EXP-066NEGATIVE
Canonical GPU engineering / SHA-256d

Constant σ0(W16/W17) hoisting

Moving invariant σ0(W16) and σ0(W17) from the nonce loop into CPU setup reduces GPU work.

Claim statusNO END-TO-END MINING ADVANTAGE DEMONSTRATED

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

Date2026-08-14
HardwareNVIDIA GeForce RTX 5080 · exact SHA-256d verifier
Run scopePreregistered campaign with sealed result and audit
ReproducibilityCompare emitted instruction families as well as balanced timing and exact hashes.
01 / Question & hypothesis

What was tested?

Moving invariant σ0(W16) and σ0(W17) from the nonce loop into CPU setup reduces GPU work.

02 / Scientific basis

Why the test is meaningful

Source-level invariance matters only if it changes generated code or runtime; modern compilers may already hoist it.

candidate passes σ0(W16),σ0(W17) as constantsΔSASS expected<0gate requires positive 4/4 timing
Constant σ0(W16/W17) hoistingVisual reading of the published metrics and gates for EXP-066; it summarizes the registered result, not a mining advantage.EXP-066 / CONSTANT Σ0(W16/W17) HOISTINGKERNEL RATIO0.998841SASS CHANGE−8 instructionsPOSITIVE HEADERS0 / 4
FIGURE / RESULT READINGVisual reading of the published metrics and gates for EXP-066; it summarizes the registered result, not a mining advantage.
03 / Method

How it was tested

Preserve nonce order and SHA arithmetic, move only two invariant expressions and run four unseen headers in 32 balanced measurements.

04 / Observed result

What happened

Kernel ratio0.998841
SASS change−8 instructions
Positive headers0 / 4

SASS fell from 2,507 to 2,499 instructions, but SHF/LOP3/IADD3 counts were identical. The candidate lost on 4/4 headers: kernel ratio 0.998841 and 95% interval 0.998082–0.999601.

05 / Validation

Exactness and statistical controls

Both binaries recovered the same B32 hit with zero discrepancies/spills and used 44 registers.

06 / Interpretation

What the result means

ptxas already hoisted the meaningful work; extra parameters do not improve the hot arithmetic.

Limitations

  • Only W16/W17 σ0 invariants were tested.
  • Source-line reduction is not execution reduction.
  • Equivalent constant combinations are closed post hoc.
07 / Reproduction

Evidence trail

Compare emitted instruction families as well as balanced timing and exact hashes.

Canonical variants

CANONICAL-EXP-142PREREGISTERED-CAMPAIGNSEALED-AUDIT

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