3 ms·
Thanks. Its always good to see real benchmarks. One "qualm" is that AMD's APU is a very interesting platform, which can share pointers and memory between the C
by dragontamer 9y ago
Thanks. Its always good to see real benchmarks.
One "qualm" is that AMD's APU is a very interesting platform, which can share pointers and memory between the CPU and GPU. Some degree of caching is shared between the CPU and GPU since both are on the same die!
Its not a big deal for cryptonight (which basically pushes AES-NI instructions: so its fastest on the CPU anyway. And L3 is so small that only 2-instances can run on the Ryzen 5 2400G). But in the general case (say: maybe Ethash, which isn't as latency bound / L3 bound), there may be custom code that can be written to take advantage of AMD's HSA platform.
Code which doesn't exist yet (and therefore can't be benchmarked). In part, the ROCm project doesn't even have Raven Ridge support yet, so HSA coding / CPU/GPU sharing can't really take place yet. But looking forward: it might be an interesting project to see EtHash implemented as HSA / ROCm code.