Mapped-host cancellation checked once per nonce
A host-mapped epoch checked for every nonce can interrupt stale GPU work without destroying useful throughput.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
A host-mapped epoch checked for every nonce can interrupt stale GPU work without destroying useful throughput.
Why the test is meaningful
Host-visible mapped memory offers immediate control semantics, but a system-call-free check can still impose a severe device-memory and coherence cost in the hottest loop.
candidate=P8-C1 / baseline=P8-C0retention=throughput(C1)/throughput(C0)gate: retention≥0.999 and response≤15 msHow it was tested
Compile exact P8 controls with and without a volatile mapped-host epoch read per nonce; require B32 correctness, then run the preregistered paired campaign only if the pilot is operationally viable.
What happened
The candidate exceeded 60 seconds for work expected to take about 2 seconds, a penalty of at least 30×. The campaign was stopped before unnecessary measurements.
Exactness and statistical controls
Exact arithmetic and the interruption path were checked, but the operational-retention gate failed decisively at pilot stage.
What the result means
Per-nonce mapped-host polling is rejected. Faster cancellation needs a control value resident in device memory and checked at a coarser cadence.
Limitations
- The early stop does not estimate a precise slowdown ratio.
- The verdict applies to this mapped-memory mechanism and CUDA stack.
- No mining-success advantage is claimed.
Evidence trail
Build P8-C0 and P8-C1 from the sealed source, verify B32, and compare equal work; stop if the registered 0.999 retention gate is already impossible.
Canonical variants
CANONICAL-EXP-162PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.