← All resources

Basic facts about GPUs

Auto-discovered

Blog Intermediate

Damek Davis

Visit resource ↗

Description

Basic facts about GPUs is a single technical essay by Damek Davis on his personal site. Last updated 2025-06-18, it consolidates GPU performance ideas drawn from deep-learning systems posts rather than serving as a product or library page.

It walks through compute and memory hierarchy on an NVIDIA A100 model, memory-bound versus compute-bound kernels, arithmetic intensity and the Roofline ridge point, and a third overhead-bound regime. Later sections cover operator fusion, tiling for matmul-like kernels, coalesced HBM loads, shared-memory banks and warps, bank conflicts, occupancy, thread divergence, and quantization notes.

Readers should be comfortable with basic parallel programming and matrix multiplication; CUDA or PyTorch experience helps but is not required to follow the diagrams. The tone targets intermediate practitioners learning why kernels stall.

The page is a free HTML article with a table of contents, ASCII diagrams, and conceptual code snippets. It cites a short list of prior posts and acknowledges reviewer comments.

It is useful as a compact mental model for diagnosing GPU bottlenecks before diving into vendor-specific CUDA optimization guides.