HardwareDSMEM
Distributed shared memory
In plain English
Access to the shared-memory storage of other thread blocks within a supported GPU cluster.
Technical definition
Distributed shared memory lets cooperating blocks in a thread block cluster address shared-memory regions belonging to other blocks in that cluster.
Engineering details
Remote access is subject to cluster synchronization and lifetime rules. Its instruction path and performance differ from local shared memory, even though both expose shared-memory addresses. Transfer size, access coalescing, and the selected instruction affect throughput and latency.
Why it matters
DSMEM can support cross-block cooperation without routing every intermediate through global memory. It should not be modeled as free access to a larger pool with local shared-memory bandwidth.
How to read it in InferenceX
The Blackwell port contrasts local LDS, cluster-scoped access, and bulk copy. A lower remote bandwidth is not proof of an HBM bottleneck; the tested path is within the cluster memory system.