| 规格 | NVIDIA H200 NVL | NVIDIA H20 |
|---|
| 架构 | NVIDIA Hopper | NVIDIA Hopper |
| 显存 | 141GB | 96GB HBM3(Flopper/VipraTech/腾讯新闻三源一致) |
| 显存类型 | HBM3e | — |
| 显存带宽 | 4.8TB/s | 4.0TB/s |
| FP64算力 | 30 TFLOPS | — |
| FP64 Tensor Core | 60 TFLOPS | — |
| FP32算力 | 60 TFLOPS | — |
| FP8算力 | 3,341 TFLOPS | — |
| FP16算力 | 1,671 TFLOPS | — |
| INT8算力 | 3,341 TFLOPS | — |
| 功耗TDP | Up to 600W (configurable) | — |
| 互联 | 2- or 4-way NVIDIA NVLink bridge: 900GB/s per GPU | PCIe Gen5: 128GB/s | NVLink 900GB/s+PCIe Gen5 x16 128GB/s |
| 形态 | PCIe Dual-slot air-cooled | HGX 8路SXM基板模组(Flopper'SXM'/VipraTech'8-way HGX'/腾讯新闻'HGX方案8路GPU配置+H20 SXM';同批L20/L2为PCIe形态) |
| MIG切片 | Up to 7 MIGs @16.5GB each | 最多7实例(VipraTech+腾讯新闻'7路MIG'双源) |
| 编解码器 | 7 NVDEC | 7 JPEG | 7 NVDEC + 7 NVJPEG(单源VipraTech) |
| 机密计算 | Supported | — |
| 算力 | — | FP8 Tensor Core 296 TFLOPS(密集)/592 TFLOPS(2:4稀疏,单源);FP16/BF16 Tensor Core 148 TFLOPS;TF32 Tensor Core 74 TFLOPS;FP32 44 TFLOPS;FP64 1.0 TFLOPS(合规裁剪:H100 SXM同表FP8 3958/FP16 1979/FP64 34);INT8 296 TOPS/592 TOPS(单源) |
| 功耗 | — | 400W TDP(Flopper/VipraTech规格表/腾讯新闻'热设计功耗400W'三源;VipraTech正文另有350W矛盾口径已弃) |
| 发布年 | — | 2024(Flopper核验Launch Year 2024;腾讯新闻2023-11-10首发报道'即将推出';具体发布日未核验双源,不录) |
| L2缓存 | — | 60MB(单源VipraTech) |
| CUDA核心 | — | 14,592 FP32 CUDA核心(单源VipraTech) |
| 出口合规背景 | — | 2022-10美国出口管制后对华合规特供卡:算力密度较H100大幅裁剪(FP8约为H100 SXM的7.5%),以显存96GB+NVLink900GB/s保推理场景(QQ:综合算力约H100的20%) |