AMD Radeon Instinct MI250 vs NVIDIA H200 SXM 141 GB

Comparison of AMD Radeon Instinct MI250 with 128 GB HBM2e and 13,312 cores vs NVIDIA H200 SXM 141 GB with 141 GB HBM3e and 16,896 cores.

Loading...

Performance Rating

NVIDIA H200 SXM 141 GB outperforms AMD Radeon Instinct MI250 by 32.4% in the overall GPU ARK performance rating

A100 A100
H200 H200
MI325X MI325X

AMD Radeon Instinct MI250

50.9

AMD Radeon Instinct MI250

50.9
RX 7900 XTX RX 7900 XTX
MI250 MI250
Instinct MI300X Instinct MI300X

NVIDIA H200 SXM 141 GB

67.4

NVIDIA H200 SXM 141 GB

67.4

Expert Comparison

AMD Radeon Instinct MI250 лучше по цене и энергоэффективности благодаря более низкому TDP и доступной памяти (128 ГБ против 141 ГБ). Однако NVIDIA H200 SXM с архитектурой Hopper имеет большее количество ядер (16896 против 13312), большую пропускную способность (4.89 TB/s против 3.28 TB/s) и выше производительность в FP32 (66.91 TFLOPS против 45.26 TFLOPS), что делает его предпочтительным для требовательных вычислительных задач.

Contents:

Memory ML Performance Compute Power Architecture & Compatibility ML Software Support Clocks & Performance Power Consumption Rendering Benchmarks Additional

Memory

Memory Size

128 ГБ
🔥 +10% 141 ГБ

Memory Type

HBM2e HBM3e

Memory Bandwidth

3.28 TB/s
🔥 +49% 4.89 TB/s

Memory Bus Width

8,192 бит 6,144 бит

ML Performance

FP16 (Half Precision)

🔥 +35% 362.1 TFLOPS
267.6 TFLOPS

BF16 (Brain Float)

No No

TF32 (TensorFloat)

No No

Compute Power

FP32 (Single Precision)

45.26 TFLOPS
🔥 +48% 66.91 TFLOPS

FP64 (Double Precision)

🔥 +35% 45.26 TFLOPS
33.45 TFLOPS

CUDA Cores

13,312
🔥 +27% 16,896

RT Cores

No No

Architecture & Compatibility

GPU Architecture

CDNA 2.0 Hopper

SM (Streaming Multiprocessor)

No
🔥 132

PCIe Version

PCIe 4.0 x16 PCIe 5.0 x16

ML Software Support

CUDA Version

No 9.0

Clocks & Performance

Base Clock

1,000
🔥 +50% 1,500

Boost Clock

1,700
🔥 +16% 1,980

Memory Clock

🔥 1,600
1,593

Power Consumption

TDP/TGP

🔥 -29% 500 W
700 W

Recommended PSU

🔥 -18% 900 W
1100 W

Power Connector

2x 8-pin 8-pin EPS

Rendering

Texture Units (TMU)

🔥 +58% 832
528

ROP

No No

L2 Cache

16 MB
🔥 +212% 50 MB

Benchmarks

MLPerf, llama2-70b-99.9 (UNSET)

3 534 tokens/s

MLPerf, llama2-70b-99.9 (fp16)

3 553 tokens/s

MLPerf, llama2-70b-99.9 (fp8)

2 444 tokens/s

MLPerf, llama3.1-405b (fp16)

40.8 tokens/s

MLPerf, llama3.1-405b (fp8)

25.3 tokens/s

MLPerf, llama3.1-8b (fp8)

5 161 tokens/s

llama.cpp, llama-2-7b-Q4_0

63.9 tokens/s

MLPerf, deepseek-r1 (fp8)

1 113 tokens/s

MLPerf, mixtral-8x7b (fp8)

7 132 tokens/s

Additional

Slots

Dual-slot
🔥 SXM Module

Release Date

Nov. 8, 2021 Nov. 18, 2024

Display Outputs

No outputs
No outputs

Renting is cheaper than buying