4 ms·
The Cebras prototype was about 0.86 PFLOPS (?) for a whole wafer (1T transistors) so this Intel chip looks like a potential viable competitor at 1PFLOP for only
by choppaface 6y ago
The Cebras prototype was about 0.86 PFLOPS (?) for a whole wafer (1T transistors) so this Intel chip looks like a potential viable competitor at 1PFLOP for only 100B transistors (even if just FP16). I'm sure Intel will want to chase NVidia but Cerebras is also a threat given it already has software support (Tensorflow, Pytorch, etc). Maybe I'm making an unfair comparison but looks Ponte Vecchio would put Intel just above where Cerebras was a couple years ago.
https://www.nextbigfuture.com/2020/11/cerebras-trillion-transistor-ai-wafer-chip-crushes-gpu-supercomputer-by-200-times.html https://www.nextbigfuture.com/2020/11/cerebras-trillion-tran...
- ipsum2 6y agoThese PFLOP measurements aren't very comparable. Benchmarks are probably the best way to see which chip is better. For example, Cerebras claims ~1PFLOPS, but has not put out any numbers showing that that actually is the case, on standard machine learning models such as ResNet, Transformers, or others. Their hardware is basically DoA.