4 ms·
Pretty cool. Shameless plug: My team at Anthropic is hiring people that can write accelerator kernels. Please reach out to <my username>@gmail.com if you want
by jbochi 2y ago
Pretty cool. Shameless plug: My team at Anthropic is hiring people that can write accelerator kernels. Please reach out to <my username>@gmail.com if you want to make state of the art models faster :)
- winwang 2y agoCurious, do you guys use Triton? I was surprised to find out PyTorch slowdown is palpable even compared to basic CUDA kernels.
- petermcneeley 2y agoqq: Would this be in CUDA?
- jbochi 2y agoNot sure what I can share publicly, but we use TPUs as well: https://cloud.google.com/blog/products/compute/announcing-cloud-tpu-v5e-in-ga https://cloud.google.com/blog/products/compute/announcing-cl...
- fpgamlirfanboy 2y agoIt's funny - former Google devs (whom I just maintain a good relationship with their former employer) are ideally positioned to profitably take advantage of the arb of TPU over GPU.
- bee_rider 2y agoGoogle should improve their abstractions until there isn’t room for that anymore, haha.
- trashtensor 2y agonot looking for a job at this time but i do this kind of work - what is the name of the team that you are working on / typically works on this?