best_index
119418898
HUMANBLOOMING · PUBLIC EVIDENCE RELEASE
ECHELEN
PASS90
An exact addressable-gate benchmark showing how structural knowledge can reduce the amount of heavy computation delivered to the same NVIDIA GPU while preserving the validated result.
PUBLIC EVIDENCEOFFLINE VERIFIABLE3 RECORDED RUNSPASS90
This release documents the HumanBlooming ECHELEN PASS90 benchmark recorded on an NVIDIA GeForce RTX 5090 Laptop GPU. The comparison is not HumanBlooming versus NVIDIA. NVIDIA is the compute platform. The experiment compares a blind full-space path with an exactly addressable path using the same heavy scoring function.
The central result is structural: 128,000,000 candidates become 62,500 valid candidates before heavy scoring. That is an exact 2048× reduction in candidates and heavy scoring work. Across three recorded runs, the resulting wall-clock speedup ranges from 516.4896× to 561.7751×, with a 543.1528× mean, while preserving the consolidated winner and score.
Original evidence package included with this release.
SHA-256:
SHA-256:
995a728ca5e0d616ecba13b2c12269244ef4f6f2fc036599226cd14b79d5a3ad01
Not a faster GPU. A more selective use of the GPU.
ECHELEN PASS90 does not attempt to demonstrate that HumanBlooming can manufacture a faster GPU than NVIDIA, outperform NVIDIA as a company, or replace CUDA as a computing platform.
It investigates an earlier optimization question: if the structure of a problem tells us exactly which candidates can belong to the valid subspace, why spend heavy GPU computation on candidates that cannot contribute to the result?
NVIDIA provides extraordinary computational capacity. ECHELEN explores how exact knowledge of the problem can reduce how much of that capacity needs to be consumed.
In this benchmark, the innovation is not a claim that the RTX 5090 itself calculates 543× faster. The measured acceleration comes from giving the same GPU dramatically less heavy work to perform.
02
The benchmark at a glance
2048×fewer candidates heavy-scored
≈1024×smaller main explicit intermediate structures
543.1528×mean recorded wall-clock speedup
3 / 3runs preserve the validated result
The benchmark uses N=128,000,000, heavy_iters=1024, gate_mod=2048 and residue=18. The exactly valid subspace contains 62,500 candidates.
03
Same scoring function. Different search path.
Both computational paths use the same heavy scoring function: echelen_score(i, heavy_iters, seed).
GPU_RAW_BLIND_FULL
Scores all 128,000,000 candidate indices and then applies the validity gate.
GPU_QPE_ADDRESSABLE_GATE
Generates only valid indices using i = residue + j × gate_mod and scores those 62,500 candidates directly.
QPE did not change the solution. It changed the search path.
The valid fraction is exactly 1 / 2048. The work avoided by the addressable path is therefore 99.951171875% of the candidate scoring defined by the benchmark.
04
2048× less heavy scoring
128,000,000blind candidates
62,500addressable candidates
131,072,000,000blind heavy scoring iterations
64,000,000addressable heavy scoring iterations
The reduction is exact: both candidate count and heavy scoring iterations fall by a factor of 2048.
Before optimizing how to execute the work, ECHELEN removes exactly the work that the benchmark's structure proves cannot belong to the valid subspace.
05
A smaller explicit intermediate footprint
The blind path allocates a float32 scores array for 128,000,000 candidates: 512,000,000 bytes, approximately 488.281 MiB.
The addressable path allocates valid_scores and valid_indices for 62,500 candidates. Together these explicitly allocated structures total 500,000 bytes, approximately 488.281 KiB.
≈488.28 MiBblind explicit scores structure
≈488.28 KiBaddressable explicit structures
≈1024×ratio for these explicit structures
Not claimed1024× less total GPU memory
The ≈1024× figure applies to the principal explicitly allocated result/intermediate structures compared in the published code. It is not a claim about total GPU memory consumption.
06
Three recorded runs above 500×
| Run | GPU speedup | Correct result | PASS_90X_GPU |
|---|---|---|---|
| X90_C_ORIGINAL | 516.4896× | True | True |
| X90_C_REPLICA_SEED28 | 561.7751× | True | True |
| X90_C_REPLICA_SEED29 | 551.1938× | True | True |
516.4896×minimum recorded
543.1528×mean recorded
551.1938×median recorded
561.7751×maximum recorded
The minimum recorded result is approximately 5.74× above the benchmark's 90× PASS threshold. The mean is approximately 6.04× above that threshold.
07
The result is preserved
PASS90 does not define success as speed alone. The benchmark requires the addressable path to preserve the winner and a compatible score.
best_score
425.7733459472656
Gate check
119418898 mod 2048 = 18
Valid position
j = 58,310, inside the 62,500-candidate valid subspace.
Eliminating 99.951% of the scoring is valuable because the winner is not eliminated.
08
What the wall-clock number includes
The official speedup uses elapsed_s measured with time.perf_counter(). The measured interval includes the kernel execution, argmax, retrieval of the winning index and score, and the synchronization required to make the result available.
