After Nvidia pulled the plug on 32-bit CUDA support with CUDA 12.0 (and later), technologies built around this platform such as PhysX subsequently reached End Of Life. A new report from PassMark at X suggests that dropping legacy CUDA support extends its impact to older OpenCL code, written for 32-bit systems. This explains the lackluster performance in PassMark's Direct Compute benchmark, part of which contains 32-bit chunks of code; likely older libraries or modules.
Nvidia hasn't flipped the switch all of a sudden with one update. Legacy CUDA support has been gradually phased out through successive updates over the past few years. Notably, programs written with 32-bit CUDA in mind can still be run on RTX 40 or older hardware. On newer RTX 50 GPUs, older CUDA applications now fallback to the CPU, which is abysmally slower than the graphics card for parallel processing, as seen with PhysX. We suspect this is mostly a software limitation, and may be overcome by 32-bit to 64-bit translation layers, but don't expect anything official from Nvidia.
Found the explanation for RTX 5090 and 5080 low compute performance.Link: https://t.co/8Rq3K3Zw1PWe found out a few hours ago that nvidia removed OpenCL 32bit support. Seems it depended on CUDA 32bit. Which is also gone. We've been unable to buy a 5090 for testing (no stock… pic.twitter.com/gIp1HiOsJNFebruary 25, 2025
PassMark reports that Nvidia's decision to terminate 32-bit CUDA compatibility also affects legacy OpenCL code. While we don't have an official statement, the firm's tweet indicates that Nvidia has outright dropped 32-bit OpenCL support. Given the widespread Blackwell shortages, the PassMark team has been unable to procure an RTX 50 GPU for testing.
In fact, Nvidia's developer documentation for OpenCL still offers samples of code written for 32-bit environments. The code reportedly works fine on older RTX 40 (Ada Lovelace) GPUs, but throws a "non-obvious" error with RTX 50 hardware. It's fair to assume that dated OpenCL programs will never work on Blackwell and future GPUs, without necessary changes to the source code.
While the main PerformanceTest application from PassMark is compatible with modern systems, it still contains several sub-benchmarks that didn't require a 64-bit address space. For the benchmark to gauge performance properly, PassMark probably needed to recompile some of their kernels in 64-bit mode.
That's easier said than done, as older libraries and dependencies made with 32-bit systems in mind can quickly pile up depending on the age of your software. Likewise, this transition may require significant code changes when dealing with low-level operations and memory management.
The latest release of PerformanceTest has resolved these issues by updating required components to 64-bit, with the whole situation taking around a week or so to clear up. It's unlikely that entire codebases will shut down and compute software will cease to work, since most developers have slowly migrated towards 64-bit code over the years.
