Dispatch to worker cores
Multiple CPU cores can work in parallel on assigned candidates.
GPU-NATIVE / MEASURED
3.7 million strategies.
Six minutes, 43 seconds.
Measured on an RTX 3090 with a real research setup, counting strategies completed across generation, full backtesting, scoring and selection.
RTX 3090 / REAL RESEARCH SETUP
Measured on 2026-09-30 with a real research setup: QHG# 37-minute data (2014-04-13 to 2026-09-25), 121,440 bars per candidate, random generation, 97% GPU utilisation with all 82 of 82 SMs (GPU compute units) active.
3,745,792Strategies evaluated in 6 min 43 s
0.108 msAverage time per strategy
11,992Live Chronograph speed: candidates / second (30-minute peak 12,590)
1.13 billionBars processed per second on average
Live speed and peak are the Chronograph’s readings at the time; averages are total strategies divided by total time. Actual speed depends on the graphics card, data length and research setup.
CPU / GPU · ARCHITECTURE
The same research question.A different compute scale.
Candidate batches, evaluated in parallel
CPU prepares the research task; GPU batches generation, backtesting and scoring.
Blocks on the left only illustrate CPU task distribution, not core counts or speedup. On the right is the AlphaPrime Chronograph dial, recorded on an RTX 3090: its 82 outer cells match the 82 SMs, and the animation only illustrates activity.