| 规格 | NVIDIA A800 | NVIDIA T4 |
|---|
| 架构 | NVIDIA Ampere | Turing (TU104-895) |
| 显存 | 80 GB HBM2e | 16GB GDDR6 with ECC (默认开启) |
| 显存带宽 | 1,935 GB/s | 320GB/s (早期一页版datasheet口径300GB/s) |
| FP64算力 | 9.7 TFLOPS | — |
| FP64 Tensor Core | 19.5 TFLOPS | — |
| FP32算力 | 19.5 TFLOPS | — |
| TF32 Tensor Core | 156 TFLOPS | — |
| FP16算力 | 312 TFLOPS | — |
| INT8算力 | 624 TOPS | — |
| Tensor核心 | 512 third-generation Tensor Cores per GPU | — |
| CUDA核心 | 8192 FP32 CUDA Cores per GPU | 2,560 (40 SM) / Tensor Core 320 / RT Core 40 |
| 互联 | NVLink: 400 GB/s | PCIe Gen3 x16 (支持x8运行, Interconnect BW 32GB/s) |
| 系统接口 | PCIe Gen 4, x16 lanes | — |
| 形态 | PCIe full height/length, double width | 单槽半高半长 Low-Profile 6.6英寸(长169.53mm), 被动散热, 板重301g |
| MIG切片 | Up to 7 GPU instances, 10GB each | — |
| 功耗TDP | 300 W | — |
| ECC | Yes | — |
| 发布年 | 2022 | 2018 (Product Brief v01 2018-10-24初版, Dec18版datasheet © 2018 NVIDIA) |
| 显存位宽 | — | 256bit |
| 算力 | — | FP32 8.1 TFLOPS / FP16·FP32混合精度 65 TFLOPS / INT8 130 TOPS / INT4 260 TOPS |
| 功耗 | — | 70W (Total board power, 默认) |
| 时钟 | — | Base 585MHz / Boost 1590MHz (Base为70W功耗墙下调校) |
| 媒体引擎 | — | 1080p30解码 H.264 32路 / HEVC 44路 / VP9 32路 |
| SR-IOV | — | Supported; 16 VF（支持） |
| 工艺 | — | 12nm FFN, 136亿晶体管, 545mm2 |
| Compute APIs | — | CUDA, NVIDIA TensorRT, ONNX |