I like this a lot. OpenCL microbenchmarks applied to many iGPU systems. What we see here is that AMD/Apple has a lot of L1 and L2 cache, while Intel does not.

I like this a lot. OpenCL microbenchmarks applied to many iGPU systems. What we see here is that AMD/Apple has a lot of L1 and L2 cache, while Intel does not. They further go into analysis, and figure out that all three systems have similar "cache-per-flops" (or at least, more-similar compared to raw results). ---------- I'm surprised to see that only Intel has an iGPU that has cache-to-cache transfers between CPU-and-GPU. I would have thought that both AMD and Apple would have had this implemented. I wonder if you need to do a "special memcopy" or maybe a shared-virtual-memory kind of setup to actually have the cache-to-cache transfers?