Compare XPUs

Select up to 5 XPUs to compare side-by-side

Select XPUs to Compare

Clear all (4)Jump to results

Filter by Vendor

Showing 72 XPUs • 4 selected

Alibaba

Hanguang 800

AMD

MI100

23.1 TFLOPs

AMD

MI210

181 TFLOPs

AMD

MI250X

383 TFLOPs

AMD

MI300A

980.6 TFLOPs

AMD

MI300X

1,307 TFLOPs

AMD

MI325X

1,400 TFLOPs

AMD

MI350X

2,100 TFLOPs

AMD

Radeon PRO W7900

122 TFLOPs

AMD

Radeon RX 7900 XT

104 TFLOPs

AMD

Radeon RX 7900 XTX

122 TFLOPs

AWS

Inferentia2

190 TFLOPs

AWS

Trainium

190 TFLOPs

AWS

Trainium2

680 TFLOPs

Baidu

Kunlun II

Biren Technology

BR100

Cambricon

MLU370

256 TFLOPs

Cerebras

WSE-3

Enflame Technology

CloudBlazer T20

Etched

Sohu

10,000 TFLOPs

FuriosaAI

Warboy

Google

TPU v4

275 TFLOPs

Google

TPU v5e

197 TFLOPs

Google

TPU v5p

459 TFLOPs

Google

TPU v6e (Trillium)

918 TFLOPs

Graphcore

Bow IPU

Graphcore

IPU-M2000

Groq

LPU Inference Engine

Huawei

Ascend 910B

Iluvatar CoreX

BI-V150

300 TFLOPs

Intel

Data Center GPU Max 1100

177 TFLOPs

Intel

Data Center GPU Max 1550

419 TFLOPs

Intel Habana

Gaudi 2

432 TFLOPs

Intel Habana

Gaudi 3

1,835 TFLOPs

Multi-Metric Comparison

Relative performance across 5 key metrics (normalized to 100 = best in comparison)

Compute Performance (BF16)

Memory Capacity

Power Consumption

Power Efficiency

Specifications

Specification	NVIDIA L40S	NVIDIA L4	Google TPU v4	Huawei Ascend 910B
Architecture	Ada Lovelace	Ada Lovelace	TPU v4	Da Vinci
Form Factor	PCIe	PCIe	Mezzanine	—
VRAM	48 GB	24 GB	—	64 GB
Memory Bandwidth	864 GB/s	300 GB/s	—	—
TFLOPs (FP32)	91.6	30.3	—	—
TFLOPs (FP16)	733	242	—	640
TFLOPs	733	121	275	—
TFLOPs (FP8)	1,466	485	—	—
TDP	350 W	72 W	300 W	400 W
Launch Date	Oct 2023	Mar 2023	May 2021	Aug 2023

Efficiency Metrics

Metric	L40S	L4	TPU v4	Ascend 910B
TFLOPs per Watt (FP32-eq)	1.05	0.84	0.46	—
Memory Bandwidth per GB	18.0 GB/s	12.5 GB/s	—	—

Performance Equivalence

How many units of each GPU are needed to match the performance of the others?

To match 1x NVIDIA L40S

NVIDIA L4

Compute (FP32-eq)

6.06x

Need 6.06x L4

FP32 Compute

3.02x

Need 3.02x L4

VRAM

2.00x

Need 2.00x L4

Memory Bandwidth

2.88x

Need 2.88x L4

Google TPU v4

Compute (FP32-eq)

2.67x

Need 2.67x TPU v4

Huawei Ascend 910B

VRAM

0.75x

Ascend 910B has 1.33x more

To match 1x NVIDIA L4

NVIDIA L40S

Compute (FP32-eq)

0.17x

L40S is 6.06x faster

FP32 Compute

0.33x

L40S is 3.02x faster

VRAM

0.50x

L40S has 2.00x more

Memory Bandwidth

0.35x

L40S has 2.88x more

Google TPU v4

Compute (FP32-eq)

0.44x

TPU v4 is 2.27x faster

Huawei Ascend 910B

VRAM

0.38x

Ascend 910B has 2.67x more

To match 1x Google TPU v4

NVIDIA L40S

Compute (FP32-eq)

0.38x

L40S is 2.67x faster

NVIDIA L4

Compute (FP32-eq)

2.27x

Need 2.27x L4

To match 1x Huawei Ascend 910B

NVIDIA L40S

VRAM

1.33x

Need 1.33x L40S

NVIDIA L4

VRAM

2.67x

Need 2.67x L4

Pricing

Price Type	L40S	L4	TPU v4	Ascend 910B
CAPEX (Street Price)	$10,000	$4,000	—	—
OPEX (per hour)	$1.50/hr	$0.80/hr	$3.00/hr	—
Price per TFLOPs (FP32-eq)	$27	$66	—	—