Hacker News Reader

One way Intel can compete with CUDA is if they allow us to write plain Numpy code, and magically compile it for their GPUs with similar performance to what we get from Nvidia when writing in CUDA.

An alternative would be for them to (again magically) modify all major DL frameworks to support their GPUs.

I don't even know which option is more realistic.

View on HN

I think #1 is the most realistic. Intel does some very significant work with Numpy as it is with MKL. #2 seems like dark voodoo magic that'd be neat, but pretty much impossible.

One way Intel can compete with CUDA is to show a benchmark where they outperform NVidia in training speed. Not speed per watt, not inference, not oddly priced speed per dollar comparisons.

If the speed is there the world will write the code.

Sure, they can "show a benchmark", however, to be more than 30% faster in reality (not just on a marketing slide) than the next generation Nvidia chip is probably just as unlikely as the first two options.

I'd love to get surprised though.

Reply on news.ycombinator.com