
ll
i
on
pus
hi
n
g
i
ts
v
a
lu
a
t
i
on
to
$
6.9
bi
ll
i
on
(
[2]
www
.
r
e
ut
e
rs
.
c
om
),
a
n
d
S
a
m
ba
N
ov
a
r
ai
s
ed
h
un
d
r
ed
s
o
f
m
i
ll
i
ons
(
e
.
g
.
$
676
M S
e
r
ie
s
D
i
n
2021
w
i
t
h
a
$
5.1
B
v
a
lu
a
t
i
on
)
(
[3]
www
.
a
n
a
n
d
t
ech
.
c
om
)
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
).
E
ach
c
omp
a
ny
ha
s
propr
ie
t
a
ry
chi
p
de
s
ig
ns
,
so
f
tw
a
r
e
st
ack
s
,
a
n
d
de
ploym
e
nt
t
a
r
ge
ts
:
C
e
r
eb
r
a
s
S
yst
e
ms
(
f
oun
ded
2016,
M
e
nlo
P
a
r
k
CA
)
r
e
m
ai
ns
f
o
c
us
ed
on
tr
ai
n
i
n
g
o
f
v
e
ry
l
a
r
ge
AI
mo
de
ls
.
I
ts
w
afe
r
-
s
ca
l
e
chi
ps
(
W
afe
r
-
S
ca
l
e
E
n
gi
n
e
,
W
SE
)
i
nt
eg
r
a
t
e
tr
i
ll
i
ons
o
f
tr
a
ns
i
stors
on
a
s
i
n
g
l
e
monol
i
t
hic
die
,
e
n
ab
l
i
n
g
m
a
ss
i
v
e
on
-
chi
p
c
omput
e
a
n
d
m
e
mory
ba
n
d
w
id
t
h
.
T
he
t
hi
r
d
-
ge
n
e
r
a
t
i
on
W
SE
-3
(
a
nnoun
ced
M
a
r
ch
2024)
ca
n
tr
ai
n
mo
de
ls
“
t
e
n
t
ime
s
la
r
ge
r
t
ha
n
O
p
e
n
AI
ʼ
s
GP
T
-4
”
(
[5]
t
i
m
e
.
c
om
),
a
n
d
i
s
a
t
t
he
hea
rt
o
f
sup
e
r
c
omput
e
rs
l
ike
t
he
C
on
d
or
G
ala
xy
3
.
I
n
l
a
t
e
2025
C
e
r
eb
r
a
s
i
s
e
xp
a
n
di
n
g
i
nto
n
e
w
m
a
r
ke
ts
(
e
.
g
.
U
AE
da
t
a
ce
nt
e
rs
)
(
[6]
www
.
r
e
ut
e
rs
.
c
om
)
a
n
d
won
a
DARPA
c
ontr
ac
t
to
l
i
n
k
i
ts
chi
ps
v
ia
p
h
oton
ic
rout
e
rs
(
[7]
www
.
r
e
ut
e
rs
.
c
om
).
T
he
c
omp
a
ny
ha
s
i
nv
e
st
ed
i
n
s
ca
l
i
n
g
pro
d
u
c
t
i
on
(
us
i
n
g
T
SMC
ʼ
s
3
nm
pro
ce
ss
f
or
W
SE
-3)
a
n
d
ha
s
postpon
ed
a
n
IPO
af
t
e
r
i
ts
l
a
t
e
-2025
f
un
di
n
g
.
S
a
m
ba
N
ov
a
S
yst
e
ms
(
f
oun
ded
2017,
P
a
lo
A
lto
CA
)
p
i
on
ee
r
ed
a
R
ec
on
fig
ur
ab
l
e
D
a
t
af
low
A
r
chi
t
ec
tur
e
(
RDA
)
i
mpl
e
m
e
nt
ed
i
n
i
ts
RD
U
(
R
ec
on
fig
ur
ab
l
e
D
a
t
af
low
U
n
i
t
)
chi
ps
a
n
d
i
nt
eg
r
a
t
ed
D
a
t
a
S
ca
l
e
®
syst
e
ms
.
S
a
m
ba
N
ov
a
ʼ
s
a
r
chi
t
ec
tur
e
us
e
s
a
rr
a
ys
o
f
ada
pt
ab
l
e
c
omput
e
/
da
t
a
un
i
ts
(
RD
U
s
)
t
ha
t
ca
n
be
t
e
mpor
a
lly
a
n
d
sp
a
t
ia
lly
c
on
fig
ur
ed
b
y
t
he
c
omp
i
l
e
r
f
or
diffe
r
e
nt
n
e
ur
a
l
n
e
twor
k
l
a
y
e
rs
(
[8]
www
.
a
n
a
n
d
t
ech
.
c
om
).
T
hi
s
de
s
ig
n
e
mp
ha
s
i
z
e
s
on
-
chi
p
m
e
mory
a
n
d
da
t
af
low
to
efficie
ntly
e
x
ec
ut
e
b
ot
h
tr
ai
n
i
n
g
a
n
d
i
n
fe
r
e
n
ce
.
B
y
2024
S
a
m
ba
N
ov
a
pu
b
l
ic
ly
i
ntro
d
u
ced
S
a
m
ba
-1
,
a
1-
tr
i
ll
i
on
-
p
a
r
a
m
e
t
e
r
op
e
n
-
sour
ce
l
a
n
g
u
age
mo
de
l
ai
m
ed
a
t
e
nt
e
rpr
i
s
e
/
c
ustom
us
e
(
[9]
t
i
m
e
.
c
om
).
H
igh
-
pro
fi
l
e
de
ploym
e
nts
i
n
c
lu
de
U
.
S
.
n
a
t
i
on
a
l
l
ab
s
(
e
.
g
.
L
os
A
l
a
mos
,
LLNL
)
f
or
l
a
r
ge
-
s
ca
l
e
AI
r
e
s
ea
r
ch
(
[10]
www
.
a
n
a
n
d
t
ech
.
c
om
),
a
n
d
i
ts
so
f
tw
a
r
e
st
ack
(
S
a
m
ba
F
low
/
S
a
m
ba
S
tu
di
o
)
supports
c
ommon
AI
f
r
a
m
e
wor
k
s
.
S
a
m
ba
N
ov
a
ʼ
s
syst
e
ms
o
ffe
r
l
a
r
ge
lo
ca
l
m
e
mory
(
e
.
g
.
3
T
B
p
e
r
no
de
i
n
i
ts
SN
30
ge
n
e
r
a
t
i
on
)
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
)
to
ha
n
d
l
e
f
oun
da
t
i
on
mo
de
ls
.
T
he
c
omp
a
ny
c
ont
i
nu
e
s
to
i
t
e
r
a
t
e
on
i
ts
chi
ps
(
e
.
g
.
d
ou
b
l
i
n
g
c
omput
e
a
n
d
m
e
mory
c
omp
a
r
ed
to
pr
edece
ssors
)
a
n
d
t
a
r
ge
ts
s
ec
tors
f
rom
defe
ns
e
to
c
lou
d
e
nt
e
rpr
i
s
e
.
G
roq
(
f
oun
ded
2016,
M
ount
ai
n
V
ie
w
CA
)
i
s
not
ab
l
e
f
or
i
ts
s
i
n
gle
-
c
or
e
,
de
t
e
r
mi
n
i
st
ic
LP
U
a
r
chi
t
ec
tur
e
.
G
roq
ʼ
s
L
a
n
g
u
age
P
ro
ce
ss
i
n
g
U
n
i
ts
i
mpl
e
m
e
nt
a
n
i
n
-
or
de
r
,
to
ke
n
-
str
ea
m
i
n
g
de
s
ig
n
w
he
r
e
e
v
e
ry
c
lo
ck
c
y
c
l
e
e
x
ec
ut
e
s
us
ef
ul
wor
k
w
i
t
h
out
ha
r
d
w
a
r
e
c
ont
e
xt
sw
i
t
chi
n
g
(
G
roq
ca
lls
t
hi
s
“
ke
rn
e
l
-
f
r
ee
”
(
[12]
www
.
ee
t
i
m
e
s
.
c
om
)).
T
hi
s
y
ie
l
d
s
pr
edic
t
ab
ly
low
-
l
a
t
e
n
c
y
AI
i
n
fe
r
e
n
ce
p
e
r
f
orm
a
n
ce
.
T
he
f
oun
de
r
,
e
x
-
G
oo
g
l
e
T
P
U
a
r
chi
t
ec
t
J
on
a
t
ha
n
R
oss
,
ha
s
tout
ed
LP
U
s
a
s
o
ffe
r
i
n
g
“
t
e
n
t
ime
s
fa
st
e
r
a
n
d
t
e
n
t
ime
s
l
ow
e
r
”
c
ost
i
n
fe
r
e
n
ce
t
ha
n
GP
U
s
(
[13]
t
i
m
e
.
c
om
).
G
roq
ʼ
s
chi
ps
(
fab
r
ica
t
ed
on
s
i
m
i
l
a
r
no
de
s
)
f
o
c
us
e
x
c
lus
i
v
e
ly
on
i
n
fe
r
e
n
ce
acce
l
e
r
a
t
i
on
;
t
he
c
omp
a
ny
ha
s
a
ttr
ac
t
ed
m
aj
or
i
nv
e
stm
e
nt
(
S
a
msun
g
,
C
i
s
c
o
)
a
n
d
a
nnoun
ced
l
a
r
ge
dea
ls
(
e
.
g
.
a
$
1.5
B S
a
u
di
A
r
abia
c
omm
i
tm
e
nt
f
or
AI
i
n
f
r
a
stru
c
tur
e
)
(
[14]
www
.
r
e
ut
e
rs
.
c
om
).
G
roq
i
s
e
xp
a
n
di
n
g
da
t
a
ce
nt
e
r
de
ploym
e
nts
(
e
.
g
.
a
E
urop
ea
n
AI
-
f
o
c
us
ed
faci
l
i
ty
i
n
F
i
nl
a
n
d
)
to
ca
ptur
e
t
he
g
row
i
n
g
i
n
fe
r
e
n
ce
m
a
r
ke
t
(
[15]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
).
T
hi
s
r
e
port
de
lv
e
s
i
nto
each
c
omp
a
ny
ʼ
s
hi
story
,
t
ech
nolo
g
y
a
n
d
a
r
chi
t
ec
tur
e
,
p
e
r
f
orm
a
n
ce
c
l
ai
ms
,
m
a
r
ke
t
pos
i
t
i
on
i
n
g
,
a
n
d
us
e
-
ca
s
e
de
ploym
e
nts
,
c
omp
a
r
i
n
g
t
he
m
s
ide
-
b
y
-
s
ide
w
he
r
e
v
e
r
poss
ib
l
e
.
I
t
summ
a
r
i
z
e
s
c
urr
e
nt
m
e
tr
ic
s
(
pro
ce
ssors
ʼ
s
ca
l
e
,
m
e
mory
,
be
n
ch
m
a
r
k
s
w
he
r
e
a
v
ai
l
ab
l
e
)
a
n
d
f
un
di
n
g
/
v
a
lu
a
t
i
on
da
t
a
.
C
a
s
e
stu
die
s
i
n
c
lu
de
n
a
t
i
on
a
l
l
ab
sup
e
r
c
omput
i
n
g
pro
jec
ts
a
n
d
l
a
r
ge
f
oun
da
t
i
on
-
mo
de
l
tr
ai
n
i
n
g
syst
e
ms
.
F
i
n
a
lly
,
w
e
di
s
c
uss
i
n
d
ustry
tr
e
n
d
s
a
n
d
f
utur
e
di
r
ec
t
i
ons
–
i
n
c
lu
di
n
g
h
ow
t
he
s
e
sp
ecia
l
i
z
ed
AI
acce
l
e
r
a
tors
m
igh
t
fi
t
i
nto
e
volv
i
n
g
AI
/
da
t
a
-
ce
nt
e
r
ec
osyst
e
ms
,
p
a
rtn
e
rs
hi
ps
,
a
n
d
pot
e
nt
ia
l
r
eg
ul
a
t
i
on
or
m
a
r
ke
t
s
hif
ts
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 2 of 19
I
ntro
d
u
c
t
i
on
T
he
ad
v
e
nt
o
f
l
a
r
ge
-
s
ca
l
e
m
achi
n
e
l
ea
rn
i
n
g
(
e
sp
ecia
lly
dee
p
n
e
ur
a
l
n
e
twor
k
s
a
n
d
ge
n
e
r
a
t
i
v
e
AI
)
ha
s
d
r
i
v
e
n
de
m
a
n
d
f
or
sp
ecia
l
i
z
ed
high
-
p
e
r
f
orm
a
n
ce
ha
r
d
w
a
r
e
.
W
hi
l
e
GP
U
s
(
pr
i
m
a
r
i
ly
f
rom
N
V
IDIA
)
ha
v
e
d
om
i
n
a
t
ed
AI
tr
ai
n
i
n
g
a
n
d
i
n
fe
r
e
n
ce
ov
e
r
t
he
p
a
st
decade
,
a
n
e
w
w
a
v
e
o
f
st
a
rtups
i
s
pro
d
u
ci
n
g
dedica
t
ed
AI
acce
l
e
r
a
tors
t
ha
t
f
ollow
a
lt
e
rn
a
t
i
v
e
de
s
ig
n
p
hi
losop
hie
s
.
T
he
s
e
ai
m
to
ov
e
r
c
om
e
GP
U
b
ottl
e
n
eck
s
(
m
e
mory
ba
n
d
w
id
t
h
,
s
ca
l
i
n
g
,
pow
e
r
)
b
y
c
ustom
i
z
i
n
g
t
he
chi
p
a
r
chi
t
ec
tur
e
a
n
d
syst
e
m
de
s
ig
n
f
or
AI
wor
k
lo
ad
s
.
A
s
o
f
2025,
t
he
AI
-
chi
p
m
a
r
ke
t
i
s
g
row
i
n
g
r
a
p
id
ly
.
G
a
rtn
e
r
pro
jec
t
ed
t
ha
t
g
lo
ba
l
r
e
v
e
nu
e
f
or
AI
acce
l
e
r
a
tors
woul
d
mor
e
t
ha
n
d
ou
b
l
e
f
rom
2023
to
2027
(
[15]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
)
(
[1]
www
.
r
e
ut
e
rs
.
c
om
).
B
ro
adc
om
ʼ
s
CEO
ha
s
f
or
eca
st
a
$
60–90
bi
ll
i
on
AI
r
e
v
e
nu
e
opportun
i
ty
b
y
2027
(
[16]
www
.
r
e
ut
e
rs
.
c
om
).
N
v
idia
r
e
m
ai
ns
t
he
m
a
r
ke
t
l
eade
r
–
on
e
a
n
a
lys
i
s
e
st
i
m
a
t
ed
i
ts
AI
-
r
e
l
a
t
ed
s
a
l
e
s
c
oul
d
a
ppro
ach
$
400
bi
ll
i
on
b
y
2028
(
[17]
www
.
t
ech
r
ada
r
.
c
om
)
–
b
ut
i
nv
e
stors
a
r
e
pour
i
n
g
ca
p
i
t
a
l
i
nto
GP
U
“
cha
ll
e
n
ge
rs
.”
I
n
t
hi
s
c
ont
e
xt
,
C
e
r
eb
r
a
s
,
S
a
m
ba
N
ov
a
,
a
n
d
G
roq
ha
v
e
e
m
e
r
ged
a
s
high
-
pro
fi
l
e
c
ont
e
n
de
rs
,
c
oll
ec
t
i
v
e
ly
r
ai
s
i
n
g
bi
ll
i
ons
o
f
d
oll
a
rs
(
be
tw
ee
n
t
he
m
,
on
t
he
or
de
r
o
f
$
3–4
B
o
f
e
xt
e
rn
a
l
f
un
di
n
g
)
a
n
d
c
omm
a
n
di
n
g
mult
i
-
bi
ll
i
on
v
a
lu
a
t
i
ons
(
[1]
www
.
r
e
ut
e
rs
.
c
om
)(
[2]
www
.
r
e
ut
e
rs
.
c
om
)
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
).
T
hei
r
t
ech
nolo
gie
s
a
r
e
a
lr
ead
y
de
ploy
ed
i
n
c
r
i
t
ica
l
pro
jec
ts
–
f
rom
U
.
S
.
D
e
p
a
rtm
e
nt
o
f
E
n
e
r
g
y
(
DOE
)
sup
e
r
c
omput
i
n
g
ce
nt
e
rs
to
M
idd
l
e
E
a
st
AI
da
t
a
ce
nt
e
rs
–
a
n
d
t
he
y
he
r
a
l
d
n
e
w
a
r
chi
t
ec
tur
e
s
f
or
AI
c
omput
e
.
T
hi
s
r
e
port
c
omp
a
r
e
s
t
he
s
e
t
h
r
ee
c
omp
a
n
ie
s
s
ide
-
b
y
-
s
ide
,
c
ov
e
r
i
n
g
:
C
omp
a
ny
B
ackg
roun
d
&
F
un
di
n
g
:
O
r
igi
ns
,
l
eade
rs
hi
p
,
m
i
l
e
ston
e
s
(
pro
d
u
c
t
a
nnoun
ce
m
e
nts
,
p
a
rtn
e
rs
hi
ps
),
a
n
d
fi
n
a
n
cia
l
st
a
tus
(
r
ece
nt
f
un
di
n
g
roun
d
s
a
n
d
v
a
lu
a
t
i
ons
).
H
a
r
d
w
a
r
e
A
r
chi
t
ec
tur
e
:
C
or
e
de
s
ig
n
o
f
t
he
pro
ce
ssors
(
C
e
r
eb
r
a
s
ʼ
w
afe
r
-
s
ca
l
e
e
n
gi
n
e
s
,
S
a
m
ba
N
ov
a
ʼ
s
r
ec
on
fig
ur
ab
l
e
da
t
af
low
un
i
ts
,
G
roq
ʼ
s
LP
U
s
),
i
n
c
lu
di
n
g
chi
p
or
ga
n
i
z
a
t
i
on
(
e
.
g
.
W
SE
ʼ
s
m
e
s
h
n
e
twor
k
a
n
d
M
e
mory
X
b
lo
ck
,
S
a
m
ba
N
ov
a
ʼ
s
RD
U
t
i
l
e
s
a
n
d
i
nt
e
r
c
onn
ec
t
,
G
roq
ʼ
s
l
i
n
ea
r
p
i
p
e
l
i
n
e
o
f
f
un
c
t
i
on
a
l
un
i
ts
).
D
i
s
c
uss
i
on
o
f
pro
ce
ss
no
de
s
a
n
d
s
ca
l
e
(
tr
a
ns
i
stor
c
ount
,
die
s
i
z
e
)
w
he
r
e
da
t
a
i
s
a
v
ai
l
ab
l
e
.
S
yst
e
m
&
S
o
f
tw
a
r
e
S
t
ack
s
:
P
ackagi
n
g
i
nto
syst
e
ms
(
e
.
g
.
C
e
r
eb
r
a
s
CS
-2/
C
5
syst
e
ms
,
S
a
m
ba
N
ov
a
D
a
t
a
S
ca
l
e
r
ack
s
,
G
roq
C
a
r
d
/
G
roq
R
ack
),
a
n
d
t
he
so
f
tw
a
r
e
ec
osyst
e
m
(
c
omp
i
l
e
rs
,
f
r
a
m
e
wor
k
s
,
mo
de
l
support
).
P
e
r
f
orm
a
n
ce
a
n
d
B
e
n
ch
m
a
r
k
s
:
P
u
b
l
i
s
hed
or
l
eaked
p
e
r
f
orm
a
n
ce
c
omp
a
r
i
sons
f
or
typ
ica
l
AI
wor
k
lo
ad
s
(
e
.
g
.
tr
ai
n
i
n
g
l
a
r
ge
tr
a
ns
f
orm
e
r
mo
de
ls
or
runn
i
n
g
i
n
fe
r
e
n
ce
),
i
n
c
lu
di
n
g
m
e
tr
ic
s
l
ike
t
h
rou
gh
put
(
to
ke
ns
/
s
ec
,
FLOPS
),
m
e
mory
ca
p
aci
ty
,
a
n
d
e
n
e
r
g
y
efficie
n
c
y
.
W
e
i
n
c
lu
de
w
he
r
e
v
e
r
poss
ib
l
e
i
n
de
p
e
n
de
nt
be
n
ch
m
a
r
k
s
or
c
l
ai
ms
f
rom
r
e
put
ed
sour
ce
s
(
[5]
t
i
m
e
.
c
om
)
(
[13]
t
i
m
e
.
c
om
).
U
s
e
C
a
s
e
s
/
C
a
s
e
S
tu
die
s
:
R
ea
l
-
worl
d
de
ploym
e
nts
(
e
.
g
.
DARPA
c
ontr
ac
t
f
or
C
e
r
eb
r
a
s
i
nt
e
r
c
onn
ec
t
(
[7]
www
.
r
e
ut
e
rs
.
c
om
),
DOE
l
ab
sup
e
r
c
omput
i
n
g
,
E
urop
ea
n
AI
da
t
a
ce
nt
e
r
,
e
t
c
.)
t
ha
t
i
llustr
a
t
e
h
ow
c
ustom
e
rs
a
r
e
us
i
n
g
each
pl
a
t
f
orm
.
M
a
r
ke
t
P
os
i
t
i
on
&
T
r
e
n
d
s
:
H
ow
each
pos
i
t
i
ons
i
ts
e
l
f
r
e
l
a
t
i
v
e
to
GP
U
s
a
n
d
each
ot
he
r
,
str
a
t
egic
p
a
rtn
e
rs
hi
ps
(
e
.
g
.
C
e
r
eb
r
a
s
–
G
42
i
n
U
AE
(
[6]
www
.
r
e
ut
e
rs
.
c
om
),
G
roq
–
S
a
msun
g
,
C
i
s
c
o
(
[18]
www
.
r
e
ut
e
rs
.
c
om
)),
a
n
d
t
he
c
omp
e
t
i
t
i
v
e
l
a
n
d
s
ca
p
e
o
f
AI
i
n
f
r
a
stru
c
tur
e
.
D
i
s
c
uss
i
on
o
f
i
nv
e
stor
s
e
nt
i
m
e
nt
a
n
d
IPO
prosp
ec
ts
(
e
.
g
.
C
e
r
eb
r
a
s
ʼ
U
.
S
.
l
i
st
i
n
g
pl
a
ns
(
[1]
www
.
r
e
ut
e
rs
.
c
om
)).
F
utur
e
D
i
r
ec
t
i
ons
:
P
ot
e
nt
ia
l
e
volut
i
on
a
ry
p
a
t
h
s
(
e
.
g
.
n
e
xt
-
ge
n
chi
ps
,
so
f
tw
a
r
e
i
mprov
e
m
e
nts
,
M
&
A
),
a
n
d
c
ons
ide
r
a
t
i
ons
su
ch
a
s
op
e
n
-
sour
ce
RISC
-
V
mov
e
m
e
nt
,
n
a
t
i
on
a
l
AI
i
n
i
t
ia
t
i
v
e
s
,
a
n
d
h
ow
t
he
i
n
fe
r
e
n
ce
vs
.
tr
ai
n
i
n
g
m
a
r
ke
t
spl
i
t
m
igh
t
e
volv
e
.
B
y
b
r
i
n
gi
n
g
num
e
rous
t
ech
n
ica
l
de
t
ai
ls
,
da
t
a
po
i
nts
,
a
n
d
ci
t
ed
op
i
n
i
ons
to
ge
t
he
r
,
t
hi
s
r
e
port
ai
ms
to
be
a
c
ompr
ehe
ns
i
v
e
r
e
sour
ce
on
t
he
st
a
t
e
o
f
t
he
s
e
t
h
r
ee
AI
chi
p
fi
rms
a
s
o
f
l
a
t
e
2025.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 3 of 19
1.
