Determine the significance of cache-coherence overhead for Orin streaming accesses

Determine whether cache-coherence overhead is negligible for streaming GPU accesses to uncached registered host memory on the NVIDIA Orin platform, as suggested by the comparable microbenchmark performance observed despite enabled coherence.

Background

On Orin, registered host memory is uncached for GPU L2 access, and the runtime does not expose a mechanism to disable GPU coherence. Consequently, the evaluated GPU read path uses Read-Alone on uncached memory.

The authors observe performance comparable to that of write-back memory in streaming microbenchmarks and conjecture that coherence overhead is negligible. However, they also report substantial penalties for workloads with temporal locality, so the generality of the conjecture remains unresolved.

References

However, in our streaming microbenchmarks, its performance remains comparable to that on WB memory; we conjecture that CC overhead is negligible here.

— Characterizing CPU-GPU Memory Sharing over Unified Memory Architecture  (2610.13125 - Yang et al., 9 Oct 2026) in Section 5.2.2, subsection “NVIDIA GPUs”