| 规格 | NVIDIA H20 | NVIDIA A30 |
|---|
| 架构 | NVIDIA Hopper(H100同架构合规裁剪版;VipraTech与腾讯新闻双源实抓) | NVIDIA Ampere |
| 显存 | 96GB HBM3(Flopper/VipraTech/腾讯新闻三源一致) | 24 GB HBM2 |
| 显存带宽 | 4.0TB/s(三源一致;显存位宽6144bit仅第三方数据库搜索摘要提及,未实抓双源,不录) | 933 GB/s |
| 算力 | FP8 Tensor Core 296 TFLOPS(密集)/592 TFLOPS(2:4稀疏,单源);FP16/BF16 Tensor Core 148 TFLOPS(密集,单源;稀疏口径未双源实抓);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) | — |
| 互联 | NVLink 900GB/s(VipraTech+腾讯新闻双源;第几代NVLink无实抓双源,不录代数)+PCIe Gen5 x16 128GB/s(VipraTech实抓,BurnCloud旁证Gen5) | PCIe Gen4: 64GB/s / Third-gen NVLINK: 200GB/s |
| 发布年 | 2024(Flopper实抓Launch Year 2024;腾讯新闻2023-11-10首发报道'即将推出';具体发布日未实抓双源,不录) | — |
| 形态 | 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) | Dual-slot, full-height, full-length (FHFL) |
| L2缓存 | 60MB(单源VipraTech) | — |
| CUDA核心 | 14,592 FP32 CUDA核心(单源VipraTech) | — |
| MIG切片 | 最多7实例(VipraTech+腾讯新闻'7路MIG'双源) | — |
| 编解码器 | 7 NVDEC + 7 NVJPEG(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%) | — |
| 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 |