HardwareGPC
Graphics processing cluster
In plain English
A GPU hardware grouping that constrains where cooperating thread block clusters can be scheduled.
Technical definition
A GPC is an NVIDIA GPU hardware grouping containing execution resources, including groups of SMs, with architecture-specific organization.
Engineering details
The Blackwell article uses cluster co-scheduling to infer logical SM groupings. Enabled resources can vary among groups because of floorsweeping. Software-visible membership does not directly identify physical coordinates or establish which die contains each group in a multi-die package.
Why it matters
A kernel that assumes all groups have identical usable capacity can leave SMs idle. Logical topology matters for scheduling even when the physical floorplan remains unknown.
How to read it in InferenceX
Keep logical cluster-mapping evidence separate from the article’s pointer-chase-based placement hypothesis. Neither the device-wide SM count nor a single timing matrix alone describes all scheduling and memory-locality constraints.