What about the Gaudi 2 GPU?
I previously compared the NVIDIA A100 (2021) and H100 (2022) GPUs, and you may recall they had similar 80 GB memory sizes, and floating point throughput of 300 TFLOPS and 2000 TFLOPS, respectively.
I just read about the new Habana Gaudi2 GPU (now owned by Intel) which has more memory (96 GB) is purported to be even faster. It seems that it is a little hard to compare it to the NVIDIA processors, in part because of a different architecture more focused on data transfer than on arithmetic speed but also because Intel is choosing to not divulge the latter! Instead they focus on comparative benchmarks, which of course were chosen by Intel. If we take the benchmarks seriously (for an independent look, see here), then the Gaudi2 looks like it has throughput perhaps double that of the A100, which puts it somewhere between the A100 and H100 if your AI application is limited by data transfer. But - if you are running a fixed LMM (essentially fixed NN for sexbot AI in Second Life) maybe not? Well of course that kind of benchmark was not included in Intel's docs...

Comments
Post a Comment