My untested rule of thumb is that if something fits inside of GPU-register space, its probably faster on GPU. (GPUs have more architectural registers than CPUs,

My untested rule of thumb is that if something fits inside of GPU-register space, its probably faster on GPU. (GPUs have more architectural registers than CPUs, since CPUs are using all those "shadow registers" to implement out-of-order stuffs). Compilers will automatically turn your most-used variables into registers today, though you might need to verify by looking at the assembly code whether or not you're actually in register-space. But if something fits inside of CPU-cache space, its probably faster on CPU. (Intel is upto 2MB L2 cache on some systems, AMD is up to 192MB L3 cache on some systems). But if something is outside of CPU-cache space but fits inside of GPU-VRAM space, its probably faster on GPU. (Better to be bandwidth-limited at 500GB/s GPU-VRAM speeds than 50GB/s DDR5 speed). Ex: 16GBs GPU-VRAM is back into GPU-sized solutions to a problem. Then if something fits inside of CPU-RAM space, which is like 2TB in practice (!!!!), then CPUs are probably faster. Because there's no point taking a 2TB DDR5 RAM limited problem, passing some of that data to 16GB or 96GB of GPU-VRAM, and then waiting for DDR5 RAM anyway. ------------ Not that I've tested out this theory. But it'd make sense in my brain at least, lol.