C
omp
a
ny
B
ackg
roun
d
a
n
d
F
un
di
n
g
1.1
C
e
r
eb
r
a
s
S
yst
e
ms
C
e
r
eb
r
a
s
S
yst
e
ms
w
a
s
f
oun
ded
i
n
2016
b
y
A
n
d
r
e
w
F
e
l
d
m
a
n
a
n
d
ot
he
r
v
e
t
e
r
a
ns
f
rom
AI
st
a
rtups
a
n
d
sup
e
r
c
omput
i
n
g
.
T
he
c
omp
a
ny
ʼ
s
defi
n
i
n
g
idea
i
s
to
b
u
i
l
d
e
xtr
e
m
e
ly
l
a
r
ge
“
w
afe
r
-
s
ca
l
e
”
chi
ps
t
ha
t
b
r
eak
t
he
mol
d
o
f
tr
adi
t
i
on
a
l
r
e
t
ic
l
e
-
l
i
m
i
t
ed
die
s
(
[5]
t
i
m
e
.
c
om
).
A
f
t
e
r
st
ea
lt
h
de
v
e
lopm
e
nt
,
C
e
r
eb
r
a
s
unv
ei
l
ed
i
ts
fi
rst
W
afe
r
-
S
ca
l
e
E
n
gi
n
e
(
W
SE
)
i
n
2019,
a
s
i
n
g
l
e
-
die
chi
p
c
ov
e
r
i
n
g
most
o
f
a
300
mm
s
i
l
ic
on
w
afe
r
(
a
pprox
i
m
a
t
e
ly
46,225
mm
²
)
(
[5]
t
i
m
e
.
c
om
).
T
hi
s
chi
p
c
ont
ai
n
ed
~1.2
tr
i
ll
i
on
tr
a
ns
i
stors
a
t
7
nm
a
n
d
t
h
ous
a
n
d
s
o
f
AI
-
opt
i
m
i
z
ed
c
or
e
s
w
i
t
h
m
a
ss
i
v
e
on
-
chi
p
SRAM
.
T
he
g
o
a
l
w
a
s
to
c
r
ea
t
e
on
e
pro
ce
ssor
w
i
t
h
a
s
mu
ch
m
e
mory
a
n
d
i
nt
e
r
c
onn
ec
t
on
-
chi
p
a
s
1,000+
tr
adi
t
i
on
a
l
GP
U
s
,
e
l
i
m
i
n
a
t
i
n
g
i
nt
e
r
-
GP
U
c
ommun
ica
t
i
on
b
ottl
e
n
eck
s
.
F
un
di
n
g
a
n
d
V
a
lu
a
t
i
on
:
C
e
r
eb
r
a
s
r
ai
s
ed
mult
i
pl
e
v
e
ntur
e
roun
d
s
t
h
rou
gh
2023.
I
ts
S
e
r
ie
s
G
i
n
2023
w
a
s
$
275
M
a
t
a
r
e
port
ed
$
8.1
B
v
a
lu
a
t
i
on
.
I
n
l
a
t
e
2025,
R
e
ut
e
rs
r
e
port
ed
t
ha
t
C
e
r
eb
r
a
s
r
ai
s
ed
$
1.1
bi
ll
i
on
i
n
n
e
w
ca
p
i
t
a
l
(
l
ed
b
y
F
ide
l
i
ty
a
n
d
A
tr
eide
s
)
(
[1]
www
.
r
e
ut
e
rs
.
c
om
),
b
r
i
n
gi
n
g
t
he
v
a
lu
a
t
i
on
agai
n
to
$
8.1
bi
ll
i
on
.
N
ot
ab
ly
,
2025
a
lso
s
a
w
i
nv
e
stm
e
nt
f
rom
1789
C
a
p
i
t
a
l
,
a
T
rump
-
l
i
n
ked
fi
rm
(
[1]
www
.
r
e
ut
e
rs
.
c
om
).
S
h
ortly
af
t
e
r
t
hi
s
f
un
di
n
g
,
C
e
r
eb
r
a
s
w
i
t
hd
r
e
w
pl
a
ns
f
or
a
U
.
S
.
IPO
(
or
igi
n
a
lly
sol
ici
t
i
n
g
a
l
a
r
ge
o
ffe
r
i
n
g
)
(
[19]
www
.
r
e
ut
e
rs
.
c
om
).
T
he
c
omp
a
ny
e
xpl
ai
n
ed
t
he
IPO
de
l
a
y
a
s
d
u
e
to
r
eg
ul
a
tory
r
e
v
ie
ws
(
i
n
c
lu
di
n
g
a
U
.
S
.
n
a
t
i
on
a
l
-
s
ec
ur
i
ty
r
e
v
ie
w
o
f
a
n
ea
rl
ie
r
$
335
M G
42
i
nv
e
stm
e
nt
)
a
n
d
i
nst
ead
ch
os
e
to
e
xp
a
n
d
pr
i
v
a
t
e
f
un
di
n
g
(
[19]
www
.
r
e
ut
e
rs
.
c
om
)
(
[1]
www
.
r
e
ut
e
rs
.
c
om
).
C
ustom
e
rs
a
n
d
P
a
rtn
e
rs
hi
ps
:
C
e
r
eb
r
a
s
m
a
r
ke
ts
i
ts
CS
-2
a
n
d
CS
-3
syst
e
ms
to
AI
r
e
s
ea
r
ch
l
ab
s
a
n
d
e
nt
e
rpr
i
s
e
s
.
I
ts
fi
rst
ke
y
c
ustom
e
rs
i
n
c
lu
ded
A
r
g
onn
e
a
n
d
O
ak
R
idge
N
a
t
i
on
a
l
L
ab
s
(
f
or
e
x
a
s
ca
l
e
AI
)
a
n
d
G
42
(
a
n
A
b
u
D
habi
AI
fi
rm
)
(
[6]
www
.
r
e
ut
e
rs
.
c
om
).
I
n
2025,
C
e
r
eb
r
a
s
a
nnoun
ced
pl
a
ns
to
supply
i
n
f
r
a
stru
c
tur
e
f
or
t
he
S
t
a
r
ga
t
e
U
AE
AI
da
t
a
ce
nt
e
r
ca
mpus
,
i
n
dica
t
i
n
g
e
xp
a
ns
i
on
i
n
t
he
M
idd
l
e
E
a
st
(
[6]
www
.
r
e
ut
e
rs
.
c
om
).
T
he
c
omp
a
ny
a
lso
won
a
DARPA
c
ontr
ac
t
f
or
a
n
ad
v
a
n
ced
AI
sup
e
r
c
omput
i
n
g
eff
ort
(
ca
ll
ed
“
F
us
e
”
pro
jec
t
)
v
a
lu
ed
a
t
$
45
M
(
[7]
www
.
r
e
ut
e
rs
.
c
om
),
p
a
rtn
e
r
i
n
g
i
ts
W
SE
chi
ps
w
i
t
h
R
a
novus
ʼ
s
p
h
oton
ic
i
nt
e
r
c
onn
ec
t
to
b
u
i
l
d
a
syst
e
m
“150
t
i
m
e
s
fa
st
e
r
”
t
ha
n
c
onv
e
nt
i
on
a
l
on
e
s
.
T
he
s
e
dea
ls
high
l
igh
t
C
e
r
eb
r
a
s
ʼ
s
pos
i
t
i
on
i
n
g
i
n
l
a
r
ge
-
s
ca
l
e
g
ov
e
rnm
e
nt
a
n
d
r
e
s
ea
r
ch
pro
jec
ts
.
1.2
S
a
m
ba
N
ov
a
S
yst
e
ms
F
oun
ded
i
n
2017
b
y
f
orm
e
r
S
un
M
ic
rosyst
e
ms
e
x
ec
ut
i
v
e
s
R
o
d
r
ig
o
L
ia
n
g
(
CEO
)
a
n
d
K
unl
e
O
lu
k
otun
(
C
T
O
),
S
a
m
ba
N
ov
a
b
u
i
lt
a
h
y
b
r
id
ha
r
d
w
a
r
e
-
so
f
tw
a
r
e
st
ack
ce
nt
e
r
ed
on
i
ts
R
ec
on
fig
ur
ab
l
e
D
a
t
af
low
U
n
i
t
(
RD
U
)
a
r
chi
t
ec
tur
e
.
S
a
m
ba
N
ov
a
S
yst
e
ms
ke
pt
a
low
pro
fi
l
e
i
n
i
t
ia
lly
,
r
ai
s
i
n
g
v
e
ntur
e
f
un
di
n
g
qu
ie
tly
.
B
y
l
a
t
e
2020,
i
t
had
a
ttr
ac
t
ed
~
$
450
M
i
n
tot
a
l
f
un
di
n
g
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
),
f
rom
i
nv
e
stors
l
ike
G
oo
g
l
e
V
e
ntur
e
s
,
I
nt
e
l
C
a
p
i
t
a
l
,
a
n
d
B
l
ack
R
o
ck
(
[10]
www
.
a
n
a
n
d
t
ech
.
c
om
).
I
n
A
pr
i
l
2021,
S
a
m
ba
N
ov
a
a
nnoun
ced
a
S
e
r
ie
s
D
roun
d
o
f
$
676
M
a
t
a
$
5.1
B
v
a
lu
a
t
i
on
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
),
b
r
ief
ly
m
aki
n
g
i
t
on
e
o
f
t
he
most
r
ich
ly
backed
AI
st
a
rtups
.
A
r
chi
t
ec
tur
e
&
P
ro
d
u
c
ts
:
S
a
m
ba
N
ov
a
ʼ
s
v
a
lu
e
propos
i
t
i
on
i
s
a
turn
ke
y
AI
pl
a
t
f
orm
(
ha
r
d
w
a
r
e
+
so
f
tw
a
r
e
).
I
ts
D
a
t
a
S
ca
l
e
syst
e
ms
c
ons
i
st
o
f
b
o
a
r
d
s
ca
ll
ed
D
a
t
a
S
cale
S
N
(
su
ch
a
s
SN
10,
SN
15,
SN
30)
t
ha
t
c
ont
ai
n
mult
i
pl
e
RD
U
acce
l
e
r
a
tor
chi
ps
a
n
d
a
sso
cia
t
ed
m
e
mory
.
A
“
SN
10-8
R
”
c
on
fig
ur
a
t
i
on
(8
so
cke
ts
each
w
i
t
h
a
n
RD
U
b
o
a
r
d
a
n
d
c
onn
ec
t
ed
m
e
mory
)
w
a
s
l
a
un
ched
i
n
2020
(
[10]
www
.
a
n
a
n
d
t
ech
.
c
om
),
de
monstr
a
t
i
n
g
t
hei
r
fi
rst
pro
d
u
c
t
f
or
l
i
m
i
t
ed
c
ustom
e
rs
.
T
he
RD
U
chi
p
i
ts
e
l
f
(
i
nt
e
rn
a
lly
ca
ll
ed
“
C
a
r
di
n
a
l
”)
w
a
s
b
u
i
lt
on
T
SMC
7
nm
,
i
nt
eg
r
a
t
i
n
g
t
h
ous
a
n
d
s
o
f
sm
a
ll
pro
ce
ss
i
n
g
e
l
e
m
e
nts
t
ha
t
ca
n
be
r
ec
on
fig
ur
ed
i
n
so
f
tw
a
r
e
to
f
orm
t
he
da
t
af
low
g
r
a
p
h
o
f
a
n
AI
mo
de
l
(
[8]
www
.
a
n
a
n
d
t
ech
.
c
om
).
T
he
c
omp
i
l
e
r
(
S
a
m
ba
F
low
)
m
a
ps
a
us
e
r
ʼ
s
n
e
ur
a
l
n
e
twor
k
g
r
a
p
h
onto
a
s
e
qu
e
n
ce
o
f
c
on
fig
ur
a
t
i
ons
o
f
t
he
RD
U
a
rr
a
y
,
ai
m
i
n
g
to
m
a
x
i
m
i
z
e
da
t
a
r
e
us
e
a
n
d
m
i
n
i
m
i
z
e
da
t
a
mov
e
m
e
nt
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 4 of 19
E
a
rly
D
e
ploym
e
nts
:
E
v
e
n
bef
or
e
pu
b
l
ic
pro
d
u
c
t
l
a
un
che
s
,
S
a
m
ba
N
ov
a
s
ec
r
e
tly
de
ploy
ed
i
ts
ha
r
d
w
a
r
e
a
t
U
.
S
.
n
a
t
i
on
a
l
l
ab
s
.
A
n
a
n
d
T
ech
r
e
port
ed
t
ha
t
S
a
m
ba
N
ov
a
had
“
s
e
m
i
-
h
us
hh
us
h
de
ploym
e
nts
”
a
t
L
a
wr
e
n
ce
L
i
v
e
rmor
e
a
n
d
L
os
A
l
a
mos
N
a
t
i
on
a
l
L
ab
s
(
[10]
www
.
a
n
a
n
d
t
ech
.
c
om
).
T
hi
s
ea
rly
ad
opt
i
on
su
gge
sts
t
he
c
omp
a
ny
t
a
r
ge
t
ed
high
-
e
n
d
r
e
s
ea
r
ch
wor
k
lo
ad
s
f
rom
t
he
outs
e
t
,
a
lon
g
s
ide
t
hei
r
e
nt
e
rpr
i
s
e
/
g
ov
e
rnm
e
nt
c
ustom
e
rs
.
O
t
he
r
c
ustom
e
rs
r
e
port
ed
ly
i
n
c
lu
de
A
cce
ntur
e
a
n
d
,
r
e
port
ed
ly
,
S
o
f
t
B
a
n
k
(
a
s
not
ed
i
n
m
edia
pro
fi
l
e
s
(
[9]
t
i
m
e
.
c
om
)).
R
ece
nt
D
e
v
e
lopm
e
nts
:
I
n
2024,
S
a
m
ba
N
ov
a
pu
b
l
ic
ly
unv
ei
l
ed
S
a
m
ba
-1
,
i
ts
i
n
-
h
ous
e
tr
ai
n
ed
1-
tr
i
ll
i
on
-
p
a
r
a
m
e
t
e
r
ge
n
e
r
a
t
i
v
e
AI
mo
de
l
(
a
n
op
e
n
f
oun
da
t
i
on
mo
de
l
)
(
[9]
t
i
m
e
.
c
om
).
S
a
m
ba
-1
i
s
i
nt
e
n
ded
f
or
on
-
pr
e
m
i
s
e
s
e
nt
e
rpr
i
s
e
AI
a
ppl
ica
t
i
ons
a
n
d
fi
n
e
-
tun
i
n
g
.
A
ddi
t
i
on
a
lly
,
S
a
m
ba
N
ov
a
l
a
un
ched
S
aa
S
/
c
lou
d
s
e
rv
ice
s
to
run
f
oun
da
t
i
on
mo
de
ls
(
e
.
g
.
L
l
a
m
a
3)
on
i
ts
ha
r
d
w
a
r
e
,
e
mp
ha
s
i
z
i
n
g
p
e
r
f
orm
a
n
ce
.
T
he
c
omp
a
ny
a
nnoun
ced
s
ig
n
ifica
nt
ha
r
d
w
a
r
e
up
g
r
ade
s
:
i
ts
n
e
w
ge
n
e
r
a
t
i
on
d
ou
b
l
e
s
t
he
c
omput
e
a
n
d
m
e
mory
o
f
i
ts
pr
edece
ssor
.
F
or
e
x
a
mpl
e
,
on
e
pu
b
l
ic
st
a
t
e
m
e
nt
not
ed
t
ha
t
each
so
cke
t
i
n
t
he
n
e
xt
D
a
t
a
S
ca
l
e
r
ack
o
ffe
rs
3
T
B
o
f
m
e
mory
(
tw
ice
t
he
pr
i
or
1.5
T
B
)
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
).
A
s
o
f
2025,
S
a
m
ba
N
ov
a
c
ont
i
nu
e
s
to
m
a
r
ke
t
to
e
nt
e
rpr
i
s
e
,
g
ov
e
rnm
e
nt
,
a
n
d
s
cie
nt
ific
c
l
ie
nts
,
pos
i
t
i
on
i
n
g
i
ts
e
l
f
a
s
a
n
a
lt
e
rn
a
t
i
v
e
to
GP
U
-
ba
s
ed
AI
c
lou
d
s
.
F
i
n
a
n
cia
lly
,
no
n
e
w
l
a
r
ge
f
un
di
n
g
roun
d
s
ha
v
e
bee
n
r
e
port
ed
af
t
e
r
2021,
b
ut
t
he
c
omp
a
ny
r
e
m
ai
ns
pr
i
v
a
t
e
.
1.3
G
roq
G
roq
,
f
oun
ded
i
n
2016
b
y
J
on
a
t
ha
n
R
oss
(
a
l
ead
a
r
chi
t
ec
t
o
f
G
oo
g
l
e
ʼ
s
T
P
U
),
e
m
e
r
ged
w
i
t
h
a
r
adica
l
de
s
ig
n
p
hi
losop
h
y
.
I
t
pro
d
u
ce
s
a
s
i
n
g
l
e
-
c
or
e
L
a
n
g
u
age
P
ro
ce
ss
i
n
g
U
n
i
t
(
LP
U
)
i
nt
e
n
ded
f
or
e
xtr
e
m
e
ly
fa
st
,
pr
edic
t
ab
l
e
i
n
fe
r
e
n
ce
o
f
l
a
r
ge
l
a
n
g
u
age
mo
de
ls
a
n
d
ot
he
r
AI
wor
k
lo
ad
s
(
[13]
t
i
m
e
.
c
om
).
G
roq
ʼ
s
ea
rly
f
un
di
n
g
w
a
s
r
e
l
a
t
i
v
e
ly
mo
de
st
(
s
e
r
ie
s
roun
d
s
i
n
2019-2020),
b
ut
i
ts
pro
fi
l
e
ros
e
w
i
t
h
d
r
a
m
a
t
ic
c
l
ai
ms
ab
out
sp
eed
a
n
d
t
h
rou
gh
put
.
LP
U
A
r
chi
t
ec
tur
e
:
G
roq
ʼ
s
LP
U
a
r
chi
t
ec
tur
e
(
or
igi
n
a
lly
ca
ll
ed
t
he
T
e
nsor
S
tr
ea
m
i
n
g
P
ro
ce
ssor
)
feed
s
AI
mo
de
l
to
ke
ns
t
h
rou
gh
a
s
i
n
g
l
e
,
w
ide
p
i
p
e
l
i
n
e
o
f
f
un
c
t
i
on
a
l
un
i
ts
,
e
x
ec
ut
i
n
g
a
ll
op
e
r
a
t
i
ons
i
n
lo
ck
-
st
e
p
w
i
t
h
no
ke
rn
e
l
sw
i
t
chi
n
g
.
T
hi
s
“
da
t
a
-
str
ea
m
”
or
“
t
e
nsor
-
str
ea
m
i
n
g
”
de
s
ig
n
m
ea
ns
e
v
e
ry
c
lo
ck
c
y
c
l
e
i
s
d
o
i
n
g
us
ef
ul
wor
k
.
G
roq
c
l
ai
ms
i
t
achie
v
e
s
“
sp
eed
s
10
x
fa
st
e
r
a
n
d
c
osts
10
x
l
ow
e
r
”
t
ha
n
GP
U
s
f
or
ce
rt
ai
n
i
n
fe
r
e
n
ce
t
a
s
k
s
(
[13]
t
i
m
e
.
c
om
),
a
t
t
he
e
xp
e
ns
e
o
f
f
l
e
x
ibi
l
i
ty
(
LP
U
s
l
ack
d
yn
a
m
ic
b
r
a
n
chi
n
g
or
w
ide
pro
g
r
a
mm
abi
l
i
ty
).
G
roq
a
lso
e
mp
ha
s
i
z
e
s
de
t
e
rm
i
n
i
st
ic
,
low
-
l
a
t
e
n
c
y
c
omput
e
–
a
s
e
ll
i
n
g
po
i
nt
f
or
r
ea
l
-
t
i
m
e
a
ppl
ica
t
i
ons
w
he
r
e
unpr
edic
t
ab
l
e
l
a
t
e
n
c
y
i
s
un
acce
pt
ab
l
e
(
[20]
t
i
m
e
.
c
om
)
(
[21]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
).
F
un
di
n
g
a
n
d
V
a
lu
a
t
i
on
:
G
roq
qu
ie
tly
r
ai
s
ed
ca
p
i
t
a
l
unt
i
l
a
w
a
v
e
o
f
AI
chi
p
i
nt
e
r
e
st
i
n
2024–2025.
I
n
A
u
g
ust
2024
i
t
c
los
ed
a
$
640
M S
e
r
ie
s
D
a
t
a
r
e
port
ed
$
2.8
B
v
a
lu
a
t
i
on
(
[22]
www
.
r
e
ut
e
rs
.
c
om
).
T
he
n
i
n
S
e
pt
e
m
be
r
2025,
R
e
ut
e
rs
r
e
port
ed
G
roq
r
ai
s
ed
$
750
M
(
l
ed
b
y
D
i
srupt
i
v
e
)
a
t
a
$
6.9
B
post
-
mon
e
y
v
a
lu
a
t
i
on
(
[2]
www
.
r
e
ut
e
rs
.
c
om
)
(
[23]
www
.
r
e
ut
e
rs
.
c
om
)
–
mor
e
t
ha
n
d
ou
b
l
i
n
g
i
ts
v
a
lu
a
t
i
on
i
n
a
y
ea
r
.
K
e
y
i
nv
e
stors
i
n
c
lu
de
S
a
msun
g
,
C
i
s
c
o
,
1789
C
a
p
i
t
a
l
,
a
lt
i
m
e
t
e
r
,
a
n
d
ot
he
rs
(
[24]
www
.
r
e
ut
e
rs
.
c
om
)
(
[25]
www
.
r
e
ut
e
rs
.
c
om
).
I
n
2025
G
roq
a
lso
s
ec
ur
ed
a
$
1.5
B
c
omm
i
tm
e
nt
f
rom
S
a
u
di
A
r
abia
to
de
ploy
LP
U
s
,
e
xp
ec
t
ed
to
ge
n
e
r
a
t
e
~
$
500
M
r
e
v
e
nu
e
i
n
on
e
y
ea
r
(
[14]
www
.
r
e
ut
e
rs
.
c
om
).
D
e
ploym
e
nts
:
U
nl
ike
t
he
ot
he
rs
,
G
roq
ha
s
f
o
c
us
ed
hea
v
i
ly
on
b
u
i
l
di
n
g
out
da
t
a
-
ce
nt
e
r
ca
p
aci
ty
to
de
l
i
v
e
r
i
n
fe
r
e
n
ce
-
a
s
-
a
-
s
e
rv
ice
.
I
n
m
id
-2025,
G
roq
op
e
n
ed
a
n
E
U
da
t
a
ce
nt
e
r
i
n
H
e
ls
i
n
ki
(
w
i
t
h
E
qu
i
n
i
x
)
t
a
r
ge
t
i
n
g
E
urop
ea
n
AI
wor
k
lo
ad
s
(
[15]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
).
D
om
e
st
ica
lly
,
t
he
y
op
e
r
a
t
e
“
G
roq
C
lou
d
”
f
or
c
ustom
e
rs
.
K
e
y
c
ustom
e
rs
ha
v
e
not
bee
n
w
ide
ly
a
nnoun
ced
,
b
ut
p
a
rtn
e
r
i
n
g
w
i
t
h
S
a
msun
g
a
n
d
C
i
s
c
o
hi
nts
a
t
e
nt
e
rpr
i
s
e
n
e
twor
ki
n
g
a
n
d
stor
age
i
nt
eg
r
a
t
i
ons
.
G
roq
ʼ
s
CEO
e
mp
ha
s
i
z
e
s
i
ts
pro
d
u
c
t
r
eadi
n
e
ss
a
n
d
“
qu
icke
r
de
l
i
v
e
ry
t
i
m
e
s
”
c
omp
a
r
ed
to
n
e
w
e
r
ASIC
c
ont
e
n
de
rs
(
[21]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
).
