| 规格 | NVIDIA H100 SXM | NVIDIA A30 |
|---|
| 架构 | NVIDIA Hopper(GH100核心,TSMC 4N定制工艺,800亿晶体管,814mm²,Compute Capability 9.0) | NVIDIA Ampere |
| 显存 | 80GB HBM3(SXM5口径,5颗HBM3堆栈);PCIe版为80GB HBM2e(5堆栈,2TB/s) | 24 GB HBM2 |
| 显存位宽 | 5120bit(SXM5与PCIe版均为10个512bit内存控制器) | — |
| 显存带宽 | 3.35TB/s(SXM5);PCIe版2TB/s | 933 GB/s |
| 算力 | FP8 Tensor Core 3,958 TFLOPS(含稀疏,密集1,979);FP16/BF16 TC 1,979*;TF32 TC 989*;FP64 TC 67;FP32 67;FP64 34(SXM5口径,*=With sparsity) | — |
| 功耗 | SXM5最高700W(可配置);PCIe版350W | — |
| 互联 | NVLink 4.0 900GB/s(SXM5,18条链路)+PCIe Gen5 x16 128GB/s;PCIe版:PCIe Gen5 128GB/s,可选双卡NVLink桥接600GB/s | PCIe Gen4: 64GB/s / Third-gen NVLINK: 200GB/s |
| 发布年 | 2022(2022-03-22 GTC发布,SXM5 Q3 2022上市) | — |
| SM与CUDA核 | 132 SM / 16,896 FP32 CUDA核(SXM5);PCIe版114 SM / 14,592 | — |
| 加速频率 | SXM5:1830MHz(FP8/FP16/BF16/TF32 TC)/1980MHz(FP64 TC与FP32/FP64);PCIe版:1620/1755MHz | — |
| L2缓存 | 50MB | — |
| MIG切片 | 第二代Secure MIG,最多7实例@10GB | — |
| 编解码器 | 7 NVDEC + 7 NVJPG | — |
| 形态 | SXM5模组(液冷/风冷由整机定,配HGX/DGX H100 4或8卡);PCIe版双槽风冷 | Dual-slot, full-height, full-length (FHFL) |
| FP32算力 | — | 10.3 teraFLOPS |
| FP64算力 | — | 5.2 teraFLOPS |
| TF32算力 | — | 82 teraFLOPS | 165 teraFLOPS |
| FP16算力 | — | 165 teraFLOPS | 330 teraFLOPS |
| INT8算力 | — | 330 TOPS | 661 TOPS |
| 功耗TDP | — | 165 W |
| MIG | — | 4 GPU instances @ 6GB each |