Definition
Plain language
NVIDIA's ready-made library of fast math routines for its graphics chips.
As stated in the literature
NVIDIA's GPU-accelerated basic linear algebra subprograms library, providing optimized matrix and vector operations; agent-written CUDA kernels are often measured against, or deliberately composed without, cuBLAS and CUTLASS templates.
Why it matters: It provides heavily optimized building blocks so developers get top graphics-chip performance without hand-tuning low-level math, and serves as the bar new custom code is measured against.
For example, when a program needs to multiply two large matrices on an NVIDIA graphics chip, it can call cuBLAS instead of writing that fast math from scratch.
Heard on the show
“No cuBLAS, no CUTLASS template.”Episode 177 — Why Raw Profiler Data Made an AI Worse at Writing GPU Code