22 ms·
Both AMD and Intel (and Qualcomm to some degree) just don't seem to get how you beat NVIDIA. If they want to grab a piece of NVIDIA's pie, they do NOT need to
by dheera 3y ago
Both AMD and Intel (and Qualcomm to some degree) just don't seem to get how you beat NVIDIA.
If they want to grab a piece of NVIDIA's pie, they do NOT need to build something better than an H100 right away. There are a million consumers who are happy with a 4090 or 4080 or even 3080 and would love for something that's equally capable at half price, and moreover, actually available for purchase, from Amazon/NewEgg/wherever, and without a "call for pricing" button. AMD and Intel are much better at making their chips available for purchase than NVIDIA. But that's not enough.
What they DO need to do to take a piece of NVIDIA's pie is to build "intelcc", "amdcc", and "qualcommcc" that accept the EXACT SAME code that people feed to "nvcc" so that it compiles as-is, with not a single function prototype being different, no questions asked, and works on the target hardware. It needs to just be a drop-in replacement for CUDA.
When that is done, recompiling PyTorch and everything else to use other chips will be trivial.
- flamedoge 3y agothats what hip is though. a recompiler for cuda code. its not good enough for ptx assembly to amdgpu yet
- DeathArrow 3y ago>Both AMD and Intel (and Qualcomm to some degree) just don't seem to get how you beat NVIDIA. That's simple, just do what Nvidia did. Be better than the competition.
- treprinum 3y agoThat's not going to work because each GPU has a different internal architecture and aligning the way data is fed with how stream processors operate is different for each architecture (stuff like how memory buffers are organized/aligned/paged etc.). AMD is very incompatible to Nvidia at the lowest level so things/approaches that are fast on Nvidia can be 10x slower on AMD and vice versa.
- dheera 3y ago> things/approaches that are fast on Nvidia can be 10x slower on AMD and vice versa. That's fine but the job of software is to abstract that out. The code should at least compile, even if it is 10x less efficient and if multiple awkward sets of instructions need to be used instead of one. If Pytorch can be recompiled for AMD overnight with zero effort (only `ln -s amdcc nvcc`, `ln -s /usr/local/cuda /usr/local/amda`) they will gain some footing against Nvidia.