3 ms·
What you mean by GPU cores here? In their marketing terminology, the jesters at NVidia are calling SIMD lanes "cores". For F32 operations, 50 AVX512 cores hav
by fulafel 6y ago
What you mean by GPU cores here?
In their marketing terminology, the jesters at NVidia are calling SIMD lanes "cores".
For F32 operations, 50 AVX512 cores have 60*(512/32)=800 SIMD lanes. And of course there are about 64 cores per x86 server CPU now at the high end.
- SatvikBeri 6y agoBy cores, I mean what NVidia calls "NVidia CUDA Cores", e.g. there are 5,120 on the Tesla V100.
- fulafel 6y agoOk. So translated to CPU terms there would seem to be 80 cores ("SPs") in a V100, that each can do up to 64-lane SIMD/SIMT with FP32 data. [1] (SIMT for NVidia seems to be just a programming model that compiles to SIMD instructions[2], a bit like what you get with ispc on CPUs) [1] https://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf https://images.nvidia.com/content/volta-architecture/pdf/vol... pages 17 and 10 [2] https://www.realworldtech.com/forum/?threadid=195094&curpostid=195099 https://www.realworldtech.com/forum/?threadid=195094&curpost...
- SatvikBeri 6y agoYeah, although a "Streaming Multiprocessor" is still less general than a CPU core IIUC.