Mach 10 ✔︎

Mach 10 on a laptop: what changes when hypersonic computation becomes fast, inspectable, and downloadable

A Human Blooming feature on Verification Hammer #02 — Mach 10 CUDA

Human Blooming’s Mach 10 campaign is not a claim that a laptop has replaced a national laboratory. It demonstrates something more immediately useful to industry: a tightly scoped numerical campaign can move from exact-reference checks and adversarial screening to an original 3-D CUDA field, while preserving the evidence needed for someone else to interrogate the result.

Evidence object	ECHELEN_MACH10_VERIFICATION_HAMMER_V2_COMPLETE.zip
ZIP SHA-256	3743c3fbec799bb9cb2dff015e7c7d8f49f3613b90f2ba7c732108cd05117d99
V10 grid	160 × 80 × 80 = 1,024,000 cells
V10 execution	1,000 iterations; 7.5743 s recorded runtime
GPU recorded	NVIDIA GeForce RTX 5090 Laptop GPU
V10 result	V10_3D_PASS; 13/13 declared falsification gates true
CPU–CUDA record	0.0 maximum discrepancy; [0,0,0] axis errors; repeated in 4/4 V9/V10 records
Original V10 payload	6 files; 27,491,217 bytes

The unusual part is not Mach 10. It is the distance between claim and evidence.

Mach 10 has been simulated before, on far more sophisticated solvers and on far larger machines. That is not the story here. The story is what happens when a small engineering organization treats computation as an auditable product rather than as a screenshot: the numerical result, the failure criteria, the histories, the original volumetric fields, the provenance record and the cryptographic identity of the files are packaged together.

Verification Hammer #02 is Human Blooming’s attempt to make that idea concrete. Its public evidence chain spans V1 through V10. V1 starts with a Mach 10 Euler reference checked against an exact oblique-shock solution. Later stages add paired comparisons, uncertainty, adversarial sweeps, equilibrium and finite-rate thermochemistry, reduced-order thermal-protection-system screening, conservation and mesh-refinement checks, then 3-D CUDA executions. V10 ends with a 160×80×80 field — 1,024,000 cells — preserved as NPZ and VTK, plus an original JSON summary, convergence history, dashboard and nested SHA-256 manifest.

EVIDENCE POINTER — Technical report, Executive Summary; EVIDENCE/V10_ORIGINAL/V10_3D_SUMMARY.json; EVIDENCE/V10_ORIGINAL/MACH10_3D_FIELDS.npz

Figure 1 — V1 public Mach 10 Euler fields. The shock structure is not only visualized; its angle is compared quantitatively with the exact weak oblique-shock solution.

Why speed matters — and what the evidence actually lets us say

The temptation is to summarize the project with a dramatic sentence: “Mach 10 in hours.” The evidence supports a more precise and, in some ways, more interesting statement. The principal V10-1000 execution records 1,000 iterations in 7.5743111 seconds. A second V10 run at 350 iterations records 5.4163727 seconds. The historical V9/V10 register preserves four 3-D executions, each ending in the PASS state for its version and each recording a maximum CPU–CUDA discrepancy of 0.0.

What the pack does not contain is a trustworthy stopwatch for the entire human project from V1 to V10. So the defensible industrial message is not that every model, test and document was created in a proven number of hours. It is that once a numerical workflow is compact, GPU-native and organized around falsifiable gates, individual screening runs can become seconds-scale events — fast enough to support an engineering loop measured in hours rather than waiting for a remote batch queue.

EVIDENCE POINTER — EVIDENCE/V7_V10_PUBLIC/EJECUCIONES_MACH10_PUBLICAS.json; V10_3D_SUMMARY.json

A consumer-class form factor, not a national supercomputer

The V10 original summary records the backend as CUDA and the device as an NVIDIA GeForce RTX 5090 Laptop GPU. NVIDIA’s current specification for that laptop GPU lists 10,496 CUDA cores and 24 GB of GDDR7 memory. That distinction matters. This is high-end hardware, but it is still a GeForce laptop GPU: a form factor available to an engineering team without procuring an institutional HPC cluster.

The broader industry is already moving in this direction. NVIDIA documents Sintavia using a single RTX PRO 6000 Blackwell workstation GPU to run a 30-million-cell, 300-iteration conjugate-heat-transfer case in seven minutes, compared with 88 minutes on a 24-core CPU. That case is larger and uses commercial industrial software; it should not be conflated with Human Blooming’s solver. What it shows is that local GPU-accelerated simulation is no longer a fringe proposition. Human Blooming’s contribution is to push the same decentralization logic toward a deliberately evidence-heavy hypersonic screening workflow.

