DAILY BRIEF / ATTENTION
Sparse attention moves the bottleneck to the indexer.
What changed
At 128K input, the indexer and Top-k take 47% of a DeepSeek V3.2 prefill layer; one key head per group avoids a 7.6x decode slowdown.
Why it matters
When profiling sparse-attention models, measure indexer and Top-k selection as their own costs, not as part of attention.
Scope and limitations
NVIDIA reports normalized runtimes with no GPU model or absolute times. The 2.1-2.4x Top-k gain over radix select is for TensorRT LLM on DeepSeek V3.2/V4 decode.
Go to the source
A sourced technical brief published 18 Nov 2026. The implication is turingfold’s interpretation. We have not independently benchmarked this change.