| 规格 | NVIDIA H800 | NVIDIA GeForce RTX 4090 |
|---|
| 架构 | NVIDIA Hopper | Ada Lovelace (AD102-300, 128 SM) |
| 显存 | 80 GB HBM2e | 24GB GDDR6X |
| 显存(94GB版) | 94 GB HBM3 | — |
| 显存带宽 | 2 TB/sec | 1008GB/s (21Gbps x 384bit) |
| 显存带宽(94GB版) | 3.9 TB/sec | — |
| FP32算力 | 51 TFLOPS | — |
| Tensor核心 | 456 fourth-generation Tensor Cores per GPU | — |
| CUDA核心 | 14592 FP32 CUDA Cores per GPU | 16384 (128 SM) / Tensor Core 512 (四代) / RT Core 128 (三代) |
| 互联 | NVLink: 400 GB/s | PCIe Gen4 x16 (不支持NVLink) |
| 系统接口 | PCIe Gen 5, x16 lanes | — |
| 形态 | PCIe full height/length, dual slot | FE三槽 304x137mm, 供电 3x8pin转接或12VHPWR(450W Gen5) |
| MIG切片 | Up to 7 GPU instances | — |
| 功耗TDP | 350 W | — |
| ECC | Yes | — |
| 发布年 | 2023 | 2022 (2022-09-20 GTC发布, 2022-10-12开售, MSRP $1599) |
| 显存位宽 | — | 384bit |
| 算力 | — | FP32 82.6 TFLOPS (官方发布口径, 现页四舍五入标83) / Tensor 1321 AI TOPS (FP8稀疏) / RT 191 TFLOPS |
| 功耗 | — | 450W TGP (平均游戏315W, 建议系统电源850W) |
| 频率 | — | Base 2.23GHz / Boost 2.52GHz |
| 工艺 | — | TSMC 4N, 763亿晶体管, 608.5mm² (AD102 满血144SM, 4090启用128) |
| 媒体引擎 | — | 2x 第八代NVENC / 1x 第五代NVDEC, AV1编解码 |
| 显示输出 | — | HDMI 2.1a x1 + DisplayPort 1.4a x3, 4K240/8K60(DSC) |
| CUDA算力等级 | — | Compute Capability 8.9 |