Perceptron branch prediction in CPUs

entry · p2026-09-06-179 · quote not found

improves misprediction rates for the SPEC 2000 benchmarks by 10.1% over the gshare predictor

tech: learned-branch-prediction
passes the definition as: a single-layer neural net per branch, fit at runtime to the program's own branch-outcome history, replacing hand-tuned saturating counters
from: 2017
region: world
domain: access
side: benefit
people: 7,470,000,000 devices (2025)
effect: 10.1 percent (2001) — fewer branch mispredictions than the gshare predictor at the same 4KB hardware budget on SPEC 2000 — the speculation that stops a pipelined CPU stalling on every branch
reversible: yes
who holds the switch: nobody
attribution: named
source kind: primary
the statistic: Jimenez & Lin (HPCA 2001) measured a 10.1% improvement in misprediction rate over gshare at a 4K byte budget on SPEC 2000. AMD shipped the technique as 'Neural Net Prediction' in Zen/Ryzen — 'an artificial intelligence neural network that learns to predict what future pathway an application will take based on past runs' (https://ir.amd.com/news-events/press-releases/detail/741/amd-takes-computing-to-a-new-horizon-with-ryzentm-processors, 2016-12-13). Reach here is smartphone application processors alone — 7,470 million smartphone subscriptions in 2025 (Ericsson Mobility Report key figures, https://www.ericsson.com/en/reports-and-papers/mobility-report/key-figures) — a floor that excludes PCs and servers.
source: https://www.cs.utexas.edu/~lin/papers/hpca01.pdf
proposed by: researcher (opus), brief 2026-09-06, quiet-adoptions
received: 2026-09-06T09:39:39-04:00
harvest report: impacts/harvest/2026-09-06-quiet-adoptions.md

Rulings, in order

2026-09-06T09:39:47-04:00 · validatorfail — quote not found on the fetched page (HTTP 200 application/pdf, 75863 chars)
2026-09-06T09:59:36-04:00 · sessionseverity — CPUs faster for everyone; nobody's switch

The proposal as it arrived (JSON)

{
 "id": "p2026-09-06-179",
 "received": "2026-09-06T09:39:39-04:00",
 "proposer": "researcher (opus), brief 2026-09-06, quiet-adoptions",
 "kind": "entry",
 "sha": "7b39c6c4139304cb",
 "data": {
  "kind": "entry",
  "tech": "learned-branch-prediction",
  "tech_name": "Perceptron branch prediction in CPUs",
  "definition": "a single-layer neural net per branch, fit at runtime to the program's own branch-outcome history, replacing hand-tuned saturating counters",
  "from": "2017",
  "region": "world",
  "domain": "access",
  "side": "benefit",
  "people": {
   "n": 7470000000,
   "unit": "devices",
   "date": "2025"
  },
  "effect": {
   "n": 10.1,
   "unit": "percent",
   "date": "2001",
   "what": "fewer branch mispredictions than the gshare predictor at the same 4KB hardware budget on SPEC 2000 — the speculation that stops a pipelined CPU stalling on every branch"
  },
  "reversible": true,
  "switch": "nobody",
  "statistic": "Jimenez & Lin (HPCA 2001) measured a 10.1% improvement in misprediction rate over gshare at a 4K byte budget on SPEC 2000. AMD shipped the technique as 'Neural Net Prediction' in Zen/Ryzen — 'an artificial intelligence neural network that learns to predict what future pathway an application will take based on past runs' (https://ir.amd.com/news-events/press-releases/detail/741/amd-takes-computing-to-a-new-horizon-with-ryzentm-processors, 2016-12-13). Reach here is smartphone application processors alone — 7,470 million smartphone subscriptions in 2025 (Ericsson Mobility Report key figures, https://www.ericsson.com/en/reports-and-papers/mobility-report/key-figures) — a floor that excludes PCs and servers.",
  "source": "https://www.cs.utexas.edu/~lin/papers/hpca01.pdf",
  "quote": "improves misprediction rates for the SPEC 2000 benchmarks by 10.1% over the gshare predictor",
  "attribution": "named",
  "source_kind": "primary"
 },
 "from_file": "impacts/harvest/2026-09-06-quiet-adoptions.md"
}