External context: NVIDIA GeForce RTX 50 Series Laptop specifications; NVIDIA/Sintavia customer story, accessed September 2026.

The campaign begins where credible CFD should begin: with something that can be wrong

In V1, a two-dimensional conservative Euler calculation over a 7° compression ramp is compared with the exact weak oblique-shock solution. The exact shock angle is 11.38172551°. As the grid is refined from 160×64 to 240×96 and 320×128, the absolute angular error falls from 0.051999° to 0.025107° and then 0.016019°. On the finest grid the final residual is 3.877×10⁻¹² and the fitted shock line reports R² = 0.9997389.

This is important because NASA’s own CFD verification guidance draws a sharp line between verification and validation: verification asks whether the equations and their implementation are being solved correctly; validation asks whether the simulation represents physical reality, which requires experimental comparison. Human Blooming’s V1 sits on the verification side of that line. It does not validate a vehicle. It establishes an exact-reference anchor for the numerical chain.

EVIDENCE POINTER — EVIDENCE/V1_V6_PUBLIC/PUBLIC_DATA/PUBLIC_RESULTS_SUMMARY.csv; ATLAS_SIGMA_MACH10_PUBLIC_EVIDENCE_DOSSIER.pdf

Figure 2 — V1 pseudo-time convergence on the finest public grid. The residual is a numerical solve metric, not experimental uncertainty.

From one clean case to hostile envelopes

The campaign then changes character. V2 freezes public acceptance gates. V3 compares paired branches from the same base field and perturbation; the final objective falls from 0.503975 in the reference branch to 0.132556 in the candidate branch, a 73.70% reduction. The principal-to-refined temporal difference is 0.05326%, while density and pressure remain positive.

V4 moves into thermodynamics and reduced-order thermal-protection screening. The public record reports a five-species equilibrium-air stagnation temperature of 3219.8 K at 1 bar and a nominal candidate surface-temperature reduction of 29.4 K. Across 200 paired uncertainty cases, 200/200 preserve candidate < reference.

V5 is deliberately harder. It combines 256 scrambled Sobol points with four exact corners, varying atmospheric density, nose radius, Sutton–Graves constant, transfer amplitude, emissivity and TPS properties simultaneously. The result is 260/260 pairs with FULL < OFF. Peak surface-temperature reduction ranges from 5.53 to 90.15 K, with a median of 46.64 K; paired heat-load reduction ranges from 0.477 to 16.791 MJ/m². In a hostile transition case with doubled heating and emissivity 0.70, the model reaches 3116.82 K OFF and 3086.81 K FULL.

Figure 3 — V4 paired uncertainty screen: 200/200 candidate cases remain below their matched reference cases. This is robustness within the reduced thermal mapping, not material qualification.

Figure 4 — V5 adversarial dashboard. The useful result is twofold: the ordering survives 260/260 cases, while the hostile thermal regime exposes where real material data and testing become mandatory.

Thermochemistry and materials: useful bounds without pretending to have a flight-qualified material

V6 bridges the frozen-gas and equilibrium limits with a finite-rate relaxation family at constant total enthalpy. The temperature bracket is 3930.36 K to 3219.78 K. The reduced reactor conserves N/O inventory and reproduces its analytic solution with error below 10⁻¹⁰ mol. The FULL < OFF ordering survives catalytic increments of 0, 10, 20 and 30%; at +30%, peak surface temperatures are 2601.74 K OFF and 2573.33 K FULL.

The same stage introduces deliberately conservative engineering bounds. Under a hostile transition factor of 2.0 and catalytic factor 1.30, with emissivity 0.70, ablation onset at 2800 K and assumed heat of ablation of 30 MJ/kg, the calculated recession bound moves from 1.019 mm OFF to 0.966 mm FULL. A fully constrained thermoelastic bound spans 0.204–0.755 GPa depending on the assumed property set.

These are not material certification numbers. The technical note explicitly calls for temperature-dependent properties, support conditions, finite-element analysis and physical testing. That restraint is part of the value: the workflow carries the calculation far enough to define the next engineering question without pretending that a reduced-order model has already answered it.

Figure 5 — V6 finite-rate thermochemistry and reacting-wall screening. It connects gas-state assumptions to surface response and conservative recession bounds without claiming a moving-boundary ablation solution.

