EXP-077EXPERIMENTAL
Canonical GPU engineering / SHA-256d

P8 energy via the cumulative NVML counter

Direct before/after energy-counter readings reduce integration error enough to decide P8 energy noninferiority.

Claim statusNO END-TO-END MINING ADVANTAGE DEMONSTRATED

Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.

Date2026-08-16
HardwareNVIDIA GeForce RTX 5080 · exact CUDA SHA-256d verifier
Run scopePreregistered paired campaign with sealed validation and audit
ReproducibilityRepeat the paired direct-counter protocol with fourfold work per run and unchanged binaries, headers policy and 1.003 gate.
01 / Question & hypothesis

What was tested?

Direct before/after energy-counter readings reduce integration error enough to decide P8 energy noninferiority.

02 / Scientific basis

Why the test is meaningful

A cumulative joule counter avoids reconstructing energy from sparse power samples, although short runs can still produce an interval too wide for a narrow noninferiority margin.

E_run=(NVML_mJ_after−NVML_mJ_before)/1000R_E=(E_P8/H_P8)/(E_P1/H_P1)promote iff headers and CI95 satisfy the 1.003 gate
P8 energy via the cumulative NVML counterVisual reading of the published metrics and gates for EXP-077; it summarizes the registered result, not a mining advantage.EXP-077 / P8 ENERGY VIA THE CUMULATIVE NVML COUNTERKERNEL SPEEDUP1.002947×ENERGY RATIO1.004283ENERGY CI950.996564–1.012061
FIGURE / RESULT READINGVisual reading of the published metrics and gates for EXP-077; it summarizes the registered result, not a mining advantage.
03 / Method

How it was tested

Measure 64 externally balanced paired runs with nvmlDeviceGetTotalEnergyConsumption before and after each process, retaining frozen P1/P8 work and exactness checks.

04 / Observed result

What happened

Kernel speedup1.002947×
Energy ratio1.004283
Energy CI950.996564–1.012061

The counter advanced in every run. P8 kernel speed was 1.002947×; direct energy/hash ratio was 1.004283, only 2/4 headers passed and CI95 was 0.996564–1.012061. Maximum temperature was 65°C.

05 / Validation

Exactness and statistical controls

B32, exact P1/P8 output, order balance and thermal gates passed. The interval crosses both parity and the noninferiority limit.

06 / Interpretation

What the result means

Direct counting works, but the energy verdict remains unresolved; four times more hashes per run are required before promotion or rejection.

Limitations

  • Processes remained too short for the registered precision target.
  • The point estimate alone cannot decide noninferiority.
  • No extrapolation to other GPUs or ASICs is supported.
07 / Reproduction

Evidence trail

Repeat the paired direct-counter protocol with fourfold work per run and unchanged binaries, headers policy and 1.003 gate.

Canonical variants

CANONICAL-EXP-153PREREGISTERED-CAMPAIGNSEALED-AUDIT

Source: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.