Validation
It computes what the textbooks compute
Fast is only useful if it is right. We run open problems with published or exact answers and put our numbers next to them. Most open as a project in the trial.
- Queues and factory physicsWaiting times and throughput match the analytical answers: Erlang C, Factory Physics' Penny Fab, CONWIP curves.
- Line balancingReaches all 13 published optimal station counts in the Scholl SALBP-1 sets.
- Scheduling and fabsReplays the known optimal FT06 schedule; Intel Mini-Fab makes 83.6 of 84 lots a week; SMT2020 at full scale lands within 3% of the published cycle times.
| Problem | Published / exact | Machsizer |
|---|---|---|
| Three parallel servers, Erlang CQueueing theory, M/M/3 | lead time 2.079 h | 2.066 h |
| Penny Fab Two, open networkHopp & Spearman, Factory Physics, ch. 7 | 30.52 h · WIP 9.16 | 30.50 ± 1.33 h · WIP 9.2 |
| Penny Fab One under CONWIPFactory Physics, ch. 7 · WIP 1 / 4 / 12 | 0.1250 / 0.2857 / 0.4000 per h | 0.1250 / 0.2842 / 0.3999 per h |
| Line balancing, Jackson and GuntherScholl SALBP-1 data sets · 13 cycle times | optimal station counts | 13 of 13 optimal |
| Intel Mini-Fab, re-entrant flow with batchesKempf 1994 · release 84 lots / week | 84 lots / week | 83.6 lots / week |
| Job shop FT06Fisher & Thompson 1963, OR-Library | optimal makespan 55 | 55 on the optimal schedule |
| SMT2020 semiconductor fabKopp et al. 2020, dataset 1 (HVLM) · 1,443 tools · up to 583 steps · year two of two, five seeds · cycle time of all lots, published by AutoSched · not in the trial | 1,227 h / 681 h | 1,258 ± 21 h / 683 ± 12 h |
Exact answers are computed independently, never from our own output.