Published onSeptember 26, 2022Load 60 BERTs onto a single T4technicalUsing partial runners to load 60 BERTs onto a T4 with no prespecified colocation.Read more →
Published onSeptember 16, 2022Getting Rid of CPU-GPU Copies in TensorFlowtechnicalPassing inputs and outputs directly through GPU memory in TensorFlow.Read more →
Published onJuly 28, 2022Increase usable cloud GPU memory by up to 6.6% through disabling ECCtechnicalMany cloud GPUs are configured by default in a way that reduces the total amount of GPU memory available.Read more →