CUDA 13.2 release adds unified memory tiering and deprecates two legacy APIs used by 8% of sampled workflows
Blackwell Ultra supply ramp clears the CoWoS packaging bottleneck
NVIDIA released CUDA 13.2, adding a unified memory tiering scheduler that allows developers to specify a device-resident working set with automatic host spill. The release also deprecates two legacy APIs introduced in CUDA 11, with migration guidance pointing to the unified memory path. The accompanying developer notes state that approximately 8 percent of sampled production workflows depend on at least one deprecated call, concentrated in inference serving frameworks. A compatibility shim is provided for two release cycles. Community discussion following the release focused on the tiering scheduler's interaction with multi-tenant inference schedulers.
Memory tiering directly targets inference serving economics. Deprecating APIs that 8% of sampled workflows use is a manageable migration cost.
Two release cycles of shim is a deadline, not a courtesy. Eight percent is small enough to ignore and large enough to break someone in production.
- Unified memory tiering scheduler with automatic host spill
- Approximately 8 percent of sampled production workflows depend on a deprecated call
- Compatibility shim provided for two release cycles
- Deprecation window capped at two releases