580.105.008
Setting CUDA_DISABLE_PERF_BOOST to 1 does not enable a CUDA program to surpass the p2 level.
When I exit the CUDA application, it instead becomes P0. Why is such a restriction imposed on the CUDA program? It prevents me from fully utilizing the graphics card’s capabilities.
Is there a feature like the one found in NVIDIA Profile Inspector on Windows, which allows setting CUDA - Force P2 State=off?
These limitations are incredibly frustrating. When the graphics card is in the P2 state, the latency of my deep learning inference becomes unstable, yet my use case demands consistent inference latency. Haven’t the official developers addressed this issue?
Even if, as suggested in some blog posts, operating in the P0 state may cause certain computational precision errors, high precision is not required for my application scenario. Could an adjustable parameter be provided to lift this restriction?
