Validate first, then compare timings

Synthetic English narration.

Fix the comparison
Use the same saved economies, parameters and random shocks for each configuration. Fix the hardware and the required accuracy before comparing one GPU with four.

Verify more than the final moments
Check arrays and residuals, and also which firms operate and when they exit. A faster implementation is useful only if it passes the same checks.

Warm up, then start the clock
Warm up compilation and solver paths before measuring. Start the clock before the whole solve, and stop it once the results are returned.

Keep all three repetitions
Run three clean repetitions for each configuration and keep every timing. These bars show saved timings of the whole workload, measured without the profiler.

Report the comparison clearly
Report all three trials and their median, with the inputs, the hardware and what the clock included, so others can repeat the comparison.