Compiling Julia to TPUs [pdf]

arxiv.org

7 pointsKenoFischer5 comments

This is a very impressive work. I am wondering why it was compared against PyTorch (just on the CPU) and not against Tensorflow (on CPU and TPU) or do I miss something?

Good question.

I wonder if the point of this article is that other tools can be used to access TPUs other than the nominal Tensorflow-based workflow. I.e. the article is to show that another workflow exists to access the capabilities of the TPU, which are (to some degree at least) already known. Keno Fischer says [1] they'll have other benchmarks out soon.

[1] https://twitter.com/KenoFischer/status/1054932277824774145

OK, I see. In any case, even if the performance would not match Tensforflow on TPU this is a major achievement. After all, Google made the TPUs specifically for Tensorflow. I would not have imagined that it was so "easy" to replace LLVM by XLA.

Still using LLVM as backend (ultimately)?

The XLA:TPU compiler is not open source, so it's anyone's guess (unless you work at Google), but my understanding is that it does NOT use LLVM under the hood.

Reply on news.ycombinator.com