3 ms·
I don't know where they make their money, but if the hope is to compete in compute, I'd suggest that they do not "play the game" at all in GPUs and instead put
by m0zg 7y ago
I don't know where they make their money, but if the hope is to compete in compute, I'd suggest that they do not "play the game" at all in GPUs and instead put together something 7nm and TPU-like. If I were them, I'd be working on that right now. 1TFLOPs/W would be a good goal, 100TFLOPs in 100W, using bfloat 16. Sidestep the GPU bullshit entirely, print money in a niche instead.
- bobajeff 7y agoI wouldn't consider TPU/NPU/NNPUs a niche. It's inevitable at this point that you'll need machine learning accelerators just for running inferencing on edge devices. Even if done only on the cloud I can see more could providers investing in their own TPU-like hardware for their clusters.