Comment by fulafel

"Thread" term in GPUs and CPUs mean different things, it's more like a SIMD lane in GPUs. A bit like ISPC can compile your code so there's eg 32 invocations of your function per CPU thread running on the same time (if you're using 16-bit datums on AVX512), and you could have 2048 executions going on after multiplying up 32 cores * 2 SMT threads/core * 32 compiler executions.