Date: 2026-09-18. Score: 63,350, four weighted-read units below the previous record. All 256 matrix outputs are verified as exact integer polynomials.
The winning branch starts from an existing 63,358 schedule produced by
blocked instruction reordering. Its multiplication order differs from the
63,354 record. A joint optimization then changes the addition trees and
assigns storage across the whole program, reaching 63,351. Removing its one
copy 1,1 produces the submitted 63,350 program; no other instruction changes
in that final cleanup.
The joint model preserves that branch’s input-copy and multiplication graph and operation times. It chooses which earlier values feed each addition or partial-sum copy within an output, while enforcing every storage tier’s live capacity. This changes when intermediate sums die and which cells can hold later values. Every product is still evaluated exactly once.
| Read source | Previous record | New program | Change |
|---|---|---|---|
| Copies | 20,696 | 20,666 | −30 |
| Multiplications | 18,348 | 18,344 | −4 |
| Additions | 20,081 | 20,082 | +1 |
| Output reads | 4,229 | 4,258 | +29 |
| Total | 63,354 | 63,350 | −4 |
The program uses 4,096 multiplications, 3,840 additions, and 2,167 copies. Its highest address is 620.
The result comes from a joint tree/storage search using scheduling, reduction trees, allocation, and redundant-copy removal.
The saved search had reached a numerical LP value of 63,351 but stopped during crossover. Completing crossover returned a valid fractional solution. The first integer repair restricted choices to its nonzero support and fixed its unit entries; that restricted model was infeasible. The successful repair instead allowed the union of the LP support and the known 63,358 incumbent, fixing unit entries only where both solutions agreed. The incumbent remains feasible in this model, and the repair found an integral 63,351 program.
The solver-reported LP optimum of 63,351 applies to the joint matching model before copy cleanup. Its allocation certificate was not independently replayed for this submission. The final 63,350 program has one fewer copy and lies outside that model. These results do not prove unrestricted matrix-multiplication optimality.
The submitted IR is a frozen artifact. Its loader checks its SHA-256; the verifier checks the official scorer, an independent polynomial implementation, operation counts, read costs, maximum address, and verifier dependencies. Verification needs no optimizer or third-party packages:
python3 -S matmul/submissions/best_63350.py
Recreating the optimization requires the original research trace and external LP/MILP tooling. Those search dependencies are not required to verify this submission, and the loader does not claim to reconstruct the optimizer run.
IR: best_63350.ir. Loader and verifier: best_63350.py.
SHA-256: 3cf3b8dea456fa0a7dba2e5fab3a31af0dc971fd11d7b04332bcfd967d3259af.