6 ms·
I'm still mad about CUDA cores. I thought I was going to be able to write some epic 1000x level parallelism running individually on all cores. It turns out, a
by dinobones 4y ago
I'm still mad about CUDA cores. I thought I was going to be able to write some epic 1000x level parallelism running individually on all cores.
It turns out, a CUDA core is not actually a "core."
- Melatonic 4y agoWhy do they call it that then? I never really looked into it that much and just took as a measure of a certain type of compute capability ( FP16 or FP64 right? )
- wmf 4y agoThe SIMT architecture makes it look to the programmer like each FPU is a separate core, but all the cores in an SM have to run in lockstep to get good performance.
- monocasa 4y agoThey use that metric because it makes their marketing specs look nicer. Ultimately it's a count of the number of SIMD lanes.