HardwareMMA
Matrix multiply-accumulate
In plain English
An operation that multiplies matrix tiles and adds the product to an accumulator.
Technical definition
Matrix multiply-accumulate computes a matrix product and accumulates it into an output tile, with supported shapes and numerical formats defined by the instruction interface.
Engineering details
Tensor-core instructions implement MMA at particular execution scopes. Input format, accumulator precision, operand placement, tile dimensions, and independent in-flight operations all affect throughput. A small shape may leave arithmetic paths unused or spend more time moving operands than performing multiplication.
Why it matters
Peak tensor FLOPs require an appropriate instruction mix and enough independent work. A GEMM kernel must also load data, synchronize stages, and store outputs, so instruction throughput is an upper bound rather than a complete kernel measurement.
How to read it in InferenceX
The Blackwell port measures instruction shapes and queue depths separately. Its formula inconsistency is retained with an editorial note, so readers should not reuse the mismatched denominator as verified arithmetic.