Rohan Anil, CoreAutoAI co-founder, defends GPU programmability using NVSHMEM as Jerry Tworek argues they are specialized GEMM engines

GEMM Processing Units

@_arohan_ 💚

This is not because of anything other than pure joy of having finished prototyping an old idea that had been simmering.

@Radipdegen No they won’t

@_arohan_ shh dont let nvidia hear u say that or theyll charge u per call next gen

@_arohan_ Sure you do to want to work with ASIC #12543 instead?

@ayzddzya I have been waiting more than half a decade to try this idea. Betryal not mine.

@_arohan_ Betryal 🙄

@NVIDIAAI Programming with it has been the most fun I’ve had in a long time!

@_arohan_ Same

@_arohan_ bro everyone is not sleeping these days?

@_arohan_ the cuda learning curve is brutal but worth it

@_arohan_ Haha I was joking 🙃 I think TPU and GPU excel in their own way.

@_arohan_ dont count your chickens before they hatch