Low performance on RX7900XTX
Posted: Mon Mar 27, 2023 1:22 pm
As noted in other threads, AMDs RX 7900XTX gets quite low performance that's barely better than the previous generations 6900XT.
The reason for that is likely that AMD changed the ratio of FP32 and FP64 throughput per WGP/CU with their recent RDNA3 GPUs. FP64 throughput was halved to 1:32 (or 1:64 when considering dual-issue) of FP32, while FP32 throughput per clock cycle has potentially been doubled from 64 FP32/CU to 128 FP32/CU, which is basically the same FP64:FP32 Ratio as Ampere/Lovelace. RDNA3 should perfom significantly better with OpenMM in Single or Mixed mode.
The reason for that is likely that AMD changed the ratio of FP32 and FP64 throughput per WGP/CU with their recent RDNA3 GPUs. FP64 throughput was halved to 1:32 (or 1:64 when considering dual-issue) of FP32, while FP32 throughput per clock cycle has potentially been doubled from 64 FP32/CU to 128 FP32/CU, which is basically the same FP64:FP32 Ratio as Ampere/Lovelace. RDNA3 should perfom significantly better with OpenMM in Single or Mixed mode.