V7–V8: the bridge to conservation, transport and refinement

The earlier V1–V6 dossier correctly recorded V7 as pending at the time it was issued. The complete Hammer contains the later record: two V7 executions with SCREENING_COMPLETE and two V8 executions with V8_PASS. Each V8 record carries ten true gates covering the equilibrium solver, mass, momentum and energy conservation, species normalization, finite-rate response, positive transport, grid-error reduction, CUDA availability and a CPU–CUDA twin.

The preserved mesh series matter because they move the campaign beyond a single-resolution success. In one V8 execution, relative L2 temperature error against the finest grid drops from 1.535×10⁻⁴ at 48 points to 3.154×10⁻⁵ at 96 and zero at 192. In the other it drops from 8.680×10⁻⁵ at 64 points to 2.051×10⁻⁵ at 128, 4.230×10⁻⁶ at 256 and zero at 512.

EVIDENCE POINTER — EVIDENCE/V7_V10_PUBLIC/EJECUCIONES_MACH10_PUBLICAS.json; BARRIDOS_RELACIONADOS_PUBLICOS.json

V10: one million cells, original fields, and a result that can be opened

The culminating V10-1000 case declares Mach 10.0, T∞ = 220 K, p∞ = 2500 Pa, γ = 1.4 and R = 287.05287 J/(kg·K). Under that declared perfect-gas state, the speed of sound is about 297.342 m/s, so Mach 10 corresponds to about 2,973.42 m/s — 2.973 km/s, or roughly 10,704 km/h.

The domain is 0.7 m long in x and ±0.22 m in y and z, with an immersed body centered at x = 0.43 m. The grid is 160×80×80. The summary records 1,000 iterations, 0.0002532917 s of simulated physical time and 7.5743 s runtime. It ends at V10_3D_PASS with all 13 declared falsification gates true.

The important point is that 1,024,000 cells is not merely metadata. The NPZ contains x(160), y(80), z(80) and volumetric pressure, density, temperature, Mach, total-enthalpy-ratio and solid arrays with shape (160,80,80). The VTK independently declares DIMENSIONS 160 80 80 and POINT_DATA 1024000.

Figure 6 — Original V10 dashboard: pressure, density, temperature and Mach slices around the immersed body. Quantitative values are carried separately in the NPZ/VTK and JSON, so the figure is not the evidence by itself.

Checks that can be recomputed from the distributed field

Direct inspection of the stored arrays finds no NaN or Inf in the inspected volumetric fields. The extrema reproduce the JSON summary: maximum pressure 291,691.125 Pa, minimum density 0.0017103914869949222 kg/m³, maximum temperature 4620.0185546875 K, maximum Mach 10.000096321105957 and maximum stored total-enthalpy ratio 1.0002001523971558.

Recomputing T = p/(ρR) with the case value R = 287.05287 J/(kg·K) gives a maximum discrepancy of 0.0 against the stored temperature array at stored precision. Reflecting the pressure field about Y and Z likewise gives maximum differences of 0.0 at stored precision. These spatial symmetry checks are distinct from the [0,0,0] axis errors in the CPU–CUDA twin record.

The energy limiter is not a hidden switch

V10 adds three gates beyond V9: total_enthalpy_bound, static_temperature_bound and limiter_energy_audited. The summary records 81,411,756 corrected cell updates and 1,344.097211 J removed by the guard. Against 24,676.906961 J of initial domain energy plus 26,792.295700 J of processed inlet energy, the stored ratio is 0.026114591674817285 — 2.611% — and the arithmetic reproduces exactly.

The maximum pre-limit total-enthalpy ratio reaches 1.2851758. The declared final ceiling is 1.0002, while the stored float32 field reaches 1.0002001524, consistent with that ceiling at stored precision. This is a better engineering artifact than a bare boolean because it leaves a quantitative trail: how often the guard acted, how much energy it removed and what the final field contains.

EVIDENCE POINTER — V10_3D_SUMMARY.json → energy_guard; CONVERGENCE_HISTORY.csv

A non-monotonic residual — published rather than polished away

The V10-1000 convergence history contains 101 preserved samples: iteration 1 and periodic samples through 1,000. The residual is finite but non-monotonic, with transient excursions; it ends at 0.0319084749. The gate is explicitly called finite_residual, not residual_converged. That distinction matters. The pack does not define an independent asymptotic-convergence criterion for V10, and the article should not invent one.

More unusually, the earlier public convergence extract and the later sealed original can be compared row by row. All 101/101 iteration–residual pairs match exactly. That creates a continuity link between what was published earlier and what was later included as original evidence.

