Parallelism
Thread block cluster
Also known as CTA cluster, cooperative grid array, CGA
In plain English
A group of GPU thread blocks scheduled together so they can coordinate through cluster-scoped mechanisms.
Technical definition
A thread block cluster groups CUDA blocks, also called CTAs, with co-scheduling guarantees and access to supported cluster-level synchronization and memory operations.
Engineering details
On the architectures discussed in the Blackwell port, blocks in a cluster are scheduled within the same GPC. Cluster dimensions affect placement, distributed shared-memory access, and multicast opportunities. A large cluster can leave resources unused if its size does not fit the available local grouping.
Why it matters
Increasing cluster size trades potential data reuse against scheduling flexibility. Divisibility by the GPU’s total SM count does not guarantee full utilization within every GPC.
How to read it in InferenceX
The Blackwell article discusses preferred and fallback cluster sizes. Those launch choices must be evaluated with the actual enabled SM grouping and the kernel’s per-block resource requirements.