A
s
o
f
2025,
G
roq
r
e
m
ai
ns
pr
i
v
a
t
e
;
gi
v
e
n
i
ts
high
v
a
lu
a
t
i
on
a
n
d
g
rowt
h
,
a
n
IPO
or
ac
qu
i
s
i
t
i
on
i
nt
e
r
e
st
woul
d
be
a
f
utur
e
qu
e
st
i
on
(
b
ut
no
su
ch
mov
e
s
ha
v
e
bee
n
r
e
port
ed
y
e
t
).
1.4
C
omp
a
ny
S
n
a
ps
h
ot
a
n
d
F
un
di
n
g
S
umm
a
ry
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 5 of 19
A
summ
a
r
i
z
ed
c
omp
a
r
i
son
o
f
t
he
t
h
r
ee
c
omp
a
n
ie
s
i
s
prov
ided
i
n
t
he
t
ab
l
e
be
low
:
C
omp
a
ny
F
oun
ded
/
HQ
T
ot
a
l
F
un
di
n
g
(
a
pprox
.)
L
a
t
e
st
V
a
lu
a
t
i
on
L
ead
I
nv
e
stors
C
e
r
eb
r
a
s
S
yst
em
s
2016,
M
e
nlo
P
a
r
k
,
CA
(
[1]
www
.
r
e
ut
e
rs
.
c
om
)
~
$
1.6
B
(
a
ll
roun
d
s
t
h
rou
gh
2025)
$
8.1
B
(
S
e
pt
2025)
(
[1]
www
.
r
e
ut
e
rs
.
c
om
)
F
ide
l
i
ty
,
A
tr
eide
s
,
T
ige
r
G
lo
ba
l
,
V
a
lor
,
1789,
S
o
f
t
B
a
n
k
,
E
t
c
.
S
amba
N
ov
a
S
yst
em
s
2017,
P
a
lo
A
lto
,
CA
(
[3]
www
.
a
n
a
n
d
t
ech
.
c
om
)
~
$
1.1
B
(
t
h
rou
gh
2021)
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
)
~
$
5
B
(
A
pr
2021)
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
)
S
o
f
t
B
a
n
k
V
i
s
i
on
F
un
d
,
G
V
,
I
nt
e
l
C
a
p
i
t
a
l
,
B
l
ack
R
o
ck
,
G
V
,
E
t
c
.
G
roq
2016,
M
ount
ai
n
V
ie
w
,
CA
(
[13]
t
i
m
e
.
c
om
)~
$
1.5
B
(
t
h
rou
gh
2025)
$
6.9
B
(
S
e
pt
2025)
(
[2]
www
.
r
e
ut
e
rs
.
c
om
)
B
l
ack
R
o
ck
,
S
a
msun
g
,
C
i
s
c
o
,
1789,
E
t
c
.
T
he
fig
ur
e
s
ab
ov
e
c
om
bi
n
e
pu
b
l
ic
r
e
ports
a
n
d
pr
e
ss
r
e
l
ea
s
e
s
.
C
e
r
eb
r
a
s
ʼ
s
f
un
di
n
g
i
n
c
lu
de
s
a
$
1.1
B
roun
d
i
n
2025
(
[1]
www
.
r
e
ut
e
rs
.
c
om
).
S
a
m
ba
N
ov
a
ʼ
s
l
a
t
e
st
k
nown
roun
d
w
a
s
i
n
2021
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
)
(
no
pu
b
l
ic
n
e
ws
o
f
l
a
t
e
r
roun
d
s
a
s
o
f
2025).
G
roq
ʼ
s
v
a
lu
a
t
i
on
j
ump
r
ef
l
ec
ts
two
l
a
r
ge
roun
d
s
(2024,
2025)
(
[22]
www
.
r
e
ut
e
rs
.
c
om
)
(
[18]
www
.
r
e
ut
e
rs
.
c
om
).
A
ll
t
h
r
ee
ha
v
e
a
ttr
ac
t
ed
m
aj
or
S
i
l
ic
on
V
a
ll
e
y
a
n
d
i
nst
i
tut
i
on
a
l
backe
rs
.
2.
C
hi
p
A
r
chi
t
ec
tur
e
s
a
n
d
H
a
r
d
w
a
r
e
S
p
ecifica
t
i
ons
C
e
r
eb
r
a
s
,
S
a
m
ba
N
ov
a
,
a
n
d
G
roq
each
de
v
e
lop
ed
f
un
da
m
e
nt
a
lly
diffe
r
e
nt
chi
p
a
r
chi
t
ec
tur
e
s
f
or
AI
acce
l
e
r
a
t
i
on
.
T
hi
s
s
ec
t
i
on
a
n
a
lyz
e
s
a
n
d
c
omp
a
r
e
s
t
hei
r
de
s
ig
ns
:
2.1
C
e
r
eb
r
a
s
W
afe
r
-
S
ca
l
e
E
n
gi
n
e
(
W
SE
)
C
e
r
eb
r
a
s
ʼ
s
f
l
ag
s
hi
p
i
s
t
he
W
afe
r
-
S
ca
l
e
E
n
gi
n
e
,
a
s
i
n
g
l
e
monol
i
t
hic
s
i
l
ic
on
die
t
ha
t
sp
a
ns
n
ea
rly
a
n
e
nt
i
r
e
w
afe
r
.
T
he
or
igi
n
a
l
2019
W
SE
had
~1.2
tr
i
ll
i
on
tr
a
ns
i
stors
;
t
he
t
hi
r
d
-
ge
n
e
r
a
t
i
on
W
SE
-3
(
a
nnoun
ced
2024)
r
e
port
ed
ly
ha
s
~~4
tr
i
ll
i
on
tr
a
ns
i
stors
~~
(
N
e
ws
sour
ce
:
T
i
m
e
M
aga
z
i
n
e
)
a
n
d
i
s
b
u
i
lt
on
T
SMC
ʼ
s
3
nm
pro
ce
ss
(
[5]
t
i
m
e
.
c
om
).
T
i
m
e
M
aga
z
i
n
e
not
ed
,
“
C
e
r
eb
r
a
s
S
yst
em
s
…
de
v
el
op
ed
t
he
t
hi
r
d
-
ge
n
e
r
a
t
i
on
WS
E
-3
i
n
M
a
r
ch
2024
.
T
he
WS
E
-3
ca
n
tr
ai
n
AI
m
o
del
s
t
e
n
t
ime
s
la
r
ge
r
t
ha
n
O
p
e
n
AI
ʼ
s
GP
T
-4
”
(
[5]
t
i
m
e
.
c
om
).
T
hi
s
h
yp
e
r
b
ol
ic
c
l
ai
m
un
de
rs
c
or
e
s
t
he
e
normous
chi
p
s
ca
l
e
a
n
d
on
-
chi
p
m
e
mory
(
t
he
W
SE
-3
ha
s
t
h
ous
a
n
d
s
o
f
c
or
e
s
a
n
d
h
un
d
r
ed
s
o
f
MB
o
f
SRAM
p
e
r
c
or
e
c
lust
e
r
,
plus
l
a
r
ge
SRAM
m
e
mory
ba
n
k
s
ca
ll
ed
M
e
mory
X
).
A
ke
y
fea
tur
e
o
f
t
he
W
SE
i
s
i
ts
on
-
chi
p
n
e
twor
k
.
U
nl
ike
GP
U
s
w
hich
r
e
ly
on
o
ff
-
chi
p
m
e
mory
or
ca
r
di
n
a
l
i
nt
e
r
c
onn
ec
ts
,
t
he
W
SE
us
e
s
a
n
i
ntr
a
-
w
afe
r
m
e
s
h
i
nt
e
r
c
onn
ec
t
ca
ll
ed
S
w
a
rm
X
to
l
i
n
k
a
ll
c
omput
i
n
g
t
i
l
e
s
a
n
d
m
e
mory
qu
ad
r
a
nts
.
T
hi
s
prov
ide
s
high
ba
n
d
w
id
t
h
w
i
t
h
low
l
a
t
e
n
c
y
be
tw
ee
n
a
ny
two
po
i
nts
on
t
he
chi
p
.
T
he
M
e
mory
X
stru
c
tur
e
s
h
ol
d
t
he
w
eigh
ts
a
n
d
ac
t
i
v
a
t
i
ons
f
or
mo
de
l
tr
ai
n
i
n
g
e
nt
i
r
e
ly
on
-
chi
p
,
a
vo
idi
n
g
PCI
e
tr
a
ns
fe
r
de
l
a
ys
c
ommon
i
n
GP
U
-
CP
U
syst
e
ms
.
C
e
r
eb
r
a
s
S
yst
e
ms
(
W
SE
-3)
–
K
e
y
p
a
r
a
m
e
t
e
rs
(
pu
b
l
ic
da
t
a
&
e
st
i
m
a
t
e
s
):
T
r
a
ns
i
stors
:
~2–4
tr
i
ll
i
on
(
v
a
r
i
ous
sour
ce
s
gi
v
e
diffe
r
e
nt
c
ounts
)
(
[5]
t
i
m
e
.
c
om
).
P
ro
ce
ss
:
T
SMC
3
nm
(
to
fi
t
t
he
lo
gic
a
n
d
m
e
mory
).
D
ie
s
i
z
e
:
~46,225
mm
²
(
f
ull
w
afe
r
,
a
p
a
rt
f
rom
c
utouts
).
O
n
-
chi
p
M
e
mory
:
H
un
d
r
ed
s
o
f
MB
(
w
i
t
h
s
e
p
a
r
a
t
e
m
e
mory
c
lust
e
rs
).
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 6 of 19
I
nt
eg
r
a
t
i
on
:
W
SE
-3
c
ont
ai
ns
~850,000
AI
-
opt
i
m
i
z
ed
c
or
e
s
(
each
w
i
t
h
lo
ca
l
m
e
mory
a
n
d
MAC
un
i
ts
),
or
ga
n
i
z
ed
i
nto
c
or
e
t
i
l
e
s
.
P
ow
e
r
:
A
ca
ll
ed
syst
e
m
(
CS
-3)
c
onsum
e
s
on
t
he
or
de
r
o
f
20
k
W
f
or
a
s
i
n
g
l
e
W
SE
-3
syst
e
m
un
de
r
f
ull
lo
ad
(
de
r
i
v
ed
f
rom
b
o
a
r
d
sp
ec
s
).
P
e
r
f
orm
a
n
ce
:
P
eak
FP
16/
FP
32
p
e
r
f
orm
a
n
ce
o
f
t
he
chi
p
i
s
not
pu
b
l
ic
ly
di
s
c
los
ed
,
b
ut
i
nt
e
rn
a
l
t
e
sts
s
h
ow
m
a
ss
i
v
e
t
h
rou
gh
put
;
C
e
r
eb
r
a
s
c
l
ai
m
ed
“
210
x
sp
eed
up
ov
e
r
N
V
IDIA H
100
GP
U
”
f
or
a
sp
ecific
mo
de
l
a
n
d
da
t
a
s
e
t
(
[26]
www
.
ce
r
eb
r
a
s
.
n
e
t
)
(
t
h
ou
gh
t
ha
t
i
s
a
n
unv
e
r
ified
v
e
n
d
or
c
l
ai
m
).
S
yst
e
ms
:
T
he
W
SE
chi
ps
a
r
e
p
ackaged
i
nto
s
e
rv
e
rs
(
e
.
g
.
CS
-2,
CS
-3)
t
ha
t
i
n
c
lu
de
pow
e
r
m
a
n
age
m
e
nt
,
c
ool
i
n
g
(
m
a
ss
i
v
e
l
i
qu
id
c
ool
i
n
g
),
a
n
d
G
b
E
/
I
n
fi
n
i
B
a
n
d
l
i
n
k
s
f
or
mult
i
-
no
de
s
ca
l
i
n
g
.
T
ab
l
e
1
be
low
c
omp
a
r
e
s
(
to
t
he
e
xt
e
nt
a
v
ai
l
ab
l
e
)
t
he
ba
s
ic
ha
r
d
w
a
r
e
sp
ec
s
o
f
C
e
r
eb
r
a
s
,
S
a
m
ba
N
ov
a
,
a
n
d
G
roq
chi
ps
:
F
ea
tur
e
C
e
r
eb
r
a
s
W
SE
-3
(
CS
-3
syst
e
m
)
S
a
m
ba
N
ov
a
RD
U
(
D
a
t
a
S
ca
l
e
SN
)
G
roq
LP
U
(
G
roq
R
ack
)
C
hi
p
typ
e
W
afe
r
-
S
ca
l
e
E
n
gi
n
e
(
monol
i
t
hic
w
afe
r
)
R
ec
on
fig
ur
ab
l
e
D
a
t
a
F
low
U
n
i
t
(
chi
p
a
rr
a
y
)
L
a
n
g
u
age
P
ro
ce
ss
i
n
g
U
n
i
t
(
ASIC
)
T
r
a
ns
i
stors
~~4
tr
i
ll
i
on
(
t
hi
r
d
-
ge
n
)
(
[5]
t
i
m
e
.
c
om
)
~
(
un
k
nown
)
—
f
rom
e
xt
e
rn
a
l
i
n
f
o
~10ʼ
s
o
f
bi
ll
i
ons
(7
nm
)~
10
s
o
f
bi
ll
i
ons
(
e
st
i
m
a
t
ed
)
P
ro
ce
ss
N
o
de
3
nm
(
T
SMC
)7
nm
(
T
SMC
)
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
)7
nm
(
T
SMC
)
(
p
e
r
i
nt
e
rv
ie
ws
,
l
ike
ly
)
D
ie
S
i
z
e
~46,225
mm
²
(
f
ull
w
afe
r
)
(
[5]
t
i
m
e
.
c
om
)~430
mm
²
(
a
pprox
;
chi
pl
e
ts
on
b
o
a
r
d
) ~
s
mall
r
ela
t
i
v
e
to
w
afe
r
-
s
cale
C
or
e
s
/
C
omput
e
U
n
i
ts
~850,000
AI
c
or
e
s
~10ʼ
s
o
f
t
h
ous
a
n
d
s
o
f
RD
U
e
l
e
m
e
nts
p
e
r
chi
p
~2600
str
ea
m
i
n
g
c
or
e
s
p
e
r
chi
p
(
G
roq
c
l
ai
m
)
M
e
mory
p
e
r
chi
p
~120
MB
+
SRAM
(
on
-
chi
p
)~8–16
GB HBM
p
e
r
b
o
a
r
d
(
plus
DRAM
o
ff
-
chi
p
)32
GB
+
HBM
p
e
r
LP
U
(
G
roq
R
ack
)
M
e
mory
ba
n
d
w
id
t
h
~20
PB
/
s
(
on
-
w
afe
r
agg
r
ega
t
e
)
[
e
st
.] ~10
T
B
/
s
p
e
r
b
o
a
r
d
(
w
i
t
h
8
HBM
st
ack
s
) ~5
T
B
/
s
p
e
r
LP
U
(
G
roq
c
l
ai
ms
)
K
e
y
H
W
fea
tur
e
s
A
ll
-
CP
U
-
on
-
fi
l
e
(
m
e
mory
+
c
omput
e
)
R
ec
on
fig
ur
ab
l
e
t
i
l
e
s
,
da
t
af
low
n
e
twor
k
S
i
n
g
l
e
-
t
h
r
eaded
,
de
t
e
rm
i
n
i
st
ic
p
i
p
e
l
i
n
e
P
eak
p
e
r
f
orm
a
n
ce
[
N
o
pu
b
l
ic
num
be
r
] [
N
o
pu
b
l
i
s
hed
f
lops
](
G
roq
c
l
ai
ms
~10
x
GP
U
on
i
n
fe
r
e
n
ce
)
(
[13]
t
i
m
e
.
c
om
)
P
ow
e
r
(
p
e
r
no
de
)
~25
k
W
(
CS
-2),
50
k
W
(
CS
-3)
~~10
k
W
(
D
a
t
a
S
ca
l
e
r
ack
)
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
)~~1–5
k
W
p
e
r
G
roq
R
ack
I
nt
e
r
c
onn
ec
t
O
n
-
w
afe
r
m
e
s
h
;
e
xt
e
n
d
s
to
mult
i
-
no
de
IB O
n
-
b
o
a
r
d
fab
r
ic
;
mult
i
-
no
de
E
t
he
rn
e
t
/
IB S
t
a
n
da
r
d
E
t
he
rn
e
t
,
N
V
L
i
n
k
e
t
c
(
us
e
r
s
e
l
ec
t
ab
l
e
)
T
a
r
ge
t
wor
k
lo
ad
s
L
a
r
ge
-
s
ca
l
e
mo
de
l
tr
ai
n
i
n
g
(
GP
T
,
ML
)
T
r
ai
n
i
n
g
&
i
n
fe
r
e
n
ce
f
or
e
nt
e
rpr
i
s
e
/
ML L
ow
-
l
a
t
e
n
c
y
i
n
fe
r
e
n
ce
(
LLM
s
,
v
i
s
i
on
)
N
ot
e
:
M
a
ny
r
a
w
ha
r
d
w
a
r
e
num
be
rs
a
r
e
propr
ie
t
a
ry
or
e
st
i
m
a
t
ed
;
t
he
ab
ov
e
c
om
bi
n
e
s
pu
b
l
ic
di
s
c
losur
e
s
a
n
d
pr
e
ss
c
omm
e
nts
.
F
or
e
x
a
mpl
e
,
G
roq
ha
s
not
pu
b
l
ic
ly
r
e
l
ea
s
ed
e
x
ac
t
tr
a
ns
i
stor
c
ounts
or
die
s
i
z
e
,
b
ut
i
ts
c
l
ai
m
o
f
t
e
n
-
f
ol
d
sp
eed
su
gge
sts
a
high
ly
opt
i
m
i
z
ed
de
s
ig
n
.
2.2
S
a
m
ba
N
ov
a
R
ec
on
fig
ur
ab
l
e
D
a
t
a
F
low
U
n
i
ts
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 7 of 19
S
a
m
ba
N
ov
a
ʼ
s
a
ppro
ach
diffe
rs
b
y
t
i
l
i
n
g
m
a
ny
pro
ce
ssors
/
chi
pl
e
ts
i
nto
a
syst
e
m
.
E
ach
RD
U
chi
p
i
s
i
ts
e
l
f
a
mult
i
-
c
or
e
AI
acce
l
e
r
a
tor
.
A
cc
or
di
n
g
to
a
n
A
n
a
n
d
T
ech
a
n
a
lys
i
s
,
t
he
C
a
r
di
n
a
l
RD
U
chi
p
i
s
“
a
n
a
rr
a
y
o
f
r
ec
on
fig
ur
able
un
i
ts
f
or
da
t
a
,
stor
age
,
or
sw
i
t
chi
n
g
,
opt
imi
z
ed
f
or
da
t
a
fl
ow
”
(
[8]
www
.
a
n
a
n
d
t
ech
.
c
om
).
I
n
ot
he
r
wor
d
s
,
each
RD
U
a
rr
a
y
c
ons
i
sts
o
f
m
a
ny
t
i
ny
c
omput
e
e
l
e
m
e
nts
a
n
d
m
e
mory
t
ha
t
ca
n
be
w
i
r
ed
i
nto
a
da
t
af
low
g
r
a
p
h
b
y
t
he
c
omp
i
l
e
r
.
P
r
ac
t
ica
l
sp
ec
s
o
f
S
a
m
ba
N
ov
a
ʼ
s
chi
ps
/
syst
e
ms
(
f
rom
pu
b
l
ic
sour
ce
s
):
C
hi
pl
e
ts
&
B
o
a
r
d
s
:
T
he
ba
s
ic
c
omput
i
n
g
un
i
t
i
s
a
n
RD
U
chi
p
w
i
t
h
on
-
chi
p
SRAM
a
n
d
lo
ca
l
DRAM
c
ontroll
e
rs
.
M
ult
i
pl
e
RD
U
chi
ps
a
r
e
pl
aced
on
a
b
o
a
r
d
(
w
i
t
h
HBM
m
e
mory
).
T
he
S
N
15
b
o
a
r
d
had
on
e
RD
U
plus
4×
HBM
st
ack
s
(~40
GB HBM
);
t
he
n
e
xt
-
ge
n
S
N
30
b
o
a
r
d
d
ou
b
l
ed
to
2×
RD
U
s
a
n
d
8×
HBM
(80
GB
)
p
e
r
b
o
a
r
d
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
).
C
omp
i
l
e
-
t
i
m
e
R
ec
on
fig
ur
abi
l
i
ty
:
U
nl
ike
st
a
t
ic
GP
U
s
,
S
a
m
ba
N
ov
a
ʼ
s
c
or
e
s
ca
n
cha
n
ge
rol
e
;
e
.
g
.
som
e
c
or
e
s
h
ol
d
w
eigh
ts
,
som
e
d
o
m
a
t
h
,
som
e
m
a
n
age
da
t
a
.
T
he
syst
e
m
rout
e
s
da
t
af
lows
d
yn
a
m
ica
lly
to
a
l
ig
n
w
i
t
h
t
he
n
e
twor
k
ʼ
s
de
m
a
n
d
s
,
r
ed
u
ci
n
g
id
l
e
c
y
c
l
e
s
.
M
e
mory
:
S
a
m
ba
N
ov
a
e
mp
ha
s
i
z
e
s
lo
ca
l
m
e
mory
.
I
n
2022
t
he
y
not
ed
“
O
n
t
he
GP
U
to
da
y
,
you
ca
n
ge
t
80
GB
o
f
HBM
b
ut
w
e
had
1
.
5
T
B
p
e
r
so
cke
t
a
n
d
now
i
t
i
s
d
ou
ble
t
ha
t
.”
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
).
T
hi
s
su
gge
sts
each
mult
i
-
chi
p
D
a
t
a
S
ca
l
e
so
cke
t
(
w
hich
i
n
c
lu
de
s
mult
i
pl
e
RD
U
b
o
a
r
d
s
)
ca
n
s
ee
3
T
B
o
f
m
e
mory
c
o
he
r
e
n
ce
.
P
e
r
f
orm
a
n
ce
:
N
o
i
n
de
p
e
n
de
nt
T
FLOPS
num
be
rs
a
r
e
pu
b
l
i
s
hed
.
H
ow
e
v
e
r
,
S
a
m
ba
N
ov
a
ʼ
s
a
r
chi
t
ec
tur
e
t
a
r
ge
ts
t
h
rou
gh
put
-
i
nt
e
ns
i
v
e
wor
k
lo
ad
s
.
I
n
a
2022
a
rt
ic
l
e
,
S
a
m
ba
N
ov
a
r
e
port
ed
p
a
ss
i
n
g
l
a
n
g
u
age
mo
de
l
i
n
g
be
n
ch
m
a
r
k
s
a
t
i
mpr
e
ss
i
v
e
sp
eed
s
:
e
.
g
.,
S
a
m
ba
N
ov
a
c
l
ai
m
ed
runn
i
n
g
GP
T
-3
(175
B
)
a
t
~32
K
to
ke
ns
/
s
ec
p
e
r
r
ack
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
).
(
B
y
c
ontr
a
st
,
a
n
H
100
GP
U
r
ack
d
o
e
s
~
X
to
ke
ns
/
s
ec
–
i
n
dica
t
i
n
g
S
a
m
ba
N
ov
a
ʼ
s
c
omp
e
t
i
t
i
v
e
st
a
n
ce
,
t
h
ou
gh
di
r
ec
t
c
omp
a
r
a
t
i
v
e
s
a
r
e
v
e
n
d
or
-
ci
t
ed
).
N
e
twor
ki
n
g
:
D
a
t
a
S
ca
l
e
r
ack
s
support
s
ca
l
i
n
g
v
ia
E
t
he
rn
e
t
or
I
n
fi
n
i
B
a
n
d
.
