Page 10 - GPUサーバ製品紹介

GPU仕様比較

H200 NVL H100 NVL H100 PCIe A100 80GB PCIe A100 40GB PCIe

FP32 CUDA Cores 2 x 19896 2 x 14592 14592 6912 6912
Boost Clock 1.78GHz 1.98GHz? 1.75GHz 1.41GHz 1.41GHz
Memory Clock 6.3Gbps HBM3e ~5.1Gbps HBM3 3.2Gbps HBM2e 3.0 Gbps HBM2 2.43Gbps HBM2

Memory Bus Width 5120-bit 6144-bit 5120-bit 5120-bit 5120-bit

Memory Bandwidth 2 x 4.8TB/sec 2 x 3.9TB/sec 2TB/sec 1.9TB/sec 1.6TB/sec
VRAM 2 x 141GB (282GB) 2 x 94GB (188GB) 80GB 80GB 40GB
FP32 Vector (単精度) 2 x 67 TFLOPS 2 x 67 TFLOPS? 51 TFLOPS 19.5 TFLOPs 19.5 TFLOPs

FP64 Vector (倍精度) 2 x 34 TFLOPS 2 x 34 TFLOPS? 26 TFLOPS 9.7 TFLOPs 9.7 TFLOPs

INT8 Tensor 2 x 3341 TFLOPS 2 x 1980 TOPS 1513 TOPS (1/2 FP32 rate) (1/2 FP32 rate)
FP16 Tensor 2 x 1671 TFLOPS 2 x 990 TFLOPS 756 TFLOPS 624 TOPs 624 TOPs

TF32 Tensor 2 x 835 TFLOPS 2 x 495 TFLOPS 378 TFLOPS 312 TFLOPs 312 TFLOPs
FP64 Tensor 2 x 60 TFLOPS 2 x 67 TFLOPS? 51 TFLOPS 156 TFLOPs 156 TFLOPs

NVLink 4 NVLink 4 NVLink 3 NVLink 3
Interconnect NVLink 4 (600GB/sec)
(900GB/sec) (600GB/sec) (600GB/sec) (600GB/sec)
GPU 2 x H200 2 x GH100 GH100 GA100 GA100
Transistor Count 2 x 80B 2 x 80B 80B 54.2B 54.2B

TDP 600W 700-800W 350W 300W 250W
Manufacturing Process TSMC 4N TSMC 4N TSMC 4N TSMC 7N TSMC 7N

2 x PCIe 5.0 (Quad 2 x PCIe 5.0 (Quad
Interface PCIe 5.0 (Dual Slot) PCIe 4.0 (Dual Slot) PCIe 4.0 (Dual Slot)
Slot) Slot)
Architecture Hopper Hopper Hopper Ampere Ampere

スケーラブルシステムズ株式会社