The benchmark performs a warm-up and uses repeat=3 for each path, retaining the minimum elapsed time. The reported path timing is therefore a best-of-3 statistic.
The measured speedup captures only part of the ideal 2048× structural reduction: approximately 25.22% to 27.43%, with a mean of approximately 26.52%. Fixed costs such as launches, reduction, argmax, synchronization and result retrieval remain.
The observed speedup is large, but not magical: it is the physical translation of an even larger algorithmic reduction.
09
Exact selection, not probabilistic pruning
The gate defines the valid domain exactly. The addressable path does not guess which candidates are likely to matter and does not introduce a precision tradeoff to decide what to discard.
It enumerates the same valid congruence class that the blind path recognizes only after scoring the full space.
In PASS90, work is not avoided because it probably does not matter. It is avoided because the benchmark definition establishes that it cannot belong to the valid subspace.
10
NVIDIA is the platform, not the adversary
The NVIDIA GeForce RTX 5090 Laptop GPU is the compute platform used by HumanBlooming in this benchmark. The experimental adversary is the blind full-space computational path running on the same NVIDIA platform.
HumanBlooming does not claim that ECHELEN is “better than CUDA,” that it “beats NVIDIA benchmarks,” or that HumanBlooming is “543× faster than NVIDIA.”
In this benchmark, ECHELEN's advantage is not a more powerful GPU. It is delivering a 2048× smaller heavy-scoring problem to the same GPU.
The defensible interpretation is complementary: NVIDIA hardware maximizes the ability to execute computation; ECHELEN investigates how exact structural knowledge can reduce the amount of computation that needs to be executed.
11
Public verification package
The release includes the frozen benchmark source, structured summary, replication report, SHA-256 inventory, offline verifier, Windows and Linux/macOS launchers, claim boundary and evidence license.
ECHELEN PASS90 Public Verification V1
SHA-256:
SHA-256:
995a728ca5e0d616ecba13b2c12269244ef4f6f2fc036599226cd14b79d5a3adPASS90_PUBLIC_EVIDENCE_VERIFIED=YES
FILES_SHA256_VERIFIED=10
RUNS_VERIFIED=3
SPEEDUPS=516.4896x,561.7751x,551.1938x
WORK_SAVED_PERCENT=99.951
HISTORICAL_RAW_DIRECTORIES_INCLUDED=NO
PUBLIC_VERIFICATION_PASS=YESThe internal SHA256SUMS.txt covers ten published artifacts and verifies 10/10.
12
Do not trust this page. Verify the evidence.
The original evidence ZIP is physically included beside this HTML page. Download it, calculate its digest and execute the included verifier offline.
1. Verify the ZIP
sha256sum ECHELEN_PASS90_PUBLIC_VERIFICATION_V1.zipExpected:
995a728ca5e0d616ecba13b2c12269244ef4f6f2fc036599226cd14b79d5a3ad2. Extract it
unzip ECHELEN_PASS90_PUBLIC_VERIFICATION_V1.zip3. Verify the internal inventory
cd PASS90_PUBLIC_RELEASE_V1
sha256sum -c SHA256SUMS.txt4. Run the offline verifier
python verify_public_evidence.py13
Claim boundary
This release supports the following benchmark-specific claim:
ECHELEN/QPE recorded a replicated 516.49× to 561.78× GPU speedup, with a 543.15× mean, in the named addressable-gate benchmark on an NVIDIA GeForce RTX 5090 Laptop GPU, while returning the same validated winner and score as the blind full-space calculation.
Not claimed. Universal 500× acceleration for arbitrary workloads; execution on quantum hardware or quantum advantage; a cryptographic break; independent third-party replication; or byte-for-byte historical raw validation without the original OUT_PASS90_* directories.
14
What survives scrutiny
PASS90 is a benchmark-specific result, not a universal performance promise. Its strongest evidence is the combination of four properties: an exact reduction in heavy work, a reduction in explicit intermediate structures, a large observed wall-clock acceleration and preservation of the validated result.
For exactly addressable useful subspaces, reducing the problem before CUDA can be more decisive than optimizing the same volume of work.
The broader research question is therefore not whether every workload can achieve this result. It is where else sufficient structure exists to identify, exactly and in advance, the work that deserves expensive computation.
15
A different question for high-performance computing
Conventional optimization asks how to execute a workload faster. ECHELEN adds an earlier question: which part of the workload needs to exist at all?
The fastest operation may be the one the structure of the problem proves no longer needs to be executed.
This ECHELEN PASS90 release is the public evidence for that benchmark.
Disclosure
The public package does not include the original historical OUT_PASS90_* raw directories. The recorded historical manifest SHA-256 values are commitments, but byte-for-byte revalidation of those historical raw directories is not possible from this ZIP alone.
The release is not an independent third-party replication. Reproduction requires a compatible CUDA/CuPy environment and suitable NVIDIA hardware.
HumanBlooming · ECHELEN PASS90 · Public Verification V1.