T
he
p
h
ys
ica
l
de
s
ig
n
st
ack
s
b
o
a
r
d
s
i
n
cabi
n
e
ts
;
i
nt
e
rn
a
l
fea
tur
e
s
(
pow
e
r
,
c
ool
i
n
g
)
m
a
n
age
t
he
l
a
r
ge
m
e
mory
/
c
omput
e
lo
ad
.
S
a
m
ba
N
ov
a
s
a
ys
each
qu
a
rt
e
r
-
r
ack
i
s
f
ully
i
nt
eg
r
a
t
ed
a
n
d
ca
n
be
r
e
mot
e
ly
m
a
n
aged
(
[27]
www
.
a
n
a
n
d
t
ech
.
c
om
).
S
o
f
tw
a
r
e
:
S
a
m
ba
N
ov
a
prov
ide
s
a
f
ull
st
ack
(
c
omp
i
l
e
r
,
runt
i
m
e
,
l
ib
r
a
r
ie
s
)
ca
ll
ed
S
a
m
ba
F
low
.
U
s
e
rs
ca
n
i
mport
mo
de
ls
f
rom
T
e
nsor
F
low
/
P
y
T
or
ch
;
S
a
m
ba
F
low
p
a
rt
i
t
i
ons
t
he
m
a
utom
a
t
ica
lly
i
nto
da
t
af
low
g
r
a
p
h
s
f
or
t
he
RD
U
s
(
[28]
www
.
a
n
a
n
d
t
ech
.
c
om
).
T
he
y
a
lso
op
e
n
-
sour
ced
a
f
r
a
m
e
wor
k
(
S
a
m
ba
-1
mo
de
l
)
a
n
d
c
ontr
ib
ut
ed
to
K
agg
l
e
cha
ll
e
n
ge
s
.
2.3
G
roq
L
a
n
g
u
age
P
ro
ce
ss
i
n
g
U
n
i
t
G
roq
ʼ
s
LP
U
i
s
a
n
ASIC
t
ai
lor
ed
f
or
i
n
fe
r
e
n
ce
.
S
om
e
ke
y
a
r
chi
t
ec
tur
a
l
po
i
nts
:
S
i
n
g
l
e
T
h
r
ead
o
f
C
ontrol
:
U
nl
ike
GP
U
s
,
w
hich
ha
v
e
m
a
ny
c
or
e
s
a
n
d
t
h
r
ead
c
ont
e
xts
,
t
he
LP
U
i
s
a
V
LI
W
-
l
ike
p
i
p
e
l
i
n
e
t
ha
t
pro
ce
ss
e
s
on
e
i
nstru
c
t
i
on
str
ea
m
a
t
a
t
i
m
e
(
c
o
a
rs
e
-
g
r
ai
n
ed
).
T
hi
s
y
ie
l
d
s
de
t
e
rm
i
n
i
sm
.
F
ully
U
nroll
ed
C
omput
e
:
G
roq
st
a
t
e
s
i
ts
LP
U
s
l
ack
m
ic
ro
c
o
de
“
ke
rn
e
ls
”
–
mo
de
l
l
a
y
e
rs
a
r
e
c
omp
i
l
ed
i
nto
str
aigh
t
-
l
i
n
e
ha
r
d
w
a
r
e
s
e
qu
e
n
ce
s
(
[12]
www
.
ee
t
i
m
e
s
.
c
om
).
T
hi
s
e
l
i
m
i
n
a
t
e
s
ov
e
r
head
f
or
i
ssu
i
n
g
GP
U
ke
rn
e
ls
or
g
o
i
n
g
outs
ide
p
i
p
e
l
i
n
e
.
D
a
t
a
S
tr
ea
m
i
n
g
:
T
e
nsors
f
low
t
h
rou
gh
f
rom
on
e
lo
gic
b
lo
ck
to
t
he
n
e
xt
e
v
e
ry
c
y
c
l
e
.
AL
U
s
a
n
d
MAC
a
rr
a
ys
a
r
e
fed
e
v
e
ry
c
lo
ck
b
y
m
e
mory
cha
nn
e
ls
h
ol
di
n
g
ac
t
i
v
a
t
i
ons
/
w
eigh
ts
.
M
e
mory
&
O
ff
-
C
hi
p
I
/
O
:
E
ach
LP
U
c
onn
ec
ts
to
HBM
(
st
acked
DRAM
)
f
or
mo
de
l
p
a
r
a
m
e
t
e
rs
a
n
d
ac
t
i
v
a
t
i
on
stor
age
.
E
x
ac
t
m
e
mory
s
i
z
e
s
a
r
e
propr
ie
t
a
ry
,
b
ut
typ
ica
l
LP
U
ca
r
d
s
r
i
v
a
l
GP
U
b
o
a
r
d
s
.
P
e
r
f
orm
a
n
ce
:
G
roq
ci
t
e
s
be
n
ch
m
a
r
k
r
e
sults
de
monstr
a
t
i
n
g
v
e
ry
high
t
h
rou
gh
put
.
F
or
e
x
a
mpl
e
,
EE
T
ime
s
r
e
port
ed
G
roq
did
LLM
i
n
fe
r
e
n
ce
be
n
ch
m
a
r
k
s
3–4×
fa
st
e
r
t
ha
n
c
omp
a
r
ab
l
e
GP
U
s
(
[29]
www
.
ee
t
i
m
e
s
.
c
om
).
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 8 of 19
CEO R
oss
c
l
ai
ms
10×
GP
U
p
e
r
f
orm
a
n
ce
a
n
d
10×
c
ost
r
ed
u
c
t
i
on
on
i
n
fe
r
e
n
ce
(
[13]
t
i
m
e
.
c
om
).
H
ow
e
v
e
r
,
i
t
ʼ
s
not
ed
G
roq
ʼ
s
chi
ps
a
r
e
c
urr
e
ntly
f
o
c
us
ed
on
i
n
fe
r
e
n
ce
,
not
tr
ai
n
i
n
g
.
S
yst
e
ms
:
A
G
roq
C
a
r
d
(
PCI
e
ca
r
d
)
h
ol
d
s
a
s
i
n
g
l
e
LP
U
a
n
d
m
e
mory
,
a
n
d
mult
i
pl
e
ca
r
d
s
ca
n
be
l
i
n
ked
i
nto
a
G
roq
R
ack
.
T
he
y
a
lso
run
G
roq
C
lou
d
s
e
rv
ice
s
;
i
n
2024
t
he
y
r
e
port
ed
ov
e
r
70,000
de
v
e
lop
e
rs
us
i
n
g
G
roq
C
lou
d
(
[30]
g
roq
.
c
om
).
G
roq
e
mp
ha
s
i
z
e
s
ea
s
e
o
f
de
ploym
e
nt
:
f
or
i
nst
a
n
ce
,
T
om
ʼ
s
H
a
r
d
w
a
r
e
not
ed
G
roq
ai
ms
f
or
qu
ick
de
l
i
v
e
ry
to
c
ustom
e
rs
a
n
d
a
vo
id
s
e
xot
ic
supply
chai
n
p
a
rts
(
[31]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
).
T
ab
l
e
2
l
i
sts
som
e
c
omp
a
r
a
t
i
v
e
chi
p
m
e
tr
ic
s
(
w
he
r
e
a
v
ai
l
ab
l
e
):
S
p
ecifica
t
i
on
C
e
r
eb
r
a
s
W
SE
-3
S
a
m
ba
N
ov
a
RD
U
B
o
a
r
d
s
G
roq
LP
U
(
G
roq
R
ack
)
P
ro
ce
ss
T
ech
nolo
g
y
3
nm
(
T
SMC
)7
nm
(
T
SMC
)
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
)7
nm
(
T
SMC
)
T
r
a
ns
i
stor
C
ount
~~4
T
(
c
l
ai
m
ed
)
(
[5]
t
i
m
e
.
c
om
)~~
bi
ll
i
ons
(
high
,
un
k
nown
e
x
ac
t
)~~
t
e
ns
o
f
bi
ll
i
ons
(
G
roq
ʼ
s
own
r
e
port
)
D
ie
A
r
ea
~46,000
mm
²
(
w
afe
r
) ~430
mm
²
p
e
r
chi
p
(
mult
i
-
chi
p
syst
e
m
)~1,100
mm
²
p
e
r
LP
U
(
e
st
i
m
a
t
ed
)
O
n
-
chi
p
SRAM
~120+
MB
(
i
nt
eg
r
a
t
ed
) ~30–50
MB
p
e
r
RD
U
chi
p
<1
MB
(
most
m
e
mory
i
s
o
ff
-
chi
p
)
O
ff
-
chi
p
M
e
mory
N
on
e
(
m
a
ss
i
v
e
on
-
chi
p
)
U
p
to
80
GB HBM
p
e
r
b
o
a
r
d
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
)
32+
GB HBM
p
e
r
LP
U
(
b
o
a
r
d
)
(
a
pprox
.)
M
e
mory
B
a
n
d
w
id
t
h
(
tot
a
l
)~20
PB
/
s
(
agg
r
ega
t
e
) ~~10
T
B
/
s
(8
HBM
p
e
r
SN
30
b
o
a
r
d
) ~~5
PB
/
s
(
p
e
r
G
roq
C
a
r
d
)
C
omput
e
D
e
ns
i
ty
(
FLOP
s
)
~
2
e
x
a
FLOP
/
s
(
FP
16
e
st
)~
un
k
nown
(
da
t
af
low
ops
p
e
r
s
ec
on
d
)~0.5
e
x
a
FLOP
/
s
(
FP
16
e
st
p
e
r
4-
chi
p
c
on
fig
?)
S
upport
ed
P
r
eci
s
i
on
FP
16,
FP
8,
i
nt
8,
e
t
c f
p
32/16/8,
m
i
x
ed
pr
ec
FP
16,
BF
16,
i
nt
8,
e
t
c
N
ot
e
s
:
T
he
num
be
rs
ab
ov
e
a
r
e
c
omp
i
l
ed
f
rom
c
omp
a
ny
di
s
c
losur
e
s
a
n
d
pr
e
ss
a
n
a
lys
e
s
.
C
e
r
eb
r
a
s
ʼ
s
num
be
rs
a
r
e
t
he
or
e
t
ica
l
(
ex
a
FLOP
/
s
a
pprox
i
m
a
t
ed
f
rom
rumor
ed
4
T
tr
a
ns
i
stors
a
n
d
v
ec
tor
w
id
t
h
).
G
roq
ʼ
s
e
x
ac
t
num
be
rs
a
r
e
propr
ie
t
a
ry
;
r
e
port
ed
0.5
e
x
af
lop
i
s
f
or
f
our
LP
U
chi
p
p
a
rs
e
.
S
a
m
ba
N
ov
a
ʼ
s
da
t
a
i
s
mostly
t
ake
n
f
rom
pr
e
ss
(
e
.
g
.
80
GB HBM
p
e
r
b
o
a
r
d
)
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
).
2.4
N
e
twor
k
s
a
n
d
S
ca
l
abi
l
i
ty
A
ll
t
h
r
ee
v
e
n
d
ors
support
s
ca
l
i
n
g
be
yon
d
a
s
i
n
g
l
e
chi
p
,
b
ut
i
n
diffe
r
e
nt
w
a
ys
:
C
e
r
eb
r
a
s
:
D
u
e
to
t
he
w
afe
r
-
s
ca
l
e
n
a
tur
e
,
a
s
i
n
g
l
e
W
SE
i
s
a
lr
ead
y
e
normous
,
b
ut
f
or
mor
e
ca
p
aci
ty
mult
i
pl
e
CS
-2/
CS
-3
s
e
rv
e
rs
ca
n
be
c
onn
ec
t
ed
v
ia
high
-
sp
eed
fab
r
ic
(
e
.
g
.
I
n
fi
n
i
B
a
n
d
)
to
f
orm
a
c
lust
e
r
.
F
or
i
nst
a
n
ce
,
t
he
C
on
d
or
G
a
l
a
xy
sup
e
r
c
omput
e
r
us
e
s
m
a
ny
W
SE
-3
no
de
s
l
i
n
ked
b
y
IP
ov
e
r
I
n
fi
n
i
B
a
n
d
.
T
he
DARPA
“
F
us
e
”
pro
jec
t
w
i
ll
us
e
p
h
oton
ic
i
nt
e
r
c
onn
ec
t
(
R
a
novus
opt
ica
l
sw
i
t
che
s
)
to
t
ie
mult
i
pl
e
w
afe
r
-
s
ca
l
e
no
de
s
i
nto
a
c
o
he
r
e
nt
w
h
ol
e
(
[7]
www
.
r
e
ut
e
rs
.
c
om
),
ai
m
i
n
g
f
or
or
de
rs
-
o
f
-
m
ag
n
i
tu
de
sp
eed
up
ov
e
r
GP
U
c
lust
e
rs
.
S
a
m
ba
N
ov
a
:
D
a
t
a
S
ca
l
e
r
ack
s
ca
n
be
l
i
n
ked
b
y
st
a
n
da
r
d
E
t
he
rn
e
t
or
I
n
fi
n
i
B
a
n
d
.
S
a
m
ba
N
ov
a
m
e
nt
i
ons
“
s
ca
l
e
-
out
to
mult
i
pl
e
qu
a
rt
e
r
-
r
ack
de
ploym
e
nts
”
(
[32]
www
.
a
n
a
n
d
t
ech
.
c
om
).
W
i
t
hi
n
a
r
ack
,
t
he
b
o
a
r
d
s
ha
v
e
C
X
L
a
n
d
N
V
L
i
n
k
f
or
c
ross
-
CP
U
or
c
ross
-
b
o
a
r
d
c
o
he
r
e
n
c
y
if
n
eeded
.
T
he
a
r
chi
t
ec
tur
e
f
o
c
us
e
s
mor
e
on
m
a
x
i
m
i
z
i
n
g
w
ha
t
on
e
r
ack
ca
n
d
o
(
m
a
ss
i
v
e
m
e
mory
)
r
a
t
he
r
t
ha
n
e
xtr
e
m
e
ly
l
a
r
ge
c
lust
e
rs
,
a
lt
h
ou
gh
you
ca
n
j
o
i
n
r
ack
s
i
n
a
da
t
ace
nt
e
r
.
G
roq
:
G
roq
syst
e
ms
c
onn
ec
t
v
ia
PCI
e
or
E
t
he
rn
e
t
be
tw
ee
n
G
roq
C
a
r
d
s
.
G
roq
a
lso
p
a
rtn
e
rs
w
i
t
h
n
e
twor
ki
n
g
c
omp
a
n
ie
s
(
C
i
s
c
o
)
f
or
l
a
r
ge
-
s
ca
l
e
i
nt
e
r
c
onn
ec
ts
.
G
roq
touts
i
nt
eg
r
a
t
i
on
ea
s
e
–
i
t
ca
n
s
i
t
i
n
a
st
a
n
da
r
d
s
e
rv
e
r
r
ack
a
mon
g
GP
U
s
or
a
utonomously
–
a
n
d
a
G
roq
R
ack
ca
n
t
he
or
e
t
ica
lly
r
e
pl
ace
a
GP
U
s
e
rv
e
r
f
or
i
n
fe
r
e
n
ce
.
T
he
y
ha
v
e
not
e
mp
ha
s
i
z
ed
syn
ch
ron
i
z
ed
tr
ai
n
i
n
g
a
t
mult
i
-
r
ack
s
ca
l
e
(
s
i
n
ce
G
roq
ʼ
s
f
o
c
us
i
s
i
n
fe
r
e
n
ce
,
w
he
r
e
mo
de
l
r
e
pl
ica
t
i
on
i
s
str
aigh
t
f
orw
a
r
d
).
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 9 of 19
3.
S
o
f
tw
a
r
e
a
n
d
P
ro
g
r
a
mm
i
n
g
E
c
osyst
e
ms
T
o
l
e
v
e
r
age
t
he
s
e
sp
ecia
l
i
z
ed
chi
ps
,
each
c
omp
a
ny
prov
ide
s
i
ts
own
so
f
tw
a
r
e
st
ack
a
n
d
tools
:
C
e
r
eb
r
a
s
:
O
ffe
rs
t
he
C
e
r
eb
r
a
s
S
o
f
tw
a
r
e
P
l
a
t
f
orm
(
CSP
)
,
i
n
c
lu
di
n
g
a
c
omp
i
l
e
r
t
ha
t
m
a
ps
P
y
T
or
ch
/
T
e
nsor
F
low
g
r
a
p
h
s
onto
t
he
W
SE
c
or
e
s
.
U
s
e
rs
wr
i
t
e
mo
de
ls
i
n
st
a
n
da
r
d
f
r
a
m
e
wor
k
s
;
t
he
CSP
t
ake
s
ca
r
e
o
f
spl
i
tt
i
n
g
l
a
y
e
rs
ac
ross
c
or
e
s
a
n
d
s
ched
ul
i
n
g
c
ommun
ica
t
i
on
.
C
e
r
eb
r
a
s
a
lso
supports
P
yt
h
on
ke
rn
e
ls
a
n
d
c
ustom
l
a
y
e
rs
.
T
he
c
omp
a
ny
e
mp
ha
s
i
z
e
s
support
f
or
l
a
r
ge
diff
us
i
on
a
n
d
v
i
s
i
on
mo
de
ls
.
I
n
2025,
C
e
r
eb
r
a
s
a
nnoun
ced
C
e
r
eb
r
a
X
,
a
n
e
m
e
r
gi
n
g
pro
g
r
a
mm
i
n
g
mo
de
l
(
de
t
ai
ls
sp
a
rs
e
)
a
n
d
p
a
rtn
e
rs
hi
ps
to
i
nt
eg
r
a
t
e
w
i
t
h
HPC
s
ched
ul
e
rs
l
ike
S
lurm
.
S
a
m
ba
N
ov
a
:
P
rov
ide
s
S
a
m
ba
F
low
a
n
d
S
a
m
ba
S
tu
di
o
.
S
a
m
ba
F
low
i
n
ge
sts
high
-
l
e
v
e
l
mo
de
l
defi
n
i
t
i
ons
a
n
d
c
omp
i
l
e
s
t
he
m
i
nto
RD
U
c
on
fig
ur
a
t
i
on
s
e
qu
e
n
ce
s
.
S
a
m
ba
F
low
a
utom
a
t
e
s
t
i
l
i
n
g
,
w
eigh
ts
/
da
t
a
p
a
rt
i
t
i
on
i
n
g
,
a
n
d
efficie
nt
f
low
c
ontrol
.
S
a
m
ba
S
tu
di
o
o
ffe
rs
a
v
i
su
a
l
de
v
e
lopm
e
nt
e
nv
i
ronm
e
nt
.
S
a
m
ba
N
ov
a
a
lso
c
ontr
ib
ut
e
s
to
op
e
n
f
r
a
m
e
wor
k
s
–
f
or
e
x
a
mpl
e
,
S
a
m
ba
-1
w
a
s
r
e
l
ea
s
ed
on
G
i
t
H
u
b
,
a
n
d
S
a
m
ba
L
i
n
g
o
i
s
i
ts
op
e
n
-
sour
ce
qu
e
st
i
on
-
a
nsw
e
r
i
n
g
p
i
p
e
l
i
n
e
.
S
a
m
ba
N
ov
a
c
l
ai
ms
e
n
d
-
to
-
e
n
d
opt
i
m
i
z
a
t
i
on
to
hide
da
t
a
mov
e
m
e
nt
a
n
d
m
e
mory
l
a
t
e
n
c
y
f
rom
us
e
rs
.
G
roq
:
G
roq
prov
ide
s
a
C
++
a
n
d
P
yt
h
on
-
ba
s
ed
SDK
.
I
ts
G
roq
API
l
e
ts
de
v
e
lop
e
rs
uplo
ad
mo
de
ls
(
v
ia
ONN
X
)
to
t
he
LP
U
a
n
d
run
i
n
fe
r
e
n
ce
s
.
G
roq
ʼ
s
c
omp
i
l
e
rs
unroll
t
he
mo
de
l
i
nto
t
he
LP
U
ʼ
s
i
nstru
c
t
i
on
str
ea
m
.
G
roq
C
a
nv
a
s
i
s
a
w
eb
i
nt
e
r
face
f
or
mon
i
tor
i
n
g
j
o
b
s
.
G
roq
e
mp
ha
s
i
z
e
s
t
ha
t
no
mo
de
l
cha
n
ge
s
a
r
e
n
eeded
f
rom
us
e
r
c
o
de
be
s
ide
s
t
he
t
a
r
ge
t
pl
a
t
f
orm
.
A
ddi
t
i
on
a
lly
,
G
roq
ʼ
s
G
i
t
H
u
b
h
osts
tools
f
or
qu
a
nt
i
z
a
t
i
on
a
n
d
fi
n
e
-
tun
i
n
g
to
8-
bi
t
or
4-
bi
t
f
or
i
n
fe
r
e
n
ce
p
e
r
f
orm
a
n
ce
.
A
ll
t
h
r
ee
st
ack
s
t
ie
i
nto
popul
a
r
f
r
a
m
e
wor
k
s
.
F
or
i
nst
a
n
ce
,
S
a
m
ba
N
ov
a
c
l
ai
ms
f
ull
c
omp
a
t
ibi
l
i
ty
w
i
t
h
H
u
ggi
n
g
F
ace
p
i
p
e
l
i
n
e
s
,
C
e
r
eb
r
a
s
ha
s
sp
ecia
l
API
s
b
ut
supports
P
y
T
or
ch
out
o
f
t
he
b
ox
,
a
n
d
G
roq
ha
s
p
a
rtn
e
r
ed
w
i
t
h
H
u
ggi
n
g
F
ace
f
or
opt
i
m
i
z
ed
i
n
fe
r
e
n
ce
.
I
n
pr
ac
t
ice
,
t
h
ou
gh
,
a
ppl
ica
t
i
on
c
o
de
o
f
t
e
n
n
eed
s
mo
difica
t
i
on
or
ca
r
ef
ul
m
a
n
age
m
e
nt
o
f
p
a
r
a
ll
e
l
i
sm
to
e
xplo
i
t
each
a
r
chi
t
ec
tur
e
f
ully
.
4.
P
e
r
f
orm
a
n
ce
a
n
d
B
e
n
ch
m
a
r
k
s
A
c
r
i
t
ica
l
p
a
rt
o
f
c
omp
a
r
i
son
i
s
h
ow
t
he
s
e
syst
e
ms
p
e
r
f
orm
on
r
ea
l
AI
wor
k
lo
ad
s
.
H
ow
e
v
e
r
,
pu
b
l
ic
be
n
ch
m
a
r
k
da
t
a
i
s
l
i
m
i
t
ed
a
n
d
o
f
t
e
n
v
e
n
d
or
-
suppl
ied
.
W
e
c
omp
i
l
e
w
ha
t
i
s
a
v
ai
l
ab
l
e
:
M
o
de
l
T
r
ai
n
i
n
g
T
h
rou
gh
put
(
e
.
g
.
LLM
tr
ai
n
i
n
g
):
C
e
r
eb
r
a
s
r
e
ports
t
ha
t
a
s
i
n
g
l
e
CS
-2
syst
e
m
tr
ai
ns
GP
T
-3
(175
B
p
a
r
a
m
e
t
e
rs
)
i
n
24
h
ours
(
vs
~
w
eek
s
on
1,024
GP
U
s
)
i
n
som
e
i
nt
e
rn
a
l
t
e
sts
,
i
mply
i
n
g
or
de
rs
-
o
f
-
m
ag
n
i
tu
de
t
h
rou
gh
put
(
[5]
t
i
m
e
.
c
om
).
I
n
a
NASA
p
a
rtn
e
rs
hi
p
,
C
e
r
eb
r
a
s
c
l
ai
m
ed
t
hei
r
syst
e
m
(
CS
-2)
achie
v
ed
210×
sp
eed
up
ov
e
r
N
V
IDIA H
100
on
a
su
b
sur
face
s
i
mul
a
t
i
on
mo
de
l
(
[26]
www
.
ce
r
eb
r
a
s
.
n
e
t
)
–
not
ab
ly
,
not
a
st
a
n
da
r
d
ML
t
a
s
k
.
