3 ms·
It's not that impressive.. The NVIDIA A100 has better efficiency: 312 TFlops @ 400W for the same precision (FP16.32).
by ml_hardware 6y ago
It's not that impressive.. The NVIDIA A100 has better efficiency: 312 TFlops @ 400W for the same precision (FP16.32).
- fluffy87 6y agoAlso the A100 is cheaper, and it’s perf is verified by MLPerf. GraphCore’s claims are just that, claims. Some of these have been debunked already by people with access to the hardware (eg promised 10x speed ups were more like 1x).
- ml_hardware 6y agoDo you have a link to those results by independent researchers? I have also been very skeptical of GraphCore’s claims (and have talked about it on reddit) but haven’t had access or heard of anyone’s experience.