Tensor FP16/BF16 performance
PFLOP/sPeak tensor 16-bit throughput per accelerator.
↑ higher is betterEpoch AI
- #01NVIDIA Vera Rubin Superchip4.00
- #02AMD Instinct MI355X2.52
- #03NVIDIA GB300 (Blackwell Ultra)2.50
- #04NVIDIA GB2002.50
- #05Google TPU v7 Ironwood2.31
- #06AMD Instinct MI350X2.31
- #07NVIDIA B300 (Blackwell Ultra)2.25
- #08NVIDIA B2002.25
- #09Intel Habana Gaudi31.68
- #10AMD Instinct MI325X1.31
- #11AMD Instinct MI300X1.31
- #12Microsoft Maia 2001.27