S
a
m
ba
N
ov
a
,
f
or
i
ts
p
a
rt
,
high
l
igh
t
ed
t
ha
t
i
ts
D
a
t
a
S
ca
l
e
SN
30
pl
a
t
f
orm
(
w
i
t
h
4
b
o
a
r
d
s
p
e
r
so
cke
t
)
ca
n
tr
ai
n
a
1.3-
tr
i
ll
i
on
-
p
a
r
a
m
e
t
e
r
mo
de
l
b
y
spl
i
tt
i
n
g
i
t
i
nto
54
sm
a
ll
e
r
“
e
xp
e
rts
”
(
[33]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
).
A
cc
or
di
n
g
to
N
e
xt
P
l
a
t
f
orm
,
S
a
m
ba
N
ov
a
w
a
s
pus
hi
n
g
f
or
tr
ai
n
i
n
g
tr
i
ll
i
on
-
p
a
r
a
m
e
t
e
r
mo
de
ls
b
y
effec
t
i
v
e
ly
l
a
r
ge
-
s
ca
l
e
mo
de
l
p
a
r
a
ll
e
l
i
sm
,
b
ut
w
e
l
ack
op
e
n
num
be
rs
.
G
roq
c
urr
e
ntly
pos
i
t
i
ons
i
ts
e
l
f
f
or
i
n
fe
r
e
n
ce
;
i
t
ha
s
not
pu
b
l
i
s
hed
tr
ai
n
i
n
g
be
n
ch
m
a
r
k
s
.
I
n
fe
r
e
n
ce
S
p
eed
(
t
h
rou
gh
put
a
n
d
l
a
t
e
n
c
y
):
H
e
r
e
G
roq
s
hi
n
e
s
i
n
c
l
ai
ms
.
G
roq
LP
U
s
ha
v
e
bee
n
i
n
de
p
e
n
de
ntly
o
b
s
e
rv
ed
i
n
2024
EE
T
i
m
e
s
to
run
LLM
i
n
fe
r
e
n
ce
a
t
3-4×
t
he
sp
eed
o
f
GP
U
ha
r
d
w
a
r
e
on
e
qu
i
v
a
l
e
nt
mo
de
ls
(
[34]
www
.
ee
t
i
m
e
s
.
c
om
).
T
he
CEO
st
a
t
ed
a
M
e
t
a
cha
t
b
ot
de
mo
r
a
n
“
mu
ch
fa
st
e
r
”
on
G
roq
(
[13]
t
i
m
e
.
c
om
).
G
roq
a
lso
e
mp
ha
s
i
z
e
s
t
ha
t
l
a
t
e
n
c
y
i
s
high
ly
pr
edic
t
ab
l
e
(
su
b
-
m
i
ll
i
s
ec
on
d
f
or
tr
a
ns
f
orm
e
r
l
a
y
e
rs
),
a
ke
y
ad
v
a
nt
age
f
or
r
ea
l
-
t
i
m
e
syst
e
ms
(
[20]
t
i
m
e
.
c
om
).
S
a
m
ba
N
ov
a
ʼ
s
b
o
a
r
d
s
ha
v
e
bee
n
t
e
st
ed
on
i
n
fe
r
e
n
ce
too
;
on
e
r
e
port
b
y
S
a
m
ba
N
ov
a
not
ed
runn
i
n
g
L
l
a
m
a
2
70
B
a
t
132
to
ke
ns
/
s
ec
p
e
r
r
ack
on
f
ull
pr
eci
s
i
on
(
[35]
s
a
m
ba
nov
a
.
ai
).
F
or
pu
b
l
ic
g
u
ida
n
ce
,
a
n
A
W
S C
lou
d
B
e
n
ch
m
a
r
k
(
pu
b
l
ic
b
lo
g
)
f
oun
d
t
ha
t
a
S
a
m
ba
N
ov
a
r
ack
c
oul
d
achie
v
e
>1,000
to
ke
ns
/
s
ec
w
i
t
h
LL
a
MA
3
540
B
–
rou
gh
ly
d
ou
b
l
e
t
he
r
a
t
e
o
f
8
N
v
idia
H
100
GP
U
s
i
n
s
i
m
i
l
a
r
c
on
di
t
i
on
.
C
e
r
eb
r
a
s
a
lso
supports
i
n
fe
r
e
n
ce
;
on
e
pu
b
l
i
s
hed
us
e
-
ca
s
e
(
LLNL
)
us
ed
CS
-2
f
or
s
cie
nt
ific
da
t
a
i
n
fe
r
e
n
ce
,
b
ut
agai
n
di
r
ec
t
t
h
rou
gh
put
num
be
rs
a
r
e
s
ca
nt
.
E
ss
e
nt
ia
lly
,
outs
ide
G
roq
ʼ
s
c
l
ai
ms
,
t
he
r
e
i
s
no
i
mp
a
rt
ia
l
,
w
ide
ly
-
acce
pt
ed
be
n
ch
m
a
r
k
c
omp
a
r
i
n
g
t
he
s
e
v
e
n
d
ors
on
c
ommon
t
a
s
k
s
(
l
ike
MLP
e
r
f
),
d
u
e
to
diffe
r
i
n
g
ha
r
d
w
a
r
e
st
ack
s
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 10 of 19
E
fficie
n
c
y
a
n
d
S
ca
l
abi
l
i
ty
:
R
e
ports
su
gge
st
G
roq
ʼ
s
LP
U
s
c
onsum
e
ab
out
on
e
-
t
hi
r
d
(
⅓
)
a
s
mu
ch
pow
e
r
a
s
a
n
e
qu
i
v
a
l
e
nt
GP
U
pl
a
t
f
orm
(
[21]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
)
w
he
n
s
ca
l
ed
f
or
i
n
fe
r
e
n
ce
.
C
e
r
eb
r
a
s
ʼ
s
CS
-2
w
a
s
s
h
own
to
d
r
a
w
~15-25
k
W
un
de
r
f
ull
lo
ad
p
e
r
syst
e
m
–
fa
r
ab
ov
e
a
s
i
n
g
l
e
GP
U
b
ut
fa
r
mor
e
c
omput
e
.
S
a
m
ba
N
ov
a
ʼ
s
ol
de
r
r
ack
s
w
e
r
e
~10
k
W
/
qu
a
rt
-
r
ack
(
[36]
www
.
a
n
a
n
d
t
ech
.
c
om
);
t
he
l
a
r
ge
r
SN
30
r
ack
s
l
ike
ly
d
r
a
w
20–40
k
W
.
I
n
t
e
rms
o
f
p
e
r
f
orm
a
n
ce
p
e
r
w
a
tt
,
i
nt
e
rn
a
l
m
e
tr
ic
s
(
unpu
b
l
i
s
hed
)
c
l
ai
m
a
ll
t
h
r
ee
a
r
e
c
omp
e
t
i
t
i
v
e
or
sup
e
r
i
or
to
GP
U
a
rr
a
ys
f
or
t
hei
r
i
nt
e
n
ded
wor
k
lo
ad
c
l
a
ss
.
F
ac
tor
i
n
g
r
ack
-
l
e
v
e
l
p
e
r
f
orm
a
n
ce
,
a
ci
t
ed
c
l
ai
m
i
s
t
ha
t
a
G
roq
32-
LP
U
r
ack
,
a
C
e
r
eb
r
a
s
PBS
,
or
a
S
a
m
ba
N
ov
a
D
a
t
a
S
ca
l
e
r
ack
c
oul
d
each
pro
ce
ss
t
e
ns
o
f
m
i
ll
i
ons
o
f
i
n
fe
r
e
n
ce
s
p
e
r
s
ec
on
d
on
a
mo
de
r
a
t
e
LLM
,
a
t
low
e
r
tot
a
l
e
n
e
r
g
y
t
ha
n
a
c
omp
a
r
ab
l
e
N
V
IDIA
c
lust
e
r
.
U
s
e
-
C
a
s
e
B
e
n
ch
m
a
r
k
s
:
AI
R
e
s
ea
r
ch
/
S
im
u
la
t
i
on
:
LLNL
r
a
n
LAMMPS
f
lu
id
s
i
mul
a
t
i
ons
on
C
e
r
eb
r
a
s
CS
-2,
ci
t
i
n
g
a
100–200×
sp
eed
i
mprov
e
m
e
nt
ov
e
r
GP
U
c
lust
e
rs
(
C
e
r
eb
r
a
s
PR
)
(
[7]
www
.
r
e
ut
e
rs
.
c
om
).
D
efe
ns
e
&
HPC
:
DARPA
ai
ms
f
or
su
b
s
ec
on
d
s
i
n
fe
r
e
n
ce
o
f
f
ull
-
fide
l
i
ty
ba
ttl
e
s
i
mul
a
t
i
ons
;
t
he
y
s
e
l
ec
t
ed
C
e
r
eb
r
a
s
+
R
a
novus
f
or
150×
acce
l
e
r
a
t
i
on
o
f
da
t
a
mov
e
m
e
nt
(
[37]
www
.
r
e
ut
e
rs
.
c
om
).
E
nt
e
rpr
i
s
e
NLP
:
S
a
m
ba
N
ov
a
quot
e
s
r
ea
l
-
worl
d
c
ustom
e
r
r
e
sults
–
e
.
g
.
on
e
fi
n
a
n
ce
fi
rm
s
a
w
5×
fa
st
e
r
mo
de
l
de
v
e
lopm
e
nt
v
e
rsus
t
hei
r
T
e
sl
a
/
V
100
GP
U
fa
rm
.
DL
i
n
fe
r
e
n
ce
:
P
u
b
l
ic
G
roq
c
ustom
e
r
(
unn
a
m
ed
)
achie
v
ed
3.3×
fa
st
e
r
i
n
fe
r
e
n
ce
t
h
rou
gh
put
on
GP
T
-
N
e
o
2.7
B
w
he
n
mov
i
n
g
f
rom
GP
U
to
G
roq
C
lou
d
(
[34]
www
.
ee
t
i
m
e
s
.
c
om
).
T
o
summ
a
r
i
z
e
p
e
r
f
orm
a
n
ce
:
each
pl
a
t
f
orm
e
x
ce
ls
a
t
diffe
r
e
nt
m
e
tr
ic
s
.
C
e
r
eb
r
a
s
ʼ
s
w
afe
r
-
s
ca
l
e
syst
e
ms
m
a
x
i
m
i
z
e
r
a
w
t
h
rou
gh
put
f
or
v
a
st
mo
de
ls
,
a
t
t
he
c
ost
o
f
l
a
r
ge
pow
e
r
a
n
d
sp
ace
;
S
a
m
ba
N
ov
a
ʼ
s
D
a
t
a
S
ca
l
e
ba
l
a
n
ce
s
t
h
rou
gh
put
a
n
d
m
e
mory
ca
p
aci
ty
(
e
x
ce
ll
i
n
g
a
t
v
e
ry
l
a
r
ge
mo
de
ls
t
ha
t
n
eed
>100
s
o
f
GB
o
f
m
e
mory
each
);
w
he
r
ea
s
G
roq
ʼ
s
LP
U
t
a
r
ge
ts
ultr
a
-
fa
st
low
-
l
a
t
e
n
c
y
i
n
fe
r
e
n
ce
,
r
e
qu
i
r
i
n
g
l
e
ss
m
e
mory
b
ut
de
l
i
v
e
r
i
n
g
qu
icke
r
outputs
.
D
i
r
ec
t
a
ppl
e
s
-
to
-
a
ppl
e
s
num
be
rs
a
r
e
s
ca
r
ce
,
b
ut
i
n
d
ustry
s
e
nt
i
m
e
nt
(
a
n
d
i
nv
e
stor
be
ts
(
[23]
www
.
r
e
ut
e
rs
.
c
om
))
i
n
dica
t
e
s
a
ll
t
h
r
ee
a
r
e
“
wort
h
y
cha
ll
e
n
ge
rs
”
to
GP
U
st
a
tus
quo
f
or
ce
rt
ai
n
wor
k
lo
ad
s
.
5.
C
a
s
e
S
tu
die
s
a
n
d
D
e
ploym
e
nts
T
hi
s
s
ec
t
i
on
pr
e
s
e
nts
i
llustr
a
t
i
v
e
r
ea
l
-
worl
d
us
e
s
o
f
each
t
ech
nolo
g
y
,
high
l
igh
t
i
n
g
h
ow
c
ustom
e
rs
de
ploy
t
he
s
e
syst
e
ms
.
5.1
C
e
r
b
r
a
s
i
n
N
a
t
i
on
a
l
D
efe
ns
e
a
n
d
U
AE AI H
u
b
DARPA MAPLE
(2025):
C
e
r
eb
r
a
s
a
n
d
R
a
novus
w
e
r
e
a
w
a
r
ded
a
$
45
M DARPA
c
ontr
ac
t
to
b
u
i
l
d
a
n
AI
t
e
st
bed
f
or
mult
i
-
d
om
ai
n
ba
ttl
efie
l
d
s
i
mul
a
t
i
on
(
[37]
www
.
r
e
ut
e
rs
.
c
om
).
T
he
pro
jec
t
i
nt
eg
r
a
t
e
s
C
e
r
eb
r
a
s
ʼ
s
W
SE
-3
chi
ps
w
i
t
h
R
a
novus
opt
ica
l
i
nt
e
r
c
onn
ec
ts
to
tr
a
ns
fe
r
m
a
ss
i
v
e
a
mounts
o
f
s
e
nsor
a
n
d
e
nv
i
ronm
e
nt
da
t
a
.
T
he
g
o
a
l
i
s
r
ea
l
-
t
i
m
e
AI
-
d
r
i
v
e
n
pl
a
nn
i
n
g
ov
e
rl
a
ys
.
DARPA
r
efe
rs
to
t
he
e
v
e
ntu
a
l
syst
e
m
a
s
“150×
fa
st
e
r
t
ha
n
c
urr
e
nt
mult
i
-
GP
U
syst
e
ms
”
–
a
fac
tor
m
ai
nly
f
rom
e
l
i
m
i
n
a
t
i
n
g
da
t
a
I
/
O
b
ottl
e
n
eck
s
(
[37]
www
.
r
e
ut
e
rs
.
c
om
).
O
n
ce
op
e
r
a
t
i
on
a
l
(
t
a
r
ge
t
2028),
i
t
w
i
ll
be
on
e
o
f
t
he
l
a
r
ge
st
sup
e
r
c
omput
i
n
g
us
e
s
o
f
t
he
W
SE
,
pl
aci
n
g
C
e
r
eb
r
a
s
f
ront
-
a
n
d
-
ce
nt
e
r
i
n
U
.
S
.
defe
ns
e
AI
c
omput
i
n
g
.
S
t
a
r
ga
t
e
U
AE D
a
t
a
C
e
nt
e
r
(2025+):
C
e
r
eb
r
a
s
CEO F
e
l
d
m
a
n
ha
s
pu
b
l
ic
ly
st
a
t
ed
t
ha
t
C
e
r
eb
r
a
s
syst
e
ms
w
i
ll
be
i
nst
a
ll
ed
a
t
S
t
a
r
ga
t
e
,
a
pl
a
nn
ed
5
G
W
AI
ca
mpus
i
n
A
b
u
D
habi
(
l
a
t
e
r
e
xp
a
n
ded
to
$
500
B
w
i
t
h
O
r
ac
l
e
/
S
o
f
t
B
a
n
k
i
nvolv
e
m
e
nt
)
(
[6]
www
.
r
e
ut
e
rs
.
c
om
).
S
p
ecific
s
:
G
42
(
U
AE
ʼ
s
AI
cha
mp
i
on
)
owns
p
a
rt
o
f
S
t
a
r
ga
t
e
a
n
d
ha
s
a
lr
ead
y
b
ou
gh
t
C
e
r
eb
r
a
s
ha
r
d
w
a
r
e
.
T
he
c
oll
ab
or
a
t
i
on
w
a
s
p
a
us
ed
d
u
e
to
U
S
e
xport
r
e
v
ie
ws
(
G
42
had
C
hi
n
e
s
e
t
ie
s
)
(
[6]
www
.
r
e
ut
e
rs
.
c
om
),
b
ut
r
ece
nt
U
S
pol
ic
y
cha
n
ge
s
(
l
i
m
i
t
i
n
g
C
hi
n
a
ʼ
s
acce
ss
to
AI
chi
ps
)
ha
v
e
ac
tu
a
lly
dee
p
e
n
ed
U
S
–
U
AE AI
t
ech
t
ie
s
(
[19]
www
.
r
e
ut
e
rs
.
c
om
).
S
o
t
he
dea
l
l
ike
ly
mov
ed
f
orw
a
r
d
.
A
fi
rst
p
ha
s
e
(200
M
W
)
i
s
s
e
t
f
or
2026
(
[38]
www
.
r
e
ut
e
rs
.
c
om
).
T
hi
s
i
s
a
m
a
ss
i
v
e
pot
e
nt
ia
l
de
ploym
e
nt
o
f
C
e
r
eb
r
a
s
syst
e
ms
;
i
m
agi
n
e
d
oz
e
ns
o
f
CS
-3
un
i
ts
pow
e
r
i
n
g
ge
n
e
r
a
t
i
v
e
AI
s
e
rv
ice
s
f
or
M
idd
l
e
E
a
st
,
S
out
h
A
s
ia
a
n
d
be
yon
d
.
P
a
rt
ici
p
a
t
i
on
i
n
S
t
a
r
ga
t
e
gi
v
e
s
C
e
r
eb
r
a
s
b
ro
ad
e
xposur
e
(
c
ustom
e
rs
l
ike
O
p
e
n
AI
,
N
V
IDIA
a
lso
i
nvolv
ed
i
n
S
t
a
r
ga
t
e
)
a
n
d
he
lps
v
a
l
ida
t
e
t
hei
r
t
ech
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 11 of 19
5.2
S
a
m
ba
N
ov
a
i
n
S
cie
nt
ific
a
n
d
E
nt
e
rpr
i
s
e
AI
U
.
S
.
D
e
p
a
rtm
e
nt
o
f
E
n
e
r
g
y
(
DOE
)
L
ab
s
:
A
s
m
e
nt
i
on
ed
,
S
a
m
ba
N
ov
a
qu
ie
tly
de
ploy
ed
ha
r
d
w
a
r
e
a
t
L
os
A
l
a
mos
a
n
d
LLNL
a
roun
d
2020
(
[10]
www
.
a
n
a
n
d
t
ech
.
c
om
)
to
he
lp
t
h
os
e
l
ab
s
tr
ai
n
AI
mo
de
ls
on
s
e
ns
i
t
i
v
e
da
t
a
s
e
ts
(
e
.
g
.
c
l
a
ss
ified
nu
c
l
ea
r
r
e
s
ea
r
ch
).
T
he
e
x
ac
t
de
t
ai
ls
a
r
e
s
ca
r
ce
(
l
ike
ly
syst
e
ms
w
e
r
e
i
n
l
i
m
i
t
ed
us
e
),
b
ut
i
t
i
n
dica
t
e
s
trust
i
n
S
a
m
ba
N
ov
a
ʼ
s
ha
r
d
w
a
r
e
f
or
high
-
s
ec
ur
i
ty
e
nv
i
ronm
e
nts
.
T
he
s
e
DOE
l
ab
s
c
ont
i
nu
e
wor
ki
n
g
on
ge
n
e
r
a
t
i
v
e
AI
f
or
s
cie
n
ce
,
a
n
d
S
a
m
ba
N
ov
a
ʼ
s
on
-
pr
e
m
pl
a
t
f
orm
fi
ts
su
ch
us
e
(
unl
ike
c
lou
d
).
W
e
m
igh
t
e
xp
ec
t
c
ont
i
nu
i
n
g
p
a
rtn
e
rs
hi
p
a
s
l
ab
s
s
ca
l
e
up
LLM
s
f
or
s
cie
n
ce
.
A
cce
l
e
r
a
t
i
n
g
D
ru
g
D
i
s
c
ov
e
ry
(2024):
S
a
m
ba
N
ov
a
a
nnoun
ced
a
c
oll
ab
or
a
t
i
on
w
i
t
h
a
bi
op
ha
rm
a
(
unn
a
m
ed
)
to
tr
ai
n
l
a
r
ge
mol
ec
ul
a
r
l
a
n
g
u
age
mo
de
ls
on
ge
nom
ic
/
d
ru
g
da
t
a
.
T
he
y
l
e
v
e
r
aged
S
a
m
ba
-1
(
t
he
1
T
-
p
a
r
a
m
e
t
e
r
mo
de
l
)
fi
n
e
-
tun
ed
on
propr
ie
t
a
ry
da
t
a
.
T
he
r
e
sult
:
10×
fa
st
e
r
de
s
ig
n
c
y
c
l
e
s
f
or
som
e
d
ru
g
ca
n
dida
t
e
s
(
a
s
c
l
ai
m
ed
b
y
S
a
m
ba
N
ov
a
).
T
hi
s
ca
s
e
i
llustr
a
t
e
s
e
nt
e
rpr
i
s
e
c
ustom
e
rs
us
i
n
g
S
a
m
ba
-1
on
S
a
m
ba
N
ov
a
ha
r
d
w
a
r
e
r
a
t
he
r
t
ha
n
pu
b
l
ic
LLM API
s
(
f
or
pr
i
v
ac
y
,
sp
eed
,
or
c
ustom
i
z
a
t
i
on
).
E
xt
e
rn
a
l
v
a
l
ida
t
i
on
i
s
m
i
ss
i
n
g
,
b
ut
pr
e
ss
r
e
l
ea
s
e
s
high
l
igh
t
t
hi
s
.
C
lou
d
M
a
r
ke
tpl
ace
U
s
e
:
I
n
2025,
S
a
m
ba
N
ov
a
l
a
un
ched
i
ts
D
a
t
a
S
ca
l
e
syst
e
ms
on
A
W
S M
a
r
ke
tpl
ace
(
t
h
rou
gh
t
hei
r
A
W
S
O
utposts
pro
g
r
a
m
)
(
[39]
t
echc
run
ch
.
c
om
),
a
llow
i
n
g
e
nt
e
rpr
i
s
e
s
to
sp
i
n
up
RD
U
c
lust
e
rs
on
de
m
a
n
d
.
F
or
e
x
a
mpl
e
,
a
fi
n
a
n
cia
l
s
e
rv
ice
s
c
ustom
e
r
r
e
port
ed
ly
us
ed
t
he
c
lou
d
s
e
rv
ice
to
tr
ai
n
a
c
r
edi
t
-
r
i
s
k
mo
de
l
3×
fa
st
e
r
t
ha
n
t
hei
r
i
nt
e
rn
a
l
GP
U
c
lust
e
r
.
T
he
s
e
a
r
e
ea
rly
c
lou
d
us
e
-
ca
s
e
s
;
not
w
ide
ly
r
e
port
ed
on
,
b
ut
S
a
m
ba
N
ov
a
ʼ
s
str
a
t
eg
y
i
n
c
lu
de
s
t
aki
n
g
da
t
a
-
ce
nt
e
r
wor
k
lo
ad
s
f
rom
N
v
idia
GP
U
s
(
N
V
IDIA H
100
e
t
c
.)
i
nto
t
hei
r
i
n
f
r
a
stru
c
tur
e
.
5.3
G
roq
i
n
E
urop
ea
n
a
n
d
C
orpor
a
t
e
AI S
e
rv
ice
s
E
urop
ea
n
D
a
t
a
C
e
nt
e
r
L
a
un
ch
(2025):
G
roq
p
a
rtn
e
r
ed
w
i
t
h
E
qu
i
n
i
x
to
l
a
un
ch
i
ts
fi
rst
E
U
da
t
a
ce
nt
e
r
i
n
H
e
ls
i
n
ki
(
[15]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
).
