算力
看 AI 算力供给、部署、电力、成本和效率怎样变化。
数据截至 2026/10/4
ANSWER
当前显示 64 条互不聚合的同指标序列。
- 覆盖时间
- 2017-05-01 – 2026-03-11 64 条独立序列
TREND
硬件 FP16 峰值能效
硬件 FP16 峰值能效,2017/5/1至2026/3/11;估算值为空心点,区间以竖线表示。
观测值估算值不同线段不跨断点连接
DATA
同口径数据
左右滑动查看完整数据。
| 日期 | 对象 | 数值 | 区间 | 性质 | 方法版本 |
|---|---|---|---|---|---|
| Google TPU v2 | 0.16 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 PCIe 16 GB | 0.45 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 SXM2 16 GB | 0.42 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA V100 | 0.42 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 SXM2 | 0.42 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 DGXS 16 GB | 0.5 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 DGXS 32 GB | 0.5 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 PCIe 32 GB | 0.45 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 SXM2 32 GB | 0.42 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100 SXM3 32 GB | 0.42 TFLOP/s/W | — | 派生值 | 1 | |
| Google TPU v3 | 0.27 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA GeForce RTX 2080 Ti 11GB | 0.46 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA Tesla V100S PCIe 32 GB | 0.52 TFLOP/s/W | — | 派生值 | 1 | |
| Google TPU v4i | 0.79 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A100 | 0.78 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A100 SXM4 40 GB | 0.78 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A100 PCIe | 1.04 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A40 PCIe | 0.5 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A100 SXM4 80 GB | 0.78 TFLOP/s/W | — | 派生值 | 1 | |
| Google TPU v4 | 0.81 TFLOP/s/W | — | 派生值 | 1 | |
| Tesla D1 Dojo | 0.91 TFLOP/s/W | — | 派生值 | 1 | |
| AMD Radeon Instinct MI250X | 0.77 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA GeForce RTX 3090 Ti | 0.36 TFLOP/s/W | — | 派生值 | 1 | |
| Intel Habana Gaudi2 | 0.75 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A800 SXM | 0.78 TFLOP/s/W | — | 派生值 | 1 | |
| Biren BR100 | 1.86 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA GeForce RTX 4090 | 0.73 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA H100 PCIe | 2.16 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA H100 SXM5 80GB | 1.41 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA L40 | 0.6 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A800 PCIe 80 GB | 1.25 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA A800 PCIe 40 GB | 1.25 TFLOP/s/W | — | 派生值 | 1 | |
| AMD Instinct MI300A | 1.29 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA GH100 | 1.41 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA H800 SXM5 | 1.41 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA L4 | 1.68 TFLOP/s/W | — | 派生值 | 1 | |
| Meta MTIA v1 (MTIA 100) | 2.05 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA GH200 | 1.41 TFLOP/s/W | — | 派生值 | 1 | |
| Google TPU v5e | 0.88 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA L20 PCle | 0.43 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA HGX H20 | 0.37 TFLOP/s/W | — | 派生值 | 1 | |
| AMD Instinct MI300X | 1.74 TFLOP/s/W | — | 派生值 | 1 | |
| AMD Radeon Instinct MI308X | 0.87 TFLOP/s/W | — | 派生值 | 1 | |
| Google TPU v5p | 0.85 TFLOP/s/W | — | 派生值 | 1 | |
| MTT S4000 | 0.22 TFLOP/s/W | — | 派生值 | 1 | |
| Meta MTIA v2 (MTIA 200) | 1.97 TFLOP/s/W | — | 派生值 | 1 | |
| Google TPU v6e Trillium | 2.42 TFLOP/s/W | — | 派生值 | 1 | |
| Maia 100 (M100) | 1.6 TFLOP/s/W | — | 派生值 | 1 | |
| Intel Habana Gaudi3 | 1.86 TFLOP/s/W | — | 派生值 | 1 | |
| AMD Instinct MI325X | 1.31 TFLOP/s/W | — | 派生值 | 1 | |
| Huawei Ascend 910C | 1.14 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA B200 | 2.25 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA H200 SXM | 1.41 TFLOP/s/W | — | 派生值 | 1 | |
| Amazon Trainium2 | 1.33 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA GB200 | 2.08 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA RTX Pro 6000 Blackwell | 0.84 TFLOP/s/W | — | 派生值 | 1 | |
| AMD Instinct MI350X | 2.31 TFLOP/s/W | — | 派生值 | 1 | |
| AMD Instinct MI355X | 1.8 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA B300 (Blackwell Ultra) | 2.05 TFLOP/s/W | — | 派生值 | 1 | |
| NVIDIA GB300 (Blackwell Ultra) | 1.79 TFLOP/s/W | — | 派生值 | 1 | |
| Google TPU v7 Ironwood | 2.4 TFLOP/s/W | — | 派生值 | 1 | |
| Amazon Trainium3 | 0.96 TFLOP/s/W | — | 派生值 | 1 | |
| Microsoft Maia 200 | 1.69 TFLOP/s/W | — | 派生值 | 1 | |
| Meta MTIA 300 | 0.75 TFLOP/s/W | — | 派生值 | 1 |
METHOD
口径说明
- 更新频率
- 硬件发布
- 覆盖时间
- 2017-05-01 – 2026-03-11
- 数值性质
- 派生值
- 版本
- 1
公开硬件规格的理论峰值,不代表应用实测效率