EVIDENCE POINTER — CONVERGENCE_HISTORY.csv SHA-256 348c2d1b80be0d60b4d20fe9d8c819a025bb9e9e4680bdf2400e3b72fa545e68; public extract SHA-256 8594c173c60e0231e7dbab3ec59bc1ea5f47f831209fdc21efac751ca72ee11d

The forensic layer: why the ZIP matters

The distributed object is not “the website” and not this article. It is the ZIP: ECHELEN_MACH10_VERIFICATION_HAMMER_V2_COMPLETE.zip. Its SHA-256 is 3743c3fbec799bb9cb2dff015e7c7d8f49f3613b90f2ba7c732108cd05117d99. Inside, SHA256SUMS.txt protects 37 objects; independent rehashing of the package produced 37/37 matches and zero mismatches. The manifest itself is the 38th file and, by definition, does not self-list.

V10 adds a second layer. EVIDENCE/V10_ORIGINAL/SHA256_MANIFEST.txt protects the five other V10 originals, and that nested manifest is itself protected by the outer manifest. The six V10 originals total exactly 27,491,217 bytes. Their hashes also match the historical inventory entries for the recorded V10-1000 output directory.

The six original V10 artifacts

Artifact	Bytes	SHA-256
CONVERGENCE_HISTORY.csv	9,625	348c2d1b80be0d60b4d20fe9d8c819a025bb9e9e4680bdf2400e3b72fa545e68
MACH10_3D_DASHBOARD.png	217,577	cfbb1b2eede181ab06359bb0fdbe1668c619a8175011361d7ce95dab9588d995
MACH10_3D_FIELDS.npz	2,683,966	53f156c8435a343ab7216ca79f402ea7d0573e52a3c5c19f07b79ed298e38908
MACH10_3D_FIELDS.vtk	24,576,552	d1e9a4efe1b193e32d10b1d751d2918e7f6afb551256c6f388b1ed7774574b70
SHA256_MANIFEST.txt	445	cce689ac9bcce4b255d82c8a7e174090db4be4230277f900014151f094cceaba
V10_3D_SUMMARY.json	3,052	eb1acecdc4d02de521803e834e7805ae67881641c96c5dd9cffefdd399917111

Where does this place Human Blooming?

Not “above NASA,” not “equivalent to a prime contractor,” and not at the same validation maturity as a program backed by wind-tunnel campaigns, reacting Navier–Stokes, LES, DSMC, materials databases and flight test. Those comparisons would collapse very different kinds of evidence into a single league table.

A more useful comparison is by capability. On rapid local GPU computation, Human Blooming is working in the same technological direction now visible in industrial accelerated-engineering programs: moving substantial simulation closer to the engineer and shortening the iteration loop. On verification culture, the campaign speaks the language used by NASA/AIAA guidance and by the High-Fidelity CFD Verification Workshop: exact or reference solutions, grid/iteration behavior, explicit test cases and transparent limitations. On artifact culture, it resembles the direction promoted by ACM reproducibility programs: preserve the artifacts, document them, make them available, and distinguish artifact availability from independent reproduction.

The differentiator is the combination. A directed review for the technical report found strong public precedents for each ingredient separately — formal CFD V&V, community verification cases, GPU-accelerated CFD, reproducibility badges and artifact repositories. It did not locate an exact public counterpart combining this particular Mach 10 V1–V10 ladder, analytic reference, paired and adversarial screening, thermochemical/TPS bounds, repeated CPU–CUDA records, an original 3-D field, an audited energy guard and multilayer SHA-256 sealing in one package. That is a search result, not a claim of world priority.

What this could change for industry

For an aerospace team, the immediate opportunity is not to replace high-fidelity certification analysis. It is to move more questions upstream. Which geometry is worth sending to an expensive solver? Which thermal assumptions dominate the result? Does a candidate ordering survive uncertainty? Where does a reduced model cross into a regime where material data become mandatory? Can a GPU and CPU implementation agree on a declared twin test? Can the exact result handed to a partner be identified months later?

Those are screening and evidence-management questions, and they occur before the most expensive part of an engineering program. If a compact local workflow can answer them quickly, the scarce resources — cluster time, specialist CFD, reacting-flow models, FEA, tunnel campaigns and material tests — can be focused on fewer, better-defined candidates.