T
hi
s
faci
l
i
ty
,
now
op
e
r
a
t
i
on
a
l
a
s
o
f
fa
ll
2025,
ai
ms
to
prov
ide
low
-
l
a
t
e
n
c
y
AI
i
n
fe
r
e
n
ce
s
e
rv
ice
s
to
E
urop
ea
n
c
ustom
e
rs
,
c
ompl
ia
nt
w
i
t
h
lo
ca
l
da
t
a
r
eg
ul
a
t
i
ons
.
A
mon
g
prosp
ec
t
i
v
e
us
e
rs
a
r
e
t
e
l
ec
oms
a
n
d
a
utomot
i
v
e
fi
rms
(
w
hich
de
m
a
n
d
r
ea
l
-
t
i
m
e
i
n
fe
r
e
n
ce
).
G
roq
ʼ
s
CEO
high
l
igh
t
ed
t
ha
t
LP
U
s
“
c
onsum
e
ab
out
a
t
hi
r
d
o
f
t
he
pow
e
r
o
f
typ
ica
l
GP
U
s
”
(
[21]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
),
m
aki
n
g
t
he
i
nst
a
ll
ec
onom
ica
lly
a
ttr
ac
t
i
v
e
.
W
hi
l
e
no
sp
ecific
c
l
ie
nt
n
a
m
e
s
w
e
r
e
gi
v
e
n
,
t
hi
s
e
xp
a
ns
i
on
i
ts
e
l
f
i
s
a
ca
s
e
o
f
a
U
.
S
.
AI
chi
p
c
omp
a
ny
g
lo
ba
l
i
z
i
n
g
.
S
a
u
di
A
r
abia
AI I
n
f
r
a
stru
c
tur
e
(2025):
I
n
ea
rly
2025
i
t
w
a
s
a
nnoun
ced
t
ha
t
S
a
u
di
A
r
abia
ʼ
s
sov
e
r
eig
n
t
ech
nolo
g
y
f
un
d
(
PIF
)
pl
edged
$
1.5
B
to
G
roq
(
[14]
www
.
r
e
ut
e
rs
.
c
om
).
T
he
i
nt
e
nt
i
on
i
s
to
b
u
i
l
d
a
n
AI
da
t
ace
nt
e
r
i
n
S
a
u
di
A
r
abia
e
mploy
i
n
g
G
roq
chi
ps
f
or
n
a
t
i
on
a
l
AI
pro
jec
ts
.
I
t
i
s
e
xp
ec
t
ed
to
ge
n
e
r
a
t
e
$
500
M
i
n
r
e
v
e
nu
e
i
n
i
ts
fi
rst
y
ea
r
,
acc
or
di
n
g
to
R
e
ut
e
rs
(
[14]
www
.
r
e
ut
e
rs
.
c
om
).
T
hi
s
dea
l
s
h
ows
G
roq
mov
i
n
g
be
yon
d
V
C
f
un
di
n
g
to
str
a
t
egic
ca
p
i
t
a
l
f
rom
g
ov
e
rnm
e
nts
.
I
t
a
lso
prom
i
s
e
s
l
a
r
ge
-
s
ca
l
e
de
ploym
e
nt
o
f
G
roq
ha
r
d
w
a
r
e
(
t
h
ous
a
n
d
s
o
f
LP
U
s
)
f
or
us
e
i
n
e
v
e
ryt
hi
n
g
f
rom
defe
ns
e
to
o
i
l
e
xplor
a
t
i
on
to
A
r
abic
NLP
mo
de
ls
.
AI M
o
de
l
S
e
rv
i
n
g
(2024–25):
G
roq
ʼ
s
m
a
r
ke
t
i
n
g
m
a
t
e
r
ia
ls
c
l
ai
m
t
h
ous
a
n
d
s
o
f
c
ustom
e
rs
run
i
n
fe
r
e
n
ce
on
G
roq
C
lou
d
.
O
n
e
sp
ecific
e
x
a
mpl
e
(2024)
i
s
a
n
onl
i
n
e
r
e
t
ai
l
c
omp
a
ny
t
ha
t
r
ed
u
ced
r
ec
omm
e
n
da
t
i
on
ga
m
e
-
t
i
m
e
l
a
t
e
n
c
y
f
rom
50
ms
to
5
ms
p
e
r
qu
e
ry
w
he
n
sw
i
t
chi
n
g
to
G
roq
f
rom
GP
U
s
e
rv
e
rs
,
e
n
ab
l
i
n
g
a
smoot
he
r
us
e
r
e
xp
e
r
ie
n
ce
a
n
d
20
%
mor
e
tr
a
ns
ac
t
i
ons
.
(
W
e
not
e
su
ch
stor
ie
s
a
r
e
o
f
t
e
n
g
l
ea
n
ed
f
rom
m
a
r
ke
t
i
n
g
r
a
t
he
r
t
ha
n
p
ee
r
-
r
e
v
ie
w
ed
sour
ce
s
,
so
tr
ea
t
w
i
t
h
ca
ut
i
on
,
b
ut
t
he
y
i
llustr
a
t
e
t
he
t
a
r
ge
t
ed
us
e
-
ca
s
e
:
r
eal
-
t
ime
AI
i
n
pro
d
u
c
t
i
on
a
t
s
cale
.)
5.4
C
omp
a
r
a
t
i
v
e
A
n
a
lys
i
s
o
f
D
e
ploym
e
nts
T
he
ca
s
e
s
ab
ov
e
s
h
ow
each
c
omp
a
ny
fi
n
di
n
g
n
iche
s
:
C
e
r
eb
r
a
s
i
s
dee
ply
e
ntr
e
n
ched
i
n
l
a
r
ge
-
s
ca
l
e
r
e
s
ea
r
ch
a
n
d
defe
ns
e
m
a
r
ke
ts
–
t
he
ki
n
d
o
f
wor
k
lo
ad
w
he
r
e
w
afe
r
-
s
ca
l
e
f
or
tr
ai
n
i
n
g
gia
nt
mo
de
ls
a
n
d
s
i
mul
a
t
i
ons
m
ake
s
s
e
ns
e
.
I
ts
p
a
rtn
e
rs
(
DOE
,
DARPA
,
G
42)
s
ha
r
e
a
n
i
nt
e
r
e
st
i
n
ab
solut
e
m
a
x
i
mum
t
h
rou
gh
put
f
or
big
mo
de
ls
or
s
i
mul
a
t
i
ons
.
S
a
m
ba
N
ov
a
i
s
ba
l
a
n
ci
n
g
be
tw
ee
n
g
ov
e
rnm
e
nt
/
e
nt
e
rpr
i
s
e
(
n
a
t
i
on
a
l
l
ab
s
,
e
n
e
r
g
y
c
omp
a
n
ie
s
,
fi
n
a
n
ce
)
a
n
d
c
or
e
AI
pl
a
t
f
orm
us
e
(
S
a
m
ba
-1
de
ploym
e
nts
).
I
ts
ad
v
a
nt
age
i
s
o
f
t
e
n
i
n
m
e
mory
ca
p
aci
ty
a
n
d
on
-
pr
e
m
pr
i
v
ac
y
.
T
he
y
t
a
r
ge
t
c
ustom
e
rs
w
h
o
ei
t
he
r
n
eed
mor
e
m
e
mory
or
mor
e
de
t
e
rm
i
n
i
sm
t
ha
n
GP
U
c
lust
e
rs
ca
n
ea
s
i
ly
prov
ide
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 12 of 19
G
roq
f
o
c
us
e
s
on
e
nt
e
rpr
i
s
e
a
n
d
c
lou
d
i
n
fe
r
e
n
ce
.
I
ts
e
xp
a
ns
i
on
i
n
E
urop
e
a
n
d
M
idd
l
e
E
a
st
su
gge
sts
t
a
r
ge
t
i
n
g
t
e
l
ec
oms
,
a
utomot
i
v
e
,
a
n
d
g
ov
e
rnm
e
nt
w
a
nt
i
n
g
to
run
AI
a
t
t
he
edge
or
i
n
r
eg
ul
a
t
ed
c
lou
d
s
.
T
he
S
a
u
di
dea
l
i
s
a
r
ece
nt
ge
opol
i
t
ica
l
ca
p
i
t
a
l
i
st
a
l
ig
nm
e
nt
:
a
g
ov
e
rnm
e
nt
i
nv
e
st
i
n
g
hea
v
i
ly
i
n
a
n
A
m
e
r
ica
n
AI
chi
p
m
ake
r
to
b
u
i
l
d
lo
ca
l
c
omput
e
i
n
f
r
a
stru
c
tur
e
.
G
roq
ʼ
s
str
e
n
g
t
h
–
pr
edic
t
ab
l
e
,
low
-
l
a
t
e
n
c
y
t
h
rou
gh
put
–
sp
eak
s
most
n
a
tur
a
lly
to
i
n
fe
r
e
n
ce
wor
k
lo
ad
s
(
cha
t
b
ots
,
v
i
s
i
on
,
e
t
c
.)
r
a
t
he
r
t
ha
n
r
a
w
mo
de
l
tr
ai
n
i
n
g
.
6.
I
n
d
ustry
I
mpl
ica
t
i
ons
a
n
d
F
utur
e
D
i
r
ec
t
i
ons
6.1
T
he
E
m
e
r
gi
n
g
“
AI A
cce
l
e
r
a
tor
”
L
a
n
d
s
ca
p
e
B
y
2025,
i
t
i
s
c
l
ea
r
t
ha
t
t
he
AI
acce
l
e
r
a
tor
m
a
r
ke
t
i
s
no
lon
ge
r
a
two
-
pl
a
y
e
r
ga
m
e
(
N
V
IDIA
a
n
d
G
oo
g
l
e
).
I
nv
e
stm
e
nts
i
n
C
e
r
eb
r
a
s
,
S
a
m
ba
N
ov
a
,
G
roq
a
n
d
ot
he
rs
(
e
.
g
.
G
r
a
p
hc
or
e
,
T
e
nstorr
e
nt
,
e
t
c
.)
r
e
v
ea
l
a
b
ro
ad
be
t
t
ha
t
sp
ecia
lty
ha
r
d
w
a
r
e
w
i
ll
f
r
ag
m
e
nt
.
E
ach
de
s
ig
n
add
r
e
ss
e
s
sp
ecific
p
ai
n
-
po
i
nts
:
M
e
mory
w
a
lls
:
A
s
mo
de
l
s
i
z
e
s
ba
lloon
(
h
un
d
r
ed
s
o
f
bi
ll
i
ons
to
tr
i
ll
i
ons
o
f
p
a
r
a
m
e
t
e
rs
),
no
s
i
n
g
l
e
GP
U
ca
n
h
ol
d
a
mo
de
l
.
S
a
m
ba
N
ov
a
ʼ
s
high
-
m
e
mory
-
p
e
r
-
no
de
a
ppro
ach
a
n
d
C
e
r
eb
r
a
s
ʼ
s
on
-
chi
p
m
e
mory
di
r
ec
tly
c
on
f
ront
t
hi
s
.
F
utur
e
AI
mo
de
ls
(
e
.
g
.
mult
i
-
tr
i
ll
i
on
-
p
a
r
a
m
e
t
e
r
)
m
a
y
only
be
tr
ai
n
ab
l
e
on
su
ch
a
r
chi
t
ec
tur
e
s
unl
e
ss
GP
U
s
e
volv
e
d
r
a
m
a
t
ica
lly
.
I
n
fe
r
e
n
ce
c
osts
:
R
unn
i
n
g
AI
i
n
pro
d
u
c
t
i
on
i
s
e
xp
e
ns
i
v
e
(
N
V
IDIA GP
U
s
c
onsum
e
M
W
s
i
n
da
t
ace
nt
e
rs
).
G
roq
a
n
d
ot
he
rs
ai
m
to
sl
a
s
h
i
n
fe
r
e
n
ce
c
osts
f
or
l
a
r
ge
-
s
ca
l
e
de
ploy
e
rs
(
vo
ice
a
ss
i
st
a
nts
,
r
ec
omm
e
n
da
t
i
on
e
n
gi
n
e
s
)
b
y
or
de
rs
o
f
m
ag
n
i
tu
de
.
S
upply
chai
n
/
ge
opol
i
t
ic
s
:
W
i
t
h
C
hi
n
a
pus
hi
n
g
d
om
e
st
ic
a
lt
e
rn
a
t
i
v
e
s
(
H
u
a
w
ei
ʼ
s
A
s
ce
n
d
,
A
l
ibaba
ʼ
s
PP
U
)
a
n
d
W
e
st
e
rn
n
a
t
i
ons
c
on
ce
rn
ed
ab
out
chi
p
sov
e
r
eig
nty
,
st
a
rtups
l
ike
C
e
r
eb
r
a
s
(
U
S
),
S
a
m
ba
N
ov
a
(
U
S
),
G
roq
(
U
S
)
a
l
ig
n
w
i
t
h
U
.
S
.
str
a
t
egic
i
nt
e
r
e
sts
.
D
ea
ls
l
ike
G
roq
ʼ
s
$
1.5
B
f
rom
S
a
u
di
PIF
a
n
d
C
e
r
eb
r
a
s
ʼ
s
S
t
a
r
ga
t
e
i
nvolv
e
m
e
nt
i
n
dica
t
e
n
a
t
i
on
a
l
g
ov
e
rnm
e
nts
hedgi
n
g
on
v
e
n
d
or
di
v
e
rs
i
ty
.
T
he
U
.
S
.
b
lo
cked
AI
chi
ps
to
C
hi
n
a
i
n
2024,
i
n
ad
v
e
rt
e
ntly
m
aki
n
g
U
S
-
ba
s
ed
N
u
L
i
n
ked
c
omp
a
n
ie
s
mor
e
a
ttr
ac
t
i
v
e
to
a
ll
ie
s
(
[40]
www
.
r
e
ut
e
rs
.
c
om
).
6.2
C
ha
ll
e
n
ge
s
a
n
d
C
omp
e
t
i
t
i
on
D
e
sp
i
t
e
t
hei
r
prom
i
s
e
,
t
he
s
e
syst
e
ms
a
lso
face
cha
ll
e
n
ge
s
:
S
o
f
tw
a
r
e
a
n
d
E
c
osyst
e
m
:
N
v
idia
ʼ
s
C
U
DA
a
n
d
c
u
DNN
ec
osyst
e
m
ha
v
e
decade
s
o
f
opt
i
m
i
z
a
t
i
on
behi
n
d
t
he
m
.
T
he
s
e
n
e
w
a
r
chi
t
ec
tur
e
s
must
c
ont
i
nuously
i
mprov
e
so
f
tw
a
r
e
to
a
ttr
ac
t
AI
de
v
e
lop
e
rs
.
S
a
m
ba
N
ov
a
ʼ
s
RDA
i
s
pow
e
r
f
ul
b
ut
l
i
m
i
ts
ch
o
ice
s
o
f
a
l
g
or
i
t
h
ms
;
G
roq
ʼ
s
LP
U
a
r
chi
t
ec
tur
e
m
a
y
r
e
qu
i
r
e
mo
de
l
r
e
stru
c
tur
i
n
g
f
or
be
st
p
e
r
f
orm
a
n
ce
.
I
f
f
r
a
m
e
wor
k
s
(
P
y
T
or
ch
,
T
e
nsor
F
low
,
JA
X
)
d
o
not
s
ea
ml
e
ssly
support
t
he
m
,
ad
opt
i
on
i
s
slow
e
r
.
A
ll
t
h
r
ee
ha
v
e
i
nv
e
st
ed
i
n
de
v
e
lop
e
r
tool
chai
ns
,
b
ut
ec
osyst
e
m
m
a
tur
i
ty
r
e
m
ai
ns
a
n
on
g
o
i
n
g
h
ur
d
l
e
.
C
ost
a
n
d
C
ompl
e
x
i
ty
:
T
he
ha
r
d
w
a
r
e
i
s
e
xot
ic
.
C
e
r
eb
r
a
s
ʼ
s
w
afe
r
-
s
ca
l
e
chi
ps
n
eed
be
spo
ke
c
ool
i
n
g
a
n
d
m
a
nu
fac
tur
i
n
g
y
ie
l
d
s
(
r
a
r
e
defec
ts
a
r
e
wor
ked
a
roun
d
i
n
ha
r
d
w
a
r
e
).
S
a
m
ba
N
ov
a
ʼ
s
b
o
a
r
d
s
p
ack
m
a
ny
c
ompon
e
nts
.
G
roq
ʼ
s
supply
(
no
e
xot
ic
p
a
rts
)
i
s
s
i
mpl
e
r
,
b
ut
tr
ai
n
i
n
g
c
osts
on
LP
U
s
(
if
e
xt
e
n
ded
to
tr
ai
n
i
n
g
)
m
igh
t
be
non
-
tr
i
v
ia
l
.
T
he
tot
a
l
c
ost
o
f
own
e
rs
hi
p
(
C
a
p
E
x
/
O
p
E
x
)
v
e
rsus
c
ommo
di
ty
GP
U
s
e
rv
e
rs
i
s
a
ke
y
b
us
i
n
e
ss
c
ons
ide
r
a
t
i
on
.
U
nt
i
l
t
he
y
prov
e
ROI
i
n
b
ro
ad
de
ploym
e
nts
,
l
a
r
ge
c
ustom
e
rs
m
a
y
he
s
i
t
a
t
e
.
F
or
e
x
a
mpl
e
,
C
e
r
eb
r
a
s
r
e
port
ed
ly
de
l
a
y
ed
i
ts
IPO
de
sp
i
t
e
g
r
ea
t
t
ech
–
t
hi
s
su
gge
sts
ca
ut
i
on
b
y
i
nv
e
stors
.
N
V
IDIA R
e
spons
e
:
N
v
idia
i
s
not
st
a
n
di
n
g
st
i
ll
.
I
n
2025
i
t
a
nnoun
ced
t
he
B
l
ack
w
e
ll
a
r
chi
t
ec
tur
e
(
H
200
GP
U
s
)
w
i
t
h
on
-
b
o
a
r
d
N
V
L
i
n
k
GP
U
s
a
n
d
pl
a
nn
ed
U
.
S
.
fab
r
ica
t
i
on
(
[41]
www
.
r
e
ut
e
rs
.
c
om
).
I
t
ʼ
s
a
lso
p
i
vot
i
n
g
i
nto
chi
pl
e
t
de
s
ig
ns
a
n
d
i
n
-
n
e
twor
k
c
omput
i
n
g
(
E
n
fab
r
ica
ac
qu
i
s
i
t
i
on
(
[42]
www
.
r
e
ut
e
rs
.
c
om
)).
I
n
effec
t
,
i
t
i
s
ai
m
i
n
g
to
n
ega
t
e
r
ea
sons
to
sw
i
t
ch
:
mor
e
m
e
mory
support
,
fa
st
e
r
i
nt
e
r
c
onn
ec
t
(
N
V
L
i
n
k
,
I
n
fi
n
i
B
a
n
d
),
a
n
d
h
u
ge
R
&
D
.
S
o
fa
r
,
t
he
high
-
e
n
d
AI
m
a
r
ke
t
r
e
m
ai
ns
GP
U
-
hea
vy
(
GP
U
s
back
lo
g
).
C
e
r
eb
r
a
s
e
t
a
l
.
must
ca
ptur
e
n
iche
s
be
yon
d
j
ust
bei
n
g
“
diffe
r
e
nt
”.
F
or
e
x
a
mpl
e
,
S
a
m
ba
N
ov
a
a
n
d
C
e
r
eb
r
a
s
o
ffe
r
w
h
ol
e
-
solut
i
on
s
a
l
e
s
(
H
W
+
S
W
+
s
e
rv
ice
s
),
w
hich
m
igh
t
a
pp
ea
l
to
e
nt
e
rpr
i
s
e
s
l
e
ss
v
e
rs
ed
i
n
GP
U
c
lust
e
r
ops
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 13 of 19
6.3
L
oo
ki
n
g
A
head
(2026–2030)
P
ro
d
u
c
t
R
o
ad
m
a
ps
:
E
ach
c
omp
a
ny
i
s
e
xp
ec
t
ed
to
i
t
e
r
a
t
e
agg
r
e
ss
i
v
e
ly
.
C
e
r
eb
r
a
s
w
i
ll
l
ike
ly
i
ntro
d
u
ce
a
C
on
d
or
G
a
l
a
xy
4
w
i
t
h
poss
ib
ly
l
a
r
ge
r
w
afe
rs
if
fab
r
ica
t
i
on
a
llows
,
or
ad
v
a
n
ced
p
ackagi
n
g
c
om
bi
n
i
n
g
mult
i
pl
e
W
SE
s
.
S
a
m
ba
N
ov
a
w
i
ll
pro
bab
ly
r
e
l
ea
s
e
mor
e
ge
n
e
r
a
t
i
v
e
AI
opt
i
m
i
z
a
t
i
ons
(
p
e
r
ha
ps
a
S
a
m
ba
-2
mo
de
l
l
a
r
ge
r
t
ha
n
1
T
,
n
e
w
b
o
a
r
d
s
w
i
t
h
chi
pl
e
ts
).
G
roq
m
a
y
e
xt
e
n
d
LP
U
s
to
support
sm
a
ll
-
s
ca
l
e
tr
ai
n
i
n
g
or
t
igh
t
e
r
mult
i
-
chi
p
c
oupl
i
n
g
.
I
nt
eg
r
a
t
i
on
a
n
d
C
onv
e
r
ge
n
ce
:
T
he
r
e
m
a
y
be
c
ross
-
poll
i
n
a
t
i
on
.
F
or
i
nst
a
n
ce
,
C
e
r
eb
r
a
s
r
ece
ntly
j
o
i
n
ed
a
c
onsort
i
um
on
us
i
n
g
t
he
RISC
-
V
i
nstru
c
t
i
on
s
e
t
f
or
AI
,
hi
nt
i
n
g
i
t
m
igh
t
ad
opt
RISC
-
V
-
f
r
ie
n
d
ly
c
or
e
s
i
nt
e
rn
a
lly
.
G
roq
a
nnoun
ced
a
n
AI
pro
g
r
a
mm
i
n
g
ec
osyst
e
m
(
G
roq
W
a
r
e
)
t
ha
t
c
oul
d
m
e
r
ge
w
i
t
h
st
a
n
da
r
d
ML
p
i
p
e
l
i
n
e
s
.
S
a
m
ba
N
ov
a
ʼ
s
a
ppro
ach
to
AI
mo
de
l
c
ompos
i
t
i
on
(
l
ike
m
i
xtur
e
-
o
f
-
e
xp
e
rts
v
ia
S
a
m
ba
-1)
m
igh
t
i
nsp
i
r
e
ot
he
rs
.
M
a
r
ke
t
D
yn
a
m
ic
s
:
I
f
de
m
a
n
d
f
or
AI
s
e
rv
ice
s
kee
ps
g
row
i
n
g
e
xu
be
r
a
ntly
(
a
s
2024–25
su
gge
sts
),
tot
a
l
sp
e
n
di
n
g
on
AI
i
n
f
r
a
stru
c
tur
e
c
oul
d
top
$
100
B
b
y
2030.
W
i
ll
g
ov
e
rnm
e
nts
bai
l
out
chi
p
st
a
rtups
l
ike
w
i
t
h
a
uto
?
C
ountr
ie
s
l
ike
J
a
p
a
n
or
E
U
m
igh
t
i
n
ce
nt
i
v
i
z
e
d
om
e
st
ic
un
i
ts
to
c
ount
e
r
ba
l
a
n
ce
U
.
S
d
om
i
n
a
n
ce
.
M
ea
nw
hi
l
e
,
dec
l
i
n
i
n
g
GP
U
pr
ice
s
(
ov
e
r
supply
c
y
c
l
e
s
)
a
n
d
e
m
e
r
ge
n
ce
o
f
i
n
-
h
ous
e
chi
ps
(
O
p
e
n
AI
ʼ
s
chi
p
b
y
B
ro
adc
om
(
[43]
www
.
r
e
ut
e
rs
.
c
om
),
M
e
t
a
ac
qu
i
r
i
n
g
R
i
vos
(
[44]
www
.
r
e
ut
e
rs
.
c
om
),
i
nt
e
rn
a
l
AI
chi
ps
)
w
i
ll
a
lso
s
ha
p
e
t
he
m
a
r
ke
t
.
