Historically, the energy efficiency of leading GPUs and TPUs has doubled every 2 years. In tensor-FP16 format, the most efficient accelerators are Meta’s MTIA, at up to 2.1 x 1012 FLOP/s per watt, and the NVIDIA H100, at up to 1.4 x 1012 FLOP/s per watt. The upcoming Blackwell series of processors may be even more efficient, depending on their power consumption.
| Hardware | Manufacturer | Energy efficiency (tensor and FP16 if available) | Release date |
|---|---|---|---|
| NVIDIA GB200 NVL2 (per GPU) | NVIDIA | 2.08 x 10^12 | 2024-06-02 |
| NVIDIA B100 | NVIDIA | 2.50 x 10^12 | 2024-03-18 |
| NVIDIA B200 | NVIDIA | 2.25 x 10^12 | 2024-03-18 |
| Google TPU v5e | 1.41 x 10^12 | 2023-08-29 | |
| NVIDIA H100 SXM5 80GB | NVIDIA | 1.41 x 10^12 | 2022-09-20 |
| Google TPU v4 | 1.62 x 10^12 | 2021-05-20 | |
| NVIDIA A100 SXM4 80 GB | NVIDIA | 7.80 x 10^11 | 2020-11-16 |
| Google TPU v4i | 7.89 x 10^11 | 2020-01-01 | |
| Google TPU v3 | 5.59 x 10^11 | 2018-05-18 | |
| NVIDIA Tesla V100 SXM2 32 GB | NVIDIA | 4.17 x 10^11 | 2018-03-27 |
| Google TPU v2 | 1.61 x 10^11 | 2017-05-01 | |
| NVIDIA P100 | NVIDIA | 7.48 x 10^10 | 2016-04-05 |
Epoch's work is free to use, distribute, and reproduce provided the source and authors are credited under the Creative Commons BY license.
Download this data
CSV, Updated Nov. 24, 2025
Explore this data
Key data on over 170 AI accelerators, such as graphics processing units (GPUs) and tensor processing units (TPUs).