Historical H100 characterization and the December 2025 numerical-engine TestFort run. Not Luxi Book. Not a current product scoreboard.
This page presents two separate evidence classes: generic H100 hardware characterization and Lu(x)iEdge numeric-engine validation. The hardware tests establish the machine's measured capabilities; the Lu(x)iEdge tests establish results for defined Lu(x)iEdge workloads. They are not interchangeable. None of the figures below are Luxi Book, inference, or option-pricing results.
These OpenBenchmarking results characterize the H100 test system. They are not measurements of Lu(x)iEdge application throughput, Lu(x)iEdge energy efficiency, or Lu(x)iEdge numerical determinism.
| Hardware Test | Result (source units) | Notes |
|---|---|---|
| Double Precision (FP64) throughput | 85,819 GFLOPS (≈85.8 TFLOPS) | Generic hardware benchmark of the H100 80GB HBM3 test host (December 2025 characterization runs) |
| Hashcat SHA-256 hashing | 15.13 Billion H/s | Generic GPU hash benchmark of the test host - not a Lu(x)iEdge verification feature |
These are hardware results, not Lu(x)iEdge application benchmarks. Hardware: NVIDIA H100 80GB HBM3 (single board) in the December 2025 test host; run with the Phoronix Test Suite / OpenBenchmarking tooling. Public OpenBenchmarking result links for these characterization runs are being consolidated and will be added here; until then, treat these values as project-recorded hardware characterization only.
TestFort QA Lab independently evaluated a defined non-linear numerical workload, repeated hash consistency, CPU/GPU output agreement in the tested H100 environment, API load behavior, and sustained GPU execution. That evaluation did not certify every function, every platform, every backend, or every possible expression.
| Measured Result | Value | Scope |
|---|---|---|
| Aggregate throughput | 286.94 Billion Ops/s | Defined non-linear function suite, H100 test environment |
| Power efficiency | 2.35 Billion Ops/J | Defined test, sustained load |
| Total operations | 444.4 Trillion | Reported for the specified one-hour API load test, under TestFort's accounting method |
| Request errors | 0.00% | No failed requests during the specified load test - not a universal mathematical-accuracy claim |
| Repeated hash consistency | Identical hashes ×5 runs | CPU and GPU outputs agreed in the tested environment, for the tested workload |
Current full-model Qwen2-7B faithful CUDA correctness evidence and the separate TRADE H100 energy research are documented on the Proof page. Those are distinct evidence classes from the numeric-engine and hardware results shown here.