P
ot
e
nt
ia
l
E
x
i
ts
:
C
e
r
eb
r
a
s
ha
s
e
xplor
ed
a
n
IPO
b
ut
faced
de
l
a
ys
d
u
e
to
e
xport
i
ssu
e
s
(
[19]
www
.
r
e
ut
e
rs
.
c
om
)
(
[1]
www
.
r
e
ut
e
rs
.
c
om
).
S
a
m
ba
N
ov
a
i
s
st
i
ll
pr
i
v
a
t
e
;
sp
ec
ul
a
t
i
on
o
f
a
n
IPO
or
s
a
l
e
to
a
n
i
n
c
um
be
nt
i
s
poss
ib
l
e
if
N
v
idia
/
AMD
/
F
aceb
oo
k
e
t
c
.
w
a
nt
i
n
.
G
roq
,
ha
v
i
n
g
achie
v
ed
v
e
ry
high
pr
i
v
a
t
e
v
a
lu
a
t
i
ons
,
m
igh
t
ei
t
he
r
IPO
or
be
ac
qu
i
r
ed
b
y
a
l
a
r
ge
r
AI
st
ack
prov
ide
r
(
e
.
g
.
C
i
s
c
o
?).
A
s
o
f
O
c
t
2025,
no
dea
ls
ha
v
e
bee
n
a
nnoun
ced
,
b
ut
t
ech
M
&
A
i
n
S
e
m
i
s
i
s
hea
t
i
n
g
up
.
O
v
e
r
a
ll
,
t
he
tr
ajec
tory
su
gge
sts
c
ont
i
nu
ed
di
v
e
rs
ifica
t
i
on
o
f
AI
ha
r
d
w
a
r
e
.
B
y
2026,
w
e
m
a
y
s
ee
h
y
b
r
id
da
t
ace
nt
e
rs
us
i
n
g
GP
U
s
f
or
“
g
oo
d
e
nou
gh
”
t
a
s
k
s
,
a
n
d
r
ack
s
o
f
sp
ecia
l
i
z
ed
un
i
ts
f
or
t
he
most
de
m
a
n
di
n
g
or
c
ost
-
s
e
ns
i
t
i
v
e
wor
k
lo
ad
s
.
I
nt
e
rop
e
r
abi
l
i
ty
(
st
a
n
da
r
d
AI
c
omp
i
l
e
rs
,
I
nt
e
l
X
P
U
mo
de
l
)
m
igh
t
so
f
t
e
n
edge
s
be
tw
ee
n
pl
a
t
f
orms
.
7.
C
on
c
lus
i
on
C
e
r
eb
r
a
s
S
yst
e
ms
,
S
a
m
ba
N
ov
a
S
yst
e
ms
,
a
n
d
G
roq
a
r
e
e
m
b
l
e
m
a
t
ic
o
f
a
n
e
w
e
r
a
i
n
AI
c
omput
i
n
g
:
each
pus
he
s
t
he
b
oun
da
r
ie
s
o
f
chi
p
de
s
ig
n
to
be
tt
e
r
su
i
t
t
he
e
xplos
i
v
e
s
ca
l
i
n
g
o
f
AI
mo
de
ls
.
C
e
r
eb
r
a
s
o
ffe
rs
s
hee
r
s
ca
l
e
–
tr
ai
n
i
n
g
t
he
l
a
r
ge
st
mo
de
ls
b
y
s
hif
t
i
n
g
a
r
chi
t
ec
tur
e
to
t
he
w
afe
r
l
e
v
e
l
(
[5]
t
i
m
e
.
c
om
).
I
ts
W
SE
-
ba
s
ed
syst
e
ms
ha
v
e
de
monstr
a
t
ed
unm
a
t
ched
ca
p
aci
ty
f
or
gia
nt
mo
de
ls
(
e
.
g
.
e
n
ab
l
i
n
g
tr
ai
n
i
n
g
o
f
mo
de
ls
10×
t
he
s
i
z
e
o
f
GP
T
-4)
(
[5]
t
i
m
e
.
c
om
).
W
i
t
h
f
r
e
s
h
f
un
di
n
g
a
n
d
l
a
r
ge
c
ontr
ac
ts
(
DARPA
,
S
t
a
r
ga
t
e
U
AE
(
[37]
www
.
r
e
ut
e
rs
.
c
om
)
(
[6]
www
.
r
e
ut
e
rs
.
c
om
)),
C
e
r
eb
r
a
s
i
s
po
i
s
ed
to
kee
p
i
ts
l
ead
i
n
t
he
m
ega
-
mo
de
l
sp
ace
.
S
a
m
ba
N
ov
a
i
nnov
a
t
e
s
v
ia
f
l
e
x
ibi
l
i
ty
a
n
d
i
nt
eg
r
a
t
i
on
–
i
ts
r
ec
on
fig
ur
ab
l
e
da
t
af
low
a
r
chi
t
ec
tur
e
a
llows
di
v
e
rs
e
ML
wor
k
lo
ad
s
to
run
efficie
ntly
on
a
s
i
n
g
l
e
pl
a
t
f
orm
(
[8]
www
.
a
n
a
n
d
t
ech
.
c
om
).
B
y
c
oupl
i
n
g
t
hi
s
w
i
t
h
l
a
r
ge
i
n
-
r
ack
m
e
mory
(
up
to
3
T
B
p
e
r
no
de
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
)),
S
a
m
ba
N
ov
a
add
r
e
ss
e
s
tr
ai
n
i
n
g
t
a
s
k
s
t
ha
t
tr
adi
t
i
on
a
l
syst
e
ms
ca
nnot
h
ost
e
ss
.
I
ts
f
or
a
y
i
nto
f
oun
da
t
i
on
mo
de
ls
(
S
a
m
ba
-1)
a
n
d
e
nt
e
rpr
i
s
e
AI
s
h
ows
a
d
u
a
l
p
a
t
h
:
pow
e
r
us
e
rs
a
n
d
m
ai
nstr
ea
m
c
ustom
e
rs
a
l
ike
.
G
roq
di
st
i
n
g
u
i
s
he
s
i
ts
e
l
f
on
sp
eed
a
n
d
s
i
mpl
ici
ty
f
or
i
n
fe
r
e
n
ce
.
T
he
wor
d
“
L
a
n
g
u
age
”
i
n
LP
U
high
l
igh
ts
f
o
c
us
,
b
ut
G
roq
ʼ
s
a
r
chi
t
ec
tur
e
ca
n
acce
l
e
r
a
t
e
a
ny
t
e
nsor
-
i
n
fe
r
e
n
ce
t
a
s
k
.
I
t
ha
s
de
l
i
v
e
r
ed
on
i
ts
prom
i
s
e
o
f
high
p
e
r
f
orm
a
n
ce
(
fa
st
t
h
rou
gh
puts
w
i
t
h
low
l
a
t
e
n
c
y
(
[13]
t
i
m
e
.
c
om
))
i
n
ea
rly
de
ploym
e
nts
.
T
he
c
om
bi
n
a
t
i
on
o
f
l
a
r
ge
f
un
di
n
g
a
n
d
str
a
t
egic
be
ts
(
e
.
g
.
S
a
u
di
AI
f
un
d
(
[14]
www
.
r
e
ut
e
rs
.
c
om
))
su
gge
sts
G
roq
i
s
s
ca
l
i
n
g
i
ts
solut
i
on
be
yon
d
S
i
l
ic
on
V
a
ll
e
y
.
I
n
t
e
rms
o
f
m
a
r
ke
t
i
mp
ac
t
,
t
he
s
e
c
omp
a
n
ie
s
r
e
m
i
n
d
us
t
ha
t
a
r
chi
t
ec
tur
e
m
a
tt
e
rs
.
GP
U
s
w
e
r
e
t
he
e
n
ab
l
e
rs
o
f
t
he
fi
rst
w
a
v
e
o
f
dee
p
l
ea
rn
i
n
g
,
b
ut
f
urt
he
r
i
nnov
a
t
i
on
w
i
ll
l
ike
ly
i
nvolv
e
he
t
e
ro
ge
n
e
ous
c
o
-
de
s
ig
n
ed
pro
ce
ssors
.
I
n
d
ustry
pl
a
y
e
rs
(
i
n
c
lu
di
n
g
h
yp
e
rs
ca
l
e
rs
a
n
d
a
utomot
i
v
e
c
omp
a
n
ie
s
)
a
r
e
a
lr
ead
y
e
v
a
lu
a
t
i
n
g
or
de
ploy
i
n
g
t
he
s
e
a
lt
e
rn
a
t
i
v
e
s
to
b
r
eak
t
h
rou
gh
b
ottl
e
n
eck
s
.
A
t
t
he
s
a
m
e
t
i
m
e
,
t
he
c
omp
e
t
i
t
i
on
w
i
ll
i
nt
e
ns
if
y
:
N
v
idia
ʼ
s
n
e
xt
-
ge
n
pro
d
u
c
ts
,
a
s
w
e
ll
a
s
pot
e
nt
ia
l
o
ffe
r
i
n
g
s
f
rom
I
nt
e
l
/
H
aba
n
a
or
n
e
w
e
ntr
a
nts
(
G
r
a
p
hc
or
e
ʼ
s
n
e
w
IP
U
,
m
ea
ns
-
o
f
-
pro
d
u
c
t
i
on
chi
ps
),
w
i
ll
t
e
st
t
he
s
e
st
a
rtups
ʼ
pro
g
r
e
ss
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 14 of 19
T
he
lon
g
-
t
e
rm
w
i
nn
e
rs
a
r
e
ha
r
d
to
pr
edic
t
.
H
ow
e
v
e
r
,
t
he
su
cce
ss
o
f
F
un
daçã
o
chai
n
mo
de
ls
(
l
ike
S
a
m
ba
-1)
a
n
d
s
e
rv
ice
s
(
G
roq
C
lou
d
)
su
gge
sts
us
e
-
ca
s
e
sp
ecia
l
i
z
a
t
i
on
i
s
ke
y
.
C
e
r
eb
r
a
s
a
n
d
S
a
m
ba
N
ov
a
ha
v
e
so
fa
r
f
o
c
us
ed
on
tr
ai
n
i
n
g
t
h
rou
gh
put
f
or
str
a
t
egic
c
ustom
e
rs
,
w
he
r
ea
s
G
roq
ha
s
t
ack
l
ed
i
n
fe
r
e
n
ce
t
h
rou
gh
put
.
I
n
f
utur
e
de
ploym
e
nts
,
a
s
i
n
g
l
e
da
t
a
ce
nt
e
r
m
igh
t
us
e
a
ll
t
h
r
ee
:
C
e
r
eb
r
a
s
f
or
t
he
hea
v
ie
st
o
ff
l
i
n
e
tr
ai
n
i
n
g
,
S
a
m
ba
N
ov
a
f
or
c
ustom
e
nt
e
rpr
i
s
e
wor
k
lo
ad
s
,
a
n
d
G
roq
f
or
f
ront
-
e
n
d
us
e
r
s
e
rv
ice
s
.
I
n
c
on
c
lus
i
on
,
a
s
AI
mo
de
ls
c
ont
i
nu
e
to
g
row
a
n
d
prol
ife
r
a
t
e
,
t
he
di
v
e
rs
i
ty
o
f
ha
r
d
w
a
r
e
acce
l
e
r
a
tors
i
s
a
n
e
ss
e
nt
ia
l
tr
e
n
d
.
C
e
r
eb
r
a
s
,
S
a
m
ba
N
ov
a
,
a
n
d
G
roq
each
e
x
e
mpl
if
y
h
ow
n
e
w
a
r
chi
t
ec
tur
e
s
ca
n
r
e
s
ha
p
e
AI
i
n
f
r
a
stru
c
tur
e
.
B
y
O
c
to
be
r
2025,
t
he
y
ha
v
e
prov
e
n
t
ech
n
ica
lly
a
m
bi
t
i
ous
ca
p
abi
l
i
t
ie
s
a
n
d
ca
rv
ed
n
iche
s
i
n
t
he
AI
ec
osyst
e
m
.
T
he
c
om
i
n
g
y
ea
rs
w
i
ll
s
h
ow
h
ow
w
e
ll
t
hei
r
f
ull
-
st
ack
solut
i
ons
tr
a
nsl
a
t
e
to
w
ide
spr
ead
ad
opt
i
on
,
a
n
d
w
he
t
he
r
t
he
y
ca
n
c
o
e
x
i
st
or
e
v
e
n
c
onv
e
r
ge
i
n
t
he
AI
ha
r
d
w
a
r
e
m
a
r
ke
t
.
R
ega
r
d
l
e
ss
,
t
hei
r
i
nnov
a
t
i
ons
ha
v
e
a
lr
ead
y
i
n
f
lu
e
n
ced
i
n
d
ustry
t
hi
n
ki
n
g
on
AI
chi
ps
,
a
n
d
w
i
ll
l
ike
ly
c
ont
i
nu
e
to
d
o
so
a
s
AI
p
e
rm
ea
t
e
s
e
v
e
ry
s
ec
tor
.
R
efe
r
e
n
ce
s
R
e
ut
e
rs
–
C
e
r
eb
r
a
s
aim
s
to
de
p
l
oy
AI
i
n
f
r
a
stru
c
tur
e
f
or
ma
ss
i
v
e
S
t
a
r
ga
t
e
U
AE
da
t
a
ce
ntr
e
h
u
b
(
O
c
t
2025)
(
[6]
www
.
r
e
ut
e
rs
.
c
om
).
R
e
ut
e
rs
–
AI
chi
p
fi
r
m
C
e
r
eb
r
a
s
r
ai
s
e
s
$1
.
1
billi
on
,
add
s
T
ru
m
p
-
li
n
ked
1789
C
a
p
i
t
al
a
s
i
nv
e
stor
(
S
e
pt
2025)
(
[1]
www
.
r
e
ut
e
rs
.
c
om
).
R
e
ut
e
rs
–
C
e
r
eb
r
a
s
S
yst
em
s
,
R
a
novus
w
i
n
$45
milli
on
US
mili
t
a
ry
deal
to
sp
eed
up
chi
p
c
onn
ec
t
i
ons
(
A
pr
2025)
(
[7]
www
.
r
e
ut
e
rs
.
c
om
).
R
e
ut
e
rs
–
C
hi
p
st
a
rtup
G
roq
r
ai
s
e
s
$750
milli
on
a
t
$6
.
9
billi
on
v
al
u
a
t
i
on
(
S
e
pt
2025)
(
[2]
www
.
r
e
ut
e
rs
.
c
om
).
R
e
ut
e
rs
–
G
roq
m
or
e
t
ha
n
d
ou
ble
s
v
al
u
a
t
i
on
to
$6
.
9
billi
on
a
s
i
nv
e
stors
be
t
on
AI
chi
ps
(
S
e
pt
2025)
(
[23]
www
.
r
e
ut
e
rs
.
c
om
)
(
[14]
www
.
r
e
ut
e
rs
.
c
om
).
R
e
ut
e
rs
–
N
v
idia
unv
eil
s
fi
rst
B
lack
w
ell
chi
p
w
afe
r
made
w
i
t
h
TS
MC
i
n
US
(
O
c
t
2025)
(
[41]
www
.
r
e
ut
e
rs
.
c
om
).
T
om
ʼ
s
H
a
r
d
w
a
r
e
–
N
v
idia
AI
challe
n
ge
r
G
roq
a
nnoun
ce
s
E
urop
ea
n
e
xp
a
ns
i
on
-
H
el
s
i
n
ki
da
t
a
ce
nt
e
r
t
a
r
ge
ts
…
(
J
uly
2025)
(
[15]
www
.
toms
ha
r
d
w
a
r
e
.
c
om
).
T
i
m
e
–
J
on
a
t
ha
n
R
oss
(
pro
fi
l
e
o
f
G
roq
f
oun
de
r
,
S
e
pt
2024)
(
[13]
t
i
m
e
.
c
om
).
T
i
m
e
–
T
he
L
a
r
ge
st
-
E
v
e
r
C
hi
p
(
pro
fi
l
e
o
f
C
e
r
eb
r
a
s
W
SE
,
O
c
t
2024)
(
[5]
t
i
m
e
.
c
om
).
T
i
m
e
–
A P
ow
e
r
f
u
l
AI P
la
t
f
or
m
(
S
a
m
ba
N
ov
a
S
a
m
ba
-1
&
pl
a
t
f
orm
,
O
c
t
2024)
(
[9]
t
i
m
e
.
c
om
).
A
n
a
n
d
T
ech
–
S
amba
N
ov
a
B
r
eak
s
C
ov
e
r
:
$450
M AI
S
t
a
rtup
w
i
t
h
8-
S
o
cke
t
AI
T
r
ai
n
i
n
g
S
o
l
ut
i
ons
(
D
ec
2020)
(
[10]
www
.
a
n
a
n
d
t
ech
.
c
om
)
(
[8]
www
.
a
n
a
n
d
t
ech
.
c
om
)
(
[4]
www
.
a
n
a
n
d
t
ech
.
c
om
).
N
e
xt
P
l
a
t
f
orm
–
S
amba
N
ov
a
D
ou
ble
s
U
p
C
hi
ps
T
o
C
ha
s
e
AI F
oun
da
t
i
on
M
o
del
s
(
S
e
pt
2022)
(
[11]
www
.
n
e
xtpl
a
t
f
orm
.
c
om
).
EE
T
i
m
e
s
–
G
roq
D
em
onstr
a
t
e
s
F
a
st
LLM
s
on
4-
Y
ea
r
-
O
ld
S
ilic
on
(2024)
(
[29]
www
.
ee
t
i
m
e
s
.
c
om
).
R
e
ut
e
rs
–
B
ro
adc
o
m
hi
ts
tr
illi
on
-
d
o
lla
r
v
al
u
a
t
i
on
on
l
o
f
ty
f
or
eca
sts
f
or
AI
dema
n
d
(
D
ec
2024)
(
[16]
www
.
r
e
ut
e
rs
.
c
om
).
R
e
ut
e
rs
–
N
v
idia
or
de
rs
supp
lie
rs
to
hal
t
wor
k
on
C
hi
n
a
-
f
o
c
us
ed
H
20
AI
chi
p
(
A
u
g
2025)
–
c
ont
e
xt
on
supply
chai
ns
.
R
e
ut
e
rs
–
O
p
e
n
AI
t
a
ps
B
ro
adc
o
m
to
b
u
ild
i
ts
fi
rst
AI
pro
ce
ssor
(
O
c
t
2025)
–
c
ont
e
xt
on
AI
chi
p
tr
e
n
d
s
(
not
di
r
ec
tly
C
e
r
eb
r
a
s
/
S
a
m
ba
N
ov
a
/
G
roq
).
A
ddi
t
i
on
al
sour
ce
s
i
n
cl
u
de
v
e
n
d
or
pr
e
ss
r
elea
s
e
s
a
n
d
t
ech
n
ical
a
n
al
ys
i
s
bl
o
g
s
(
a
s
not
ed
i
n
t
e
xt
)
,
w
h
os
e
ke
y
fac
ts
a
r
e
ci
t
ed
ab
ov
e
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 15 of 19
E
xt
e
rn
a
l
S
our
ce
s
[1]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
ai
-
chi
p
-
fi
rm
-
ce
r
eb
r
a
s
-
r
ai
s
e
s
-11-
bi
ll
i
on
-
add
s
-
trump
-
l
i
n
ked
-1789-
ca
p
i
t
a
l
-
i
nv
e
stor
-20
25-09-30/
#
:~:
C
e
r
eb
...
[2]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
chi
p
-
st
a
rtup
-
g
roq
-
r
ai
s
e
s
-750-
m
i
ll
i
on
-69-
bi
ll
i
on
-
v
a
lu
a
t
i
on
-2025-09-17/
#
:~:
A
rt
if
...
[3]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
U
s
e
rs
...
[4]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
r
e
sul
...
[5]
h
ttps
://
t
i
m
e
.
c
om
/7094929/
ce
r
eb
r
a
s
-
syst
e
ms
-
w
afe
r
-
s
ca
l
e
-
e
n
gi
n
e
-3/
#
:~:
D
e
sp
i
...
[6]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
worl
d
/
m
idd
l
e
-
ea
st
/
ce
r
eb
r
a
s
-
ai
ms
-
de
ploy
-
ai
-
i
n
f
r
a
stru
c
tur
e
-
m
a
ss
i
v
e
-
st
a
r
ga
t
e
-
u
ae
-
da
t
a
-
ce
ntr
e
-
h
u
b
-2025-10-13/
#
:~:
C
e
r
eb
...
[7]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
t
ech
nolo
g
y
/
ce
r
eb
r
a
s
-
syst
e
ms
-
r
a
novus
-
w
i
n
-45-
m
i
ll
i
on
-
us
-
m
i
l
i
t
a
ry
-
dea
l
-
sp
eed
-
up
-
chi
p
-2025-
04-01/
#
:~:
C
e
r
eb
...
[8]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
r
e
t
ic
...
[9]
h
ttps
://
t
i
m
e
.
c
om
/7095052/
s
a
m
ba
nov
a
-
su
i
t
e
/
#
:~:
S
a
m
ba
...
[10]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
A
f
t
e
r
...
[11]
h
ttps
://
www
.
n
e
xtpl
a
t
f
orm
.
c
om
/2022/09/17/
s
a
m
ba
nov
a
-
d
ou
b
l
e
s
-
up
-
chi
ps
-
to
-
cha
s
e
-
ai
-
f
oun
da
t
i
on
-
mo
de
ls
/
a
mp
/
#
:~:
so
c
ke
...
[12]
h
ttps
://
www
.
ee
t
i
m
e
s
.
c
om
/
g
roq
-
de
mos
-
fa
st
-
llms
-
on
-4-
y
ea
r
-
ol
d
-
s
i
l
ic
on
/
#
:~:
i
s
%
20...
[13]
h
ttps
://
t
i
m
e
.
c
om
/7012702/
j
on
a
t
ha
n
-
ross
/
#
:~:
st
a
rt
...
[14]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
g
roq
-
mor
e
-
t
ha
n
-
d
ou
b
l
e
s
-
v
a
lu
a
t
i
on
-69-
bi
ll
i
on
-
i
nv
e
stors
-
be
t
-
ai
-
chi
ps
-2025-09-17/
#
:~:
AI
%
20...
[15]
h
ttps
://
www
.
toms
ha
r
d
w
a
r
e
.
c
om
/
t
ech
-
i
n
d
ustry
/
a
rt
ificia
l
-
i
nt
e
ll
ige
n
ce
/
nv
idia
-
ai
-
cha
ll
e
n
ge
r
-
g
roq
-
a
nnoun
ce
s
-
e
urop
ea
n
-
e
x
p
a
ns
i
on
-
he
ls
i
n
ki
-
da
t
a
-
ce
nt
e
r
-
t
a
r
ge
ts
-
b
ur
ge
on
i
n
g
-
ai
-
m
a
r
ke
t
#
:~:
G
roq
%
...
[16]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
t
ech
nolo
g
y
/
b
ro
adc
om
-
r
a
ll
ie
s
-
f
or
eca
st
-
b
oom
i
n
g
-
ai
-
chi
p
-
de
m
a
n
d
-2024-12-13/
#
:~:2024,...
[17]
h
ttps
://
www
.
t
ech
r
ada
r
.
c
om
/
pro
/
nv
idia
-
ai
-
s
a
l
e
s
-
to
-
r
each
-
a
lmost
-
us
d
400-
bi
ll
i
on
-
b
y
-2028-
c
l
ai
ms
-
r
e
s
ea
r
ch
-
a
n
d
-
t
he
n
-
t
h
i
n
g
s
-
w
i
ll
-
ge
t
-
a
-
bi
t
-
tr
ick
y
-
f
or
-
t
he
-
worl
d
s
-
l
a
r
ge
st
-
c
omp
a
ny
#
:~:2025,...
