External-kernel launch geometry autotune
A different threads-per-block and nonces-per-thread geometry can improve the frozen external kernel by at least 1%.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
A different threads-per-block and nonces-per-thread geometry can improve the frozen external kernel by at least 1%.
Why the test is meaningful
Geometry changes can alter occupancy, scheduling and amortization without changing SHA semantics; selection must still clear a preregistered effect gate on every dataset.
G(T,N)=geomean throughput(T,N)/throughput(512,32)promote iff G≥1.01 and datasets≥4/4hash(candidate)=hash(baseline)How it was tested
Test TPB 128/256/512 crossed with 16/32/64 nonces per thread on four fresh headers, two repeats per combination and 72 measured runs.
What happened
T256-N16 was best at 6.348510 GH/s versus 6.332001 GH/s, ratio 1.002607. It failed the 1.01 gate and improved by at least 1% on 0/4 datasets. All nine binaries verified exact B32 nonces with zero discrepancies and zero spills.
Exactness and statistical controls
An initial runner-only failure was archived before measurement; the unused optional field read was removed and the whole campaign was resealed and repeated.
What the result means
The small maximum is compatible with tuning noise. T512-N32 remains the baseline.
Limitations
- One GPU architecture was tested.
- The search covers only nine registered geometries.
- A sub-gate maximum is not promoted.
Evidence trail
Rebuild all nine binaries, preserve the upstream arithmetic, repeat the balanced four-header matrix and enforce the 1% gate.
Canonical variants
CANONICAL-EXP-133PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.