Public standard (draft)
A proposed independent grading and certification standard for semiconductor hardware. Complex measurements stay rigorous; buyers also get a simple, recognizable overall grade.
Measurement standard, not marketing. Not a live certification product. Numeric thresholds TBD pending statistically defensible hardware testing. No personal email published on this site.
The DeBruin Scale could provide a common language for manufacturers, data centers, brokers, recyclers, refurbishers, AI infrastructure companies, and secondary-market buyers — analogous to diamond grading.
Semiconductor manufacturers already test and bin chips extensively. A device that cannot meet one product’s requirements may qualify for a lower-performance product. Other devices enter refurbishment, recycling, or secondary markets.
Buyers still lack a universal, manufacturer-independent way to compare condition and usable capability. Two accelerators with the same model number can have very different histories — nearly new vs years under heavy thermal load. The DeBruin Scale would attempt to make that difference measurable.
The standard should explicitly identify the level of hardware being certified.
Individual die: percentage of functional compute units, defect map, maximum stable frequency, voltage requirements, leakage, thermal behavior, I/O functionality, stability under sustained load. Memory capacity normally N/A unless integrated.
After packaging / chiplets / interposers: die-to-die connectivity, package integrity, interconnect performance, thermal characteristics, chiplet functionality, integrated memory interfaces.
Complete accelerator (e.g. NVIDIA H100/H200/B200-class or AMD/Intel peers). Memory belongs here: compute, HBM/VRAM capacity and integrity, bandwidth, interconnect (PCIe/NVLink), thermal, power efficiency, stability, error rates, device health/history when available.
GPU servers, AI appliances, rack-scale, refurbished nodes, wafer-scale systems — plus networking, cooling, power delivery, storage, multi-device reliability.
Usable computational capability vs expected spec: functional units, stable clocks, sustained throughput, tensor/AI engines, error-free workload execution, degradation.
Usable memory and related capacity (HBM/VRAM, errors, bandwidth, cache). N/A or die-appropriate substitute for bare die.
PCIe, NVLink/equivalent, memory interfaces, chiplet/interposer links, high-speed I/O, error rates, bandwidth vs spec.
Stress-test stability, thermal/power behavior, error history, age, operating hours when verifiable, thermal cycling, degradation indicators, physical/package condition. Especially valuable for used/refurbished hardware.
| Grade | Conceptual meaning |
|---|---|
| AAA | Exceptional / essentially full-spec quality |
| AA | Excellent |
| A | High quality |
| B | Reliable compute-grade with measurable limitations |
| C | Value compute; selected workloads |
| D | Salvage-grade; specialized applications |
| F | Not suitable for certified compute use |
Exact numeric thresholds must await statistically defensible hardware testing.
Example only — not an issued certificate and not a claim about any specific device.
Immediate opportunity is less fab scrap and more the secondary AI hardware market: hyperscaler → enterprise → secondary → specialized → recycling. Model number alone does not answer residual value.
Certification fees · testing software licenses · certified labs · marketplace transaction fees · health/yield/value data · API for marketplaces/insurers/fleet ops · enterprise fleet assessment.
Progression: GPU testing → certification → secondary marketplace → industry standard → broader semiconductor certification.
No “buy cert” CTA on this site. Credibility first on used/refurbished/decommissioned compute; fabs as future partners. Do not depend on manufacturers releasing partial silicon.
Methodology · test database · RUL/failure models · lab network · marketplace recognition · cryptographic certificates · eventual industry adoption. The name alone is not the moat.
Start with finished AI accelerators. Build a software suite producing Compute / Capacity / Connectivity / Condition on identical GPUs (new vs heavy-use vs refurbished vs known issues). Correlate with reliability, performance, power, and resale value.
The DeBruin Scale — Every Chip Gets a Grade.
Alt line: The Independent Standard for Semiconductor Quality.
Ultimate vision: between perfect silicon and scrap lies economic value. The DeBruin Scale exists to measure, certify, and make that value visible.