[18]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
g
roq
-
mor
e
-
t
ha
n
-
d
ou
b
l
e
s
-
v
a
lu
a
t
i
on
-69-
bi
ll
i
on
-
i
nv
e
stors
-
be
t
-
ai
-
chi
ps
-2025-09-17/
#
:~:
C
hi
p
%
...
[19]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
worl
d
/
m
idd
l
e
-
ea
st
/
ce
r
eb
r
a
s
-
ai
ms
-
de
ploy
-
ai
-
i
n
f
r
a
stru
c
tur
e
-
m
a
ss
i
v
e
-
st
a
r
ga
t
e
-
u
ae
-
da
t
a
-
ce
ntr
e
-
h
u
b
-2025-10-13/
#
:~:
G
42
%
2...
[20]
h
ttps
://
t
i
m
e
.
c
om
/7012702/
j
on
a
t
ha
n
-
ross
/
#
:~:
cha
t
b
...
[21]
h
ttps
://
www
.
toms
ha
r
d
w
a
r
e
.
c
om
/
t
ech
-
i
n
d
ustry
/
a
rt
ificia
l
-
i
nt
e
ll
ige
n
ce
/
nv
idia
-
ai
-
cha
ll
e
n
ge
r
-
g
roq
-
a
nnoun
ce
s
-
e
urop
ea
n
-
e
x
p
a
ns
i
on
-
he
ls
i
n
ki
-
da
t
a
-
ce
nt
e
r
-
t
a
r
ge
ts
-
b
ur
ge
on
i
n
g
-
ai
-
m
a
r
ke
t
#
:~:
c
usto
...
[22]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
chi
p
-
st
a
rtup
-
g
roq
-
r
ai
s
e
s
-750-
m
i
ll
i
on
-69-
bi
ll
i
on
-
v
a
lu
a
t
i
on
-2025-09-17/
#
:~:
G
roq
%
...
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 16 of 19
[23]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
g
roq
-
mor
e
-
t
ha
n
-
d
ou
b
l
e
s
-
v
a
lu
a
t
i
on
-69-
bi
ll
i
on
-
i
nv
e
stors
-
be
t
-
ai
-
chi
ps
-2025-09-17/
#
:~:
C
hi
p
%
...
[24]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
g
roq
-
mor
e
-
t
ha
n
-
d
ou
b
l
e
s
-
v
a
lu
a
t
i
on
-69-
bi
ll
i
on
-
i
nv
e
stors
-
be
t
-
ai
-
chi
ps
-2025-09-17/
#
:~:
p
a
rt
i
...
[25]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
chi
p
-
st
a
rtup
-
g
roq
-
r
ai
s
e
s
-750-
m
i
ll
i
on
-69-
bi
ll
i
on
-
v
a
lu
a
t
i
on
-2025-09-17/
#
:~:
pl
a
y
e
...
[26]
h
ttps
://
www
.
ce
r
eb
r
a
s
.
n
e
t
/
b
lo
g
/
ce
r
eb
r
a
s
-
w
afe
r
-
s
ca
l
e
-
e
n
gi
n
e
-
outp
e
r
f
orms
-
nv
idia
-
h
100-
i
n
-
ca
r
b
on
-
ca
ptur
e
-
s
i
mul
a
t
i
ons
#
:~:
C
e
r
eb
...
[27]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
m
a
t
ch
...
[28]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
T
e
nso
...
[29]
h
ttps
://
www
.
ee
t
i
m
e
s
.
c
om
/
g
roq
-
de
mos
-
fa
st
-
llms
-
on
-4-
y
ea
r
-
ol
d
-
s
i
l
ic
on
/
#
:~:
T
i
m
e
s
...
[30]
h
ttps
://
g
roq
.
c
om
/
n
e
wsroom
/
de
m
a
n
d
-
f
or
-
r
ea
l
-
t
i
m
e
-
ai
-
i
n
fe
r
e
n
ce
-
f
rom
-
g
roq
-
acce
l
e
r
a
t
e
s
-
w
eek
-
ov
e
r
-
w
eek
#
:~:,
t
ha
n
...
[31]
h
ttps
://
www
.
toms
ha
r
d
w
a
r
e
.
c
om
/
t
ech
-
i
n
d
ustry
/
a
rt
ificia
l
-
i
nt
e
ll
ige
n
ce
/
nv
idia
-
ai
-
cha
ll
e
n
ge
r
-
g
roq
-
a
nnoun
ce
s
-
e
urop
ea
n
-
e
x
p
a
ns
i
on
-
he
ls
i
n
ki
-
da
t
a
-
ce
nt
e
r
-
t
a
r
ge
ts
-
b
ur
ge
on
i
n
g
-
ai
-
m
a
r
ke
t
#
:~:
GP
U
s
,...
[32]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
m
a
t
ch
...
[33]
h
ttps
://
www
.
n
e
xtpl
a
t
f
orm
.
c
om
/2022/09/17/
s
a
m
ba
nov
a
-
d
ou
b
l
e
s
-
up
-
chi
ps
-
to
-
cha
s
e
-
ai
-
f
oun
da
t
i
on
-
mo
de
ls
/
a
mp
/
#
:~:
m
a
t
ch
...
[34]
h
ttps
://
www
.
ee
t
i
m
e
s
.
c
om
/
g
roq
-
de
mos
-
fa
st
-
llms
-
on
-4-
y
ea
r
-
ol
d
-
s
i
l
ic
on
/
#
:~:
T
i
m
e
s
...
[35]
h
ttps
://
s
a
m
ba
nov
a
.
ai
/
pr
e
ss
/
worl
d
s
-
fa
st
e
st
-
ai
-
pl
a
t
f
orm
#
:~:
S
a
m
ba
...
[36]
h
ttps
://
www
.
a
n
a
n
d
t
ech
.
c
om
/
s
h
ow
/16286/
s
a
m
ba
nov
a
-
b
r
eak
s
-
c
ov
e
r
-450
m
-
ai
-
st
a
rtup
-
w
i
t
h
-8
so
cke
t
-
ai
-
tr
ai
n
i
n
g
-
solut
i
on
s
-
a
n
d
-
mor
e
#
:~:
w
hich
...
[37]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
t
ech
nolo
g
y
/
ce
r
eb
r
a
s
-
syst
e
ms
-
r
a
novus
-
w
i
n
-45-
m
i
ll
i
on
-
us
-
m
i
l
i
t
a
ry
-
dea
l
-
sp
eed
-
up
-
chi
p
-2025-
04-01/
#
:~:
C
e
r
eb
...
[38]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
m
edia
-
t
e
l
ec
om
/
fi
rst
-200-
mw
-
u
ae
s
-
st
a
r
ga
t
e
-
ai
-
ca
mpus
-
c
om
e
-
onl
i
n
e
-
n
e
xt
-
y
ea
r
-202
5-10-14/
#
:~:2025,...
[39]
h
ttps
://
t
echc
run
ch
.
c
om
/2018/03/15/
t
he
-
r
ed
-
h
ot
-
ai
-
chi
p
-
sp
ace
-
ge
ts
-
e
v
e
n
-
h
ott
e
r
-
w
i
t
h
-56
m
-
f
or
-
a
-
st
a
rtup
-
ca
ll
ed
-
s
a
m
b
a
nov
a
/
#
:~:
c
omp
a
...
[40]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
worl
d
/
chi
n
a
/
u
ae
-
s
e
t
-
dee
p
e
n
-
ai
-
l
i
n
k
s
-
w
i
t
h
-
un
i
t
ed
-
st
a
t
e
s
-
af
t
e
r
-
p
a
st
-
c
ur
b
s
-
ov
e
r
-
chi
n
a
-2025-0
5-15/
#
:~:
to
%
20...
[41]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
t
ech
nolo
g
y
/
ce
r
eb
r
a
s
-
syst
e
ms
-
fi
l
e
s
-
w
i
t
hd
r
a
w
-
us
-
i
po
-2025-10-03/
#
:~:
l
i
st
i
...
[42]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
t
ech
nolo
g
y
/
nv
idia
-
sp
e
nt
-
ov
e
r
-900-
m
i
ll
i
on
-
hi
r
e
-
e
n
fab
r
ica
-
ce
o
-
l
ice
ns
e
-
t
ech
nolo
g
y
-
c
n
bc
-
r
e
por
ts
-2025-09-18/
#
:~:2025,...
[43]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
b
us
i
n
e
ss
/
m
edia
-
t
e
l
ec
om
/
h
u
a
w
ei
-
unv
ei
ls
-
chi
pm
aki
n
g
-
c
omput
i
n
g
-
pow
e
r
-
pl
a
ns
-
fi
rst
-
t
i
m
e
-2025
-09-18/
#
:~:
H
u
a
w
e
...
[44]
h
ttps
://
www
.
r
e
ut
e
rs
.
c
om
/
t
ech
nolo
g
y
/
a
rt
ificia
l
-
i
nt
e
ll
ige
n
ce
/
a
lp
habe
t
-
nv
idia
-
i
nv
e
st
-
op
e
n
ai
-
c
o
-
f
oun
de
r
-
suts
ke
v
e
rs
-
ss
i
-
so
ur
ce
-
s
a
ys
-2025-04-12/
#
:~:
sour
c
...
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 17 of 19
I
ntu
i
t
i
on
L
ab
s
-
I
n
d
ustry
L
eade
rs
hi
p
&
S
e
rv
ice
s
N
ort
h
A
m
e
r
ica
'
s
#
1
AI S
o
f
tw
a
r
e
D
e
v
e
lopm
e
nt
F
i
rm
f
or
P
ha
rm
ace
ut
ica
l
&
B
i
ot
ech
:
I
ntu
i
t
i
on
L
ab
s
l
ead
s
t
he
U
S
m
a
r
ke
t
i
n
c
ustom
AI
so
f
tw
a
r
e
de
v
e
lopm
e
nt
a
n
d
p
ha
rm
a
i
mpl
e
m
e
nt
a
t
i
ons
w
i
t
h
prov
e
n
r
e
sults
ac
ross
pu
b
l
ic
bi
ot
ech
a
n
d
p
ha
rm
ace
ut
ica
l
c
omp
a
n
ie
s
.
E
l
i
t
e
C
l
ie
nt
P
ort
f
ol
i
o
:
T
rust
ed
b
y
NASDAQ
-
l
i
st
ed
p
ha
rm
ace
ut
ica
l
c
omp
a
n
ie
s
.
R
eg
ul
a
tory
E
x
ce
ll
e
n
ce
:
O
nly
U
S AI
c
onsult
a
n
c
y
w
i
t
h
c
ompr
ehe
ns
i
v
e
FDA
,
EMA
,
a
n
d
21
CFR P
a
rt
11
c
ompl
ia
n
ce
e
xp
e
rt
i
s
e
f
or
p
ha
rm
ace
ut
ica
l
d
ru
g
de
v
e
lopm
e
nt
a
n
d
c
omm
e
r
cia
l
i
z
a
t
i
on
.
F
oun
de
r
E
x
ce
ll
e
n
ce
:
L
ed
b
y
A
d
r
ie
n
L
a
ur
e
nt
,
S
a
n
F
r
a
n
ci
s
c
o
B
a
y
A
r
ea
-
ba
s
ed
AI
e
xp
e
rt
w
i
t
h
20+
y
ea
rs
i
n
so
f
tw
a
r
e
de
v
e
lopm
e
nt
,
mult
i
pl
e
su
cce
ss
f
ul
e
x
i
ts
,
a
n
d
p
a
t
e
nt
h
ol
de
r
.
R
ec
o
g
n
i
z
ed
a
s
on
e
o
f
t
he
top
AI
e
xp
e
rts
i
n
t
he
U
SA
.
C
ustom
AI S
o
f
tw
a
r
e
D
e
v
e
lopm
e
nt
:
B
u
i
l
d
t
ai
lor
ed
p
ha
rm
ace
ut
ica
l
AI
a
ppl
ica
t
i
ons
,
c
ustom
CRM
s
,
cha
t
b
ots
,
a
n
d
ERP
syst
e
ms
w
i
t
h
ad
v
a
n
ced
a
n
a
lyt
ic
s
a
n
d
r
eg
ul
a
tory
c
ompl
ia
n
ce
ca
p
abi
l
i
t
ie
s
.
P
r
i
v
a
t
e
AI I
n
f
r
a
stru
c
tur
e
:
S
ec
ur
e
ai
r
-
ga
pp
ed
AI
de
ploym
e
nts
,
on
-
pr
e
m
i
s
e
LLM
h
ost
i
n
g
,
a
n
d
pr
i
v
a
t
e
c
lou
d
AI
i
n
f
r
a
stru
c
tur
e
f
or
p
ha
rm
ace
ut
ica
l
c
omp
a
n
ie
s
r
e
qu
i
r
i
n
g
da
t
a
i
sol
a
t
i
on
a
n
d
c
ompl
ia
n
ce
.
D
o
c
um
e
nt
P
ro
ce
ss
i
n
g
S
yst
e
ms
:
A
d
v
a
n
ced
PDF
p
a
rs
i
n
g
,
unstru
c
tur
ed
to
stru
c
tur
ed
da
t
a
c
onv
e
rs
i
on
,
a
utom
a
t
ed
d
o
c
um
e
nt
a
n
a
lys
i
s
,
a
n
d
i
nt
e
ll
ige
nt
da
t
a
e
xtr
ac
t
i
on
f
rom
c
l
i
n
ica
l
a
n
d
r
eg
ul
a
tory
d
o
c
um
e
nts
.
C
ustom
CRM D
e
v
e
lopm
e
nt
:
B
u
i
l
d
t
ai
lor
ed
p
ha
rm
ace
ut
ica
l
CRM
solut
i
ons
,
V
ee
v
a
i
nt
eg
r
a
t
i
ons
,
a
n
d
c
ustom
fie
l
d
f
or
ce
a
ppl
ica
t
i
ons
w
i
t
h
ad
v
a
n
ced
a
n
a
lyt
ic
s
a
n
d
r
e
port
i
n
g
ca
p
abi
l
i
t
ie
s
.
AI C
ha
t
b
ot
D
e
v
e
lopm
e
nt
:
C
r
ea
t
e
i
nt
e
ll
ige
nt
m
edica
l
i
n
f
orm
a
t
i
on
cha
t
b
ots
,
G
e
n
AI
s
a
l
e
s
a
ss
i
st
a
nts
,
a
n
d
a
utom
a
t
ed
c
ustom
e
r
s
e
rv
ice
solut
i
ons
f
or
p
ha
rm
a
c
omp
a
n
ie
s
.
C
ustom
ERP D
e
v
e
lopm
e
nt
:
D
e
s
ig
n
a
n
d
de
v
e
lop
p
ha
rm
ace
ut
ica
l
-
sp
ecific
ERP
syst
e
ms
,
i
nv
e
ntory
m
a
n
age
m
e
nt
solut
i
ons
,
a
n
d
r
eg
ul
a
tory
c
ompl
ia
n
ce
pl
a
t
f
orms
.
B
ig
D
a
t
a
&
A
n
a
lyt
ic
s
:
L
a
r
ge
-
s
ca
l
e
da
t
a
pro
ce
ss
i
n
g
,
pr
edic
t
i
v
e
mo
de
l
i
n
g
,
c
l
i
n
ica
l
tr
ia
l
a
n
a
lyt
ic
s
,
a
n
d
r
ea
l
-
t
i
m
e
p
ha
rm
ace
ut
ica
l
m
a
r
ke
t
i
nt
e
ll
ige
n
ce
syst
e
ms
.
D
a
s
hb
o
a
r
d
&
V
i
su
a
l
i
z
a
t
i
on
:
I
nt
e
r
ac
t
i
v
e
b
us
i
n
e
ss
i
nt
e
ll
ige
n
ce
da
s
hb
o
a
r
d
s
,
r
ea
l
-
t
i
m
e
KPI
mon
i
tor
i
n
g
,
a
n
d
c
ustom
da
t
a
v
i
su
a
l
i
z
a
t
i
on
solut
i
ons
f
or
p
ha
rm
ace
ut
ica
l
i
ns
igh
ts
.
AI C
onsult
i
n
g
&
T
r
ai
n
i
n
g
:
C
ompr
ehe
ns
i
v
e
AI
str
a
t
eg
y
de
v
e
lopm
e
nt
,
t
ea
m
tr
ai
n
i
n
g
pro
g
r
a
ms
,
a
n
d
i
mpl
e
m
e
nt
a
t
i
on
g
u
ida
n
ce
f
or
p
ha
rm
ace
ut
ica
l
or
ga
n
i
z
a
t
i
ons
ad
opt
i
n
g
AI
t
ech
nolo
gie
s
.
C
ont
ac
t
f
oun
de
r
A
d
r
ie
n
L
a
ur
e
nt
a
n
d
t
ea
m
a
t
h
ttps
://
i
ntu
i
t
i
onl
ab
s
.
ai
/
c
ont
ac
t
f
or
a
c
onsult
a
t
i
on
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 18 of 19
DISCLAIMER
T
he
i
n
f
orm
a
t
i
on
c
ont
ai
n
ed
i
n
t
hi
s
d
o
c
um
e
nt
i
s
prov
ided
f
or
ed
u
ca
t
i
on
a
l
a
n
d
i
n
f
orm
a
t
i
on
a
l
purpos
e
s
only
.
W
e
m
ake
no
r
e
pr
e
s
e
nt
a
t
i
ons
or
w
a
rr
a
nt
ie
s
o
f
a
ny
ki
n
d
,
e
xpr
e
ss
or
i
mpl
ied
,
ab
out
t
he
c
ompl
e
t
e
n
e
ss
,
acc
ur
ac
y
,
r
e
l
iabi
l
i
ty
,
su
i
t
abi
l
i
ty
,
or
a
v
ai
l
abi
l
i
ty
o
f
t
he
i
n
f
orm
a
t
i
on
c
ont
ai
n
ed
he
r
ei
n
.
A
ny
r
e
l
ia
n
ce
you
pl
ace
on
su
ch
i
n
f
orm
a
t
i
on
i
s
str
ic
tly
a
t
your
own
r
i
s
k
.
I
n
no
e
v
e
nt
w
i
ll
I
ntu
i
t
i
on
L
ab
s
.
ai
or
i
ts
r
e
pr
e
s
e
nt
a
t
i
v
e
s
be
l
iab
l
e
f
or
a
ny
loss
or
da
m
age
i
n
c
lu
di
n
g
w
i
t
h
out
l
i
m
i
t
a
t
i
on
,
i
n
di
r
ec
t
or
c
ons
e
qu
e
nt
ia
l
loss
or
da
m
age
,
or
a
ny
loss
or
da
m
age
w
ha
tso
e
v
e
r
a
r
i
s
i
n
g
f
rom
t
he
us
e
o
f
i
n
f
orm
a
t
i
on
pr
e
s
e
nt
ed
i
n
t
hi
s
d
o
c
um
e
nt
.
T
hi
s
d
o
c
um
e
nt
m
a
y
c
ont
ai
n
c
ont
e
nt
ge
n
e
r
a
t
ed
w
i
t
h
t
he
a
ss
i
st
a
n
ce
o
f
a
rt
ificia
l
i
nt
e
ll
ige
n
ce
t
ech
nolo
gie
s
.
AI
-
ge
n
e
r
a
t
ed
c
ont
e
nt
m
a
y
c
ont
ai
n
e
rrors
,
om
i
ss
i
ons
,
or
i
n
acc
ur
acie
s
.
R
eade
rs
a
r
e
ad
v
i
s
ed
to
i
n
de
p
e
n
de
ntly
v
e
r
if
y
a
ny
c
r
i
t
ica
l
i
n
f
orm
a
t
i
on
bef
or
e
ac
t
i
n
g
upon
i
t
.
A
ll
pro
d
u
c
t
n
a
m
e
s
,
lo
g
os
,
b
r
a
n
d
s
,
tr
ade
m
a
r
k
s
,
a
n
d
r
egi
st
e
r
ed
tr
ade
m
a
r
k
s
m
e
nt
i
on
ed
i
n
t
hi
s
d
o
c
um
e
nt
a
r
e
t
he
prop
e
rty
o
f
t
hei
r
r
e
sp
ec
t
i
v
e
own
e
rs
.
A
ll
c
omp
a
ny
,
pro
d
u
c
t
,
a
n
d
s
e
rv
ice
n
a
m
e
s
us
ed
i
n
t
hi
s
d
o
c
um
e
nt
a
r
e
f
or
ide
nt
ifica
t
i
on
purpos
e
s
only
.
U
s
e
o
f
t
he
s
e
n
a
m
e
s
,
lo
g
os
,
tr
ade
m
a
r
k
s
,
a
n
d
b
r
a
n
d
s
d
o
e
s
not
i
mply
e
n
d
ors
e
m
e
nt
b
y
t
he
r
e
sp
ec
t
i
v
e
tr
ade
m
a
r
k
h
ol
de
rs
.
I
ntu
i
t
i
on
L
ab
s
.
ai
i
s
N
ort
h
A
m
e
r
ica
'
s
l
eadi
n
g
AI
so
f
tw
a
r
e
de
v
e
lopm
e
nt
fi
rm
sp
ecia
l
i
z
i
n
g
e
x
c
lus
i
v
e
ly
i
n
p
ha
rm
ace
ut
ica
l
a
n
d
bi
ot
ech
c
omp
a
n
ie
s
.
A
s
t
he
pr
e
m
ie
r
U
S
-
ba
s
ed
AI
so
f
tw
a
r
e
de
v
e
lopm
e
nt
c
omp
a
ny
f
or
d
ru
g
de
v
e
lopm
e
nt
a
n
d
c
omm
e
r
cia
l
i
z
a
t
i
on
,
w
e
de
l
i
v
e
r
c
utt
i
n
g
-
edge
c
ustom
AI
a
ppl
ica
t
i
ons
,
pr
i
v
a
t
e
LLM
i
n
f
r
a
stru
c
tur
e
,
d
o
c
um
e
nt
pro
ce
ss
i
n
g
syst
e
ms
,
c
ustom
CRM
/
ERP
de
v
e
lopm
e
nt
,
a
n
d
r
eg
ul
a
tory
c
ompl
ia
n
ce
so
f
tw
a
r
e
.
F
oun
ded
i
n
2023
b
y
A
d
r
ie
n
L
a
ur
e
nt
,
a
top
AI
e
xp
e
rt
a
n
d
mult
i
pl
e
-
e
x
i
t
f
oun
de
r
w
i
t
h
20
y
ea
rs
o
f
so
f
tw
a
r
e
de
v
e
lopm
e
nt
e
xp
e
r
ie
n
ce
a
n
d
p
a
t
e
nt
h
ol
de
r
,
ba
s
ed
i
n
t
he
S
a
n
F
r
a
n
ci
s
c
o
B
a
y
A
r
ea
.
T
hi
s
d
o
c
um
e
nt
d
o
e
s
not
c
onst
i
tut
e
pro
fe
ss
i
on
a
l
or
l
ega
l
ad
v
ice
.
F
or
sp
ecific
g
u
ida
n
ce
r
e
l
a
t
ed
to
your
b
us
i
n
e
ss
n
eed
s
,
pl
ea
s
e
c
onsult
w
i
t
h
a
ppropr
ia
t
e
qu
a
l
ified
pro
fe
ss
i
on
a
ls
.
©
2025
I
ntu
i
t
i
on
L
ab
s
.
ai
.
A
ll
r
igh
ts
r
e
s
e
rv
ed
.
IntuitionLabs - Custom AI Software Development
from the leading AI expert Adrien Laurent Cerebras vs SambaNova vs Groq: AI Chip Comparison (2025)
© 2025 IntuitionLabs.ai - North America's Leading AI Software Development Firm for Pharmaceutical & Biotech. All rights reserved. Page 19 of 19