[pixl97]

With the number of operations and the error rate in GPUs this is going to be interesting in SOTA models.

[1024core]

Don't forget quantization..