Potential applications include early hypersonic configuration screening, shock-structure studies, thermal-protection trade studies, uncertainty envelopes, solver regression tests, GPU port verification, supplier/customer evidence exchange, and pre-test definition. The Hammer does not make all of those applications validated; it shows a workflow architecture that can support them at the appropriate screening level.

The most important limitation is also the reason the result is credible

V10 is a first-order Cartesian finite-volume Euler solver with a stair-step immersed body and calorically perfect frozen gas. The 3-D update does not include viscosity, turbulence, finite-rate chemistry, radiation or material response. V4–V8 add thermodynamic, chemical and wall layers as separate screening models; they do not magically turn V10 into a 3-D reacting Navier–Stokes solver.

The package therefore does not claim experimental validation, flight qualification, material qualification or independent third-party reproduction. NASA’s V&V framework is explicit that validation requires comparison with physical reality. ACM’s reproducibility framework is equally explicit that “results reproduced” means the main results have been obtained by a person or team other than the authors. Human Blooming has not claimed either threshold here.

That boundary should not be read as weakness. It tells an industrial reader exactly what is available now and exactly what the next evidence layer must be.

Human Blooming’s proposition: engineering velocity with an audit trail

The broader proposition behind Human Blooming is that advanced engineering should not be forced to choose between speed and traceability. Modern local GPUs make it possible to collapse some numerical loops from hours or days toward minutes or seconds; disciplined evidence packaging can make those fast loops reviewable rather than ephemeral.

Verification Hammer #02 is a concrete demonstration of that philosophy. The most striking number may be 7.5743 seconds for the recorded V10-1000 run. But the more consequential number may be 37/37: every object protected by the outer manifest rehashing correctly. Speed creates options. Evidence creates trust.

The industrial promise is therefore not “press a button and certify a hypersonic vehicle.” It is more practical: ask more engineering questions earlier, run them locally, attack the result with explicit gates, preserve the original bytes, and hand the next team something they can inspect rather than a slide they must believe.

A result designed to be challenged

At the end of the campaign, the strongest statement is not simply V10_3D_PASS. It is that the one-million-cell result is connected to multiple distinct consistency layers: an exact-reference beginning, refinement behavior, paired and adversarial tests, thermochemical and material bounds, repeated local CUDA records, inspectable volumetric fields, an algebraically consistent equation of state, recomputable Y/Z symmetry, quantitative energy accounting, public-to-original convergence continuity and a cryptographic chain tying the current originals to the historical inventory.

That is what makes this publication worth attention. Not because Mach 10 is new. Not because a GeForce laptop is secretly a supercomputer. But because the cost of asking a serious computational question is falling — and Human Blooming is arguing that the cost of showing your work should fall with it.

Verification box

Evidence Pack: ECHELEN_MACH10_VERIFICATION_HAMMER_V2_COMPLETE.zip

SHA-256: 3743c3fbec799bb9cb2dff015e7c7d8f49f3613b90f2ba7c732108cd05117d99

Outer manifest: SHA256SUMS.txt — 37 protected objects; 37/37 independently rehashed in the audit.

V10 original set: 6 files; 27,491,217 bytes.

Primary V10 summary: EVIDENCE/V10_ORIGINAL/V10_3D_SUMMARY.json — SHA-256 eb1acecdc4d02de521803e834e7805ae67881641c96c5dd9cffefdd399917111

Original volumetric field: EVIDENCE/V10_ORIGINAL/MACH10_3D_FIELDS.npz — SHA-256 53f156c8435a343ab7216ca79f402ea7d0573e52a3c5c19f07b79ed298e38908

Sources and context

Human Blooming internal evidence pack and reviewed technical report: Primary source for all Hammer-specific numerical results, hashes, filenames and limitations.

NASA Glenn Research Center — Overview of CFD Verification & Validation: https://www.grc.nasa.gov/www/wind/valid/tutorial/overview.html

NASA Glenn Research Center — Verification Assessment: https://www.grc.nasa.gov/WWW/wind/valid/tutorial/verassess.html

High-Fidelity CFD Verification Workshop — SciTech 2024: https://highfidelitycfdverificationworkshop.github.io/

ACM SIGSIM/PADS — Reproducibility and Artifact Evaluation: https://sigsim.acm.org/conf/pads/2024/blog/artifact-evaluation/

NVIDIA — GeForce RTX 50 Series Laptop specifications: https://www.nvidia.com/en-us/geforce/laptops/50-series/

NVIDIA — Sintavia aerospace accelerated-computing case study: https://www.nvidia.com/en-us/case-studies/sintavia-lightweight-aerospace-nvidia-accelerated-desig