howtospark
Training

N:M (2:4) sparsity

Pruning / sparsityReference

2-of-4 structured sparsity — the one sparsity pattern NVIDIA actually accelerates.

Relieves
CapacityBandwidthCompute
Targets
Weights · Structure
Format
2:4 semi-structured
Granularity
2-of-4 pattern
Structure
semi-structured
Lifecycle
PTQ · no gradients
Calibration
Small calibration set
Compression
~2× with real speedup
Quality
Moderate loss; hardware-accelerated
Hardware
NVIDIA Ampere+ sparse tensor cores
Runtimes
TensorRT-LLMvLLM

A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.