Hardware
Two-SM MMA
Also known as 2SM MMA, cta_group::2
In plain English
A Blackwell matrix instruction mode in which a coordinated pair of blocks uses two SMs together.
Technical definition
Two-SM MMA is a Blackwell tensor-core execution mode in which a cooperating CTA pair performs matrix multiply-accumulate work across two SMs.
Engineering details
The larger execution scope supports different instruction shapes and data-sharing arrangements than a single-SM operation. Evaluating scaling requires accounting for the extra compute resources and the amount of work performed. Operand layout and shared-memory traffic can change the bottleneck alongside arithmetic capacity.
Why it matters
A twofold throughput gain with two SMs is not a twofold improvement in per-SM efficiency. Gains beyond that ratio can reflect relief of a single-SM memory bottleneck for particular shapes.
How to read it in InferenceX
The Blackwell port distinguishes SS and TS layouts and reports shape-dependent scaling. Preserve those conditions rather than treating its best ratio as a universal kernel or inference speedup.