Determine the cause of Raptor GPU cache behavior under coherence

Determine whether GPU requests on the Intel Raptor platform traverse the ring’s coherence agents and use the ring coherence protocol to access the shared last-level cache, thereby explaining why enabling GPU-to-CPU cache coherence has no discernible effect on GPU read performance.

Background

The paper evaluates CPU–GPU memory sharing on the Raptor platform, which uses a ring interconnect and a shared last-level cache. Raptor lacks GPU instructions for fine-grained memory and cache control and supports only driver-initiated cache invalidation at inter-kernel boundaries.

The authors observe that enabling cache coherence does not measurably affect GPU read performance on Raptor. They offer a hypothesis about the underlying coherence path, but do not establish it experimentally or through hardware documentation, leaving the precise mechanism unresolved.