Why architecture matters here

Tensor core failures come from picking the wrong dtype path (falling back to CUDA cores), wrong tile size, or missing sparsity opportunities. Architecture matters because kernel choice + dtype align the workload to hardware.

Advertisement

The architecture: every piece explained

The top strip is the primitive. Kernel launcher sets up GEMM. WMMA / MMA is the warp-level matrix multiply-accumulate. Tile size depends on architecture. Accumulator dtype typically fp32 for stability.

The middle row is the details. Input dtype bf16/fp16/fp8. Sparsity 2:4 structured for 2x on Ampere+. Multi-buffering hides memory latency. Kernel selection uses cuBLAS/cuDNN/CUTLASS.

The lower rows are ops. Nsight profiling for MFU per kernel. Failure modes — dtype fallback. Ops covers hardware matrix + library version.

Tensor cores — WMMA + mixed precision + dtype paths + shapesthe matmul engine of modern GPUsKernel launcherGEMM setupWMMA / MMAwarp matrix opsTile size16x16 / 32x8 / 64x8Accumulator dtypefp32 typicalInput dtypebf16 / fp16 / fp8Sparsity 2:4structuredMulti-bufferinghide latencyKernel selectionvendor librariesNsight profilingMFU per kernelFailure modeswrong dtype pathOps — hardware matrix + library version + tuningcastprunepipelinepickprofileavoidavoidoperateoperate
Tensor core kernel selection with dtype and sparsity.
Advertisement

End-to-end flow

End-to-end: training loop launches GEMM. Kernel picks CUTLASS with bf16 input + fp32 accum + tile 128x128. WMMA executes; multi-buffering hides HBM read latency. Nsight shows 60% MFU. Sparsity 2:4 applied to weights; 40% speedup on inference. Kernel matrix versioned per driver.