P32 block-shared epoch cancellation
One global epoch load per block cached in shared memory can outperform the warp-broadcast cancellation design.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
One global epoch load per block cached in shared memory can outperform the warp-broadcast cancellation design.
Why the test is meaningful
A block-shared value reduces global loads further, but synchronization around shared state can cost more than the saved memory traffic.
epoch_block=global_load→shared_memoryretention=throughput(block)/throughput(P32)selection requires retention≥0.9995 and response≤15 msHow it was tested
Add a single block-level global load, shared-memory broadcast and barrier to exact P32; perform 2^32 validation, 24 balanced benchmarks and four cancellation trials.
What happened
All responses passed at 7.13–12.05 ms, with zero exactness discrepancies. Retention was 0.999277 with CI95 0.998796–0.999759 and 0/4 favorable headers.
Exactness and statistical controls
B32, factors, 44 registers and zero spills passed. The barrier-based candidate failed the frozen throughput selection gate.
What the result means
Block sharing improves response consistency but not throughput. The warp design remains the lighter candidate for replication.
Limitations
- Barrier cost is architecture-specific.
- The experiment compares one block geometry.
- Passing cancellation does not establish mining advantage.
Evidence trail
Rebuild the sealed P32 baseline and block-shared variant; preserve full audit coverage and report every paired header and cancellation trial.
Canonical variants
CANONICAL-EXP-166PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.