Die
Individual die: functional units, clocks, leakage, thermal, I/O, stability.
Semiconductor hardware grading
The DeBruin Scale — Every Chip Gets a Grade.
A proposed independent grading language for semiconductor hardware — analogous to diamond grading: measure the hard parts, then give buyers a simple overall grade.
Research / standards idea. Not a live certification authority. Not a fab yield product. Not affiliated with any chip vendor. Thresholds TBD pending real hardware stats.
The problem
Manufacturers already test and bin chips. The same model number can still hide very different histories — nearly new vs years under heavy thermal load. Buyers lack a manufacturer-independent way to compare usable capability and condition.
Four certification levels
Package
Packaged / chiplet devices: interconnect, integrity, interfaces.
Accelerator
Complete AI accelerators — where memory and interconnect matter most.
System
Servers, appliances, rack-scale — plus power, cooling, multi-device reliability.
Four dimensions (0–100)
Compute
Usable capability vs expected spec under sustained load.
Capacity
Memory and related capacity, errors, bandwidth.
Connectivity
PCIe, NVLink/equivalent, chiplet and high-speed I/O health.
Condition
Stability, thermal/power behavior, age, degradation — critical for used gear.
Overall grade (AAA→F)
Conceptual meanings only — numeric cutoffs wait on statistically defensible testing.
| Grade | Meaning |
|---|---|
| AAA | Exceptional / essentially full-spec |
| AA | Excellent |
| A | High quality |
| B | Reliable compute-grade with measurable limits |
| C | Value compute; selected workloads |
| D | Salvage-grade; specialized applications |
| F | Not suitable for certified compute use |
Where it starts
Immediate focus is the secondary AI hardware market — finished accelerators first — not fab scrap. Model number alone does not answer residual value.
Business model (Idea·WIP): certification · labs · marketplace · API · fleet assessment. No “buy a certificate” CTA yet — the standard comes first.