2 ms·
Compute performance is fine. 20+ TFlops (or T-Integer ops) on 128MB L3 cache on 16GBs GDDR6. I don't think anyone can complain about the raw compute performance
by dragontamer 6y ago
Compute performance is fine. 20+ TFlops (or T-Integer ops) on 128MB L3 cache on 16GBs GDDR6. I don't think anyone can complain about the raw compute performance of these cards.
Infrastructure: yeah. Without ROCm support, its hard to recommend. You still get Windows DirectCompute, OpenCL, and Vulkan compute. But ROCm is much easier to use than those other frameworks.
---------
For compute: With 40 WGP x 4 SIMDs per WGP x 32 Threads per WGP x Occupany 4, you're probably aiming at 20,480 SIMD-threads if you're going to be using compute on this card.
Occupancy 20 is supported: 20-threads per VALU, kinda like hyperthreading although you split your registers between the threads. 20 is probably too much (too few VALUs per thread), but 1 is too few (you'll be waiting on VRAM latency). So 4 to 8 is an estimate for how much occupancy a typical program probably needs to hide memory latency.
Those 16GBs / 20kiloThreads == 780kB per thread. Sooooo... yeah. RAM goes quick... really quick... on compute applications. Compute programmers really want more VRAM.