Compute Performance Problems with the NVIDIA GeForce RTX 50 Series Owing to the Removal of 32-bit Support

The underlying reason behind the surprisingly poor computational performance of NVIDIA’s new GeForce RTX 5090, RTX 5080, and RTX 5070 Ti GPUs has been found by PassMark Software. The reason for the incompatibility with current benchmarking tools and applications is that NVIDIA has quietly stopped supporting 32-bit OpenCL and CUDA in its “Blackwell” architecture. On RTX 5000 series GPUs, the problem appeared when the error code “CL_OUT_OF_RESOURCES (-5)” was reported by PassMark’s DirectCompute benchmark. Developers looked into it and found that although the benchmark’s main use case has been 64-bit for years, a few of compute sub-benchmarks continue to use 32-bit code that used to work properly on RTX 4000 and older GPUs. NVIDIA’s developer website still shows 32-bit code examples and documentation even after actual support was removed, indicating that they did not adequately document this architectural change.

The effects go beyond software benchmarking. As computational activities revert to CPU processing instead of leveraging the GPU’s parallel design, applications built on outdated CUDA infrastructure—including technologies like PhysX—will suffer from severe performance deterioration. Older apps can still run on the RTX 40 series and older hardware thanks to this fallback mechanism, but the RTX 5000 series only uses the CPU to execute these activities, which results in significantly worse performance. In order to properly test the compute capabilities of the latest GPUs, PassMark is now working to upgrade the impacted OpenCL code to 64-bit. They caution, nonetheless, that without source code changes, many current programs with 32-bit OpenCL components might never run correctly on RTX 5000 series cards. The benchmark developer also points out that this modification doesn’t adequately account for DirectX9’s subpar performance, implying that other architectural adjustments might have an impact on legacy rendering paths. Even with today’s software upgrade, legacy benchmarks may still be negatively impacted by PassMark. The earlier benchmark run below, which does not have the most recent PassMark V11.1 build 1004 patches, demonstrates how much the newest generations suffer in the absence of adequate software support.