Skip to content

How we test

Nine stages. 24 to 72 hours. One signed result.

Every card we buy runs the same bench on our own test servers: Ubuntu LTS, NVIDIA’s datacenter driver, DCGM and CUDA. One orchestrator runs the stages in order, records time series, and hands the results to the signer.

01 — The bench

Most of the time is burn-in. That’s the point.

Grade B
Burn-in 24 h
≈ 28 h
Grade A
Burn-in 72 h
≈ 76 h
0 h20406080 h
Stages 01–06, up to 4 hStage 07 burn-in at TDP
No.StageToolDurationPass criteria
01IntakeBarcode and serial scan, photos of both sides, connectors and label, weight5 minSerial matches the probe report. No physical damage.
02Identitynvidia-smi -q1 minBoard part number and VBIOS family match the SKU.
03Health snapshotECC, row remapper, PCIe link1 minNo remap failure, no pending remaps, link at full generation and width.
04NVIDIA diagnosticsdcgmi diag -r 3 (and -r 4 for Grade A)30–90 minAll tests pass.
05Memory stressDCGM memory tests and HBM stress1–2 hZero uncorrectable errors. No new remapped rows.
06PerformanceBF16/FP16 GEMM, HBM bandwidth, PCIe host↔device with nvbandwidth20 minAt least 95% of our reference for the SKU.
07Burn-ingpu-burn or DCGM targeted power at TDP24 h (B) · 72 h (A)No Xid errors, no ECC growth, temperature under limit, no thermal throttling.
08Re-snapshotRepeat of stage 031 minNothing changed since stage 03.
09Grade and signPlatform signerInstantGrade A, B, C or Reject, signed with ed25519.

Thresholds are where we start, and will be tuned against our own warranty outcomes. Performance is measured against a reference we record on known-good cards of the same SKU, not the spec sheet.

02 — Why these tests

Built on published failure data.

During a 54-day snapshot of Llama 3 pre-training on 16,384 H100 GPUs, Meta logged 268 GPU-related interruptions. Each one maps to a stage on our bench.

Source: The Llama 3 Herd of Models (Meta, 2024), Table 5. These were interruptions on new cards under full training load, and not all were permanent failures.

Faulty GPU148 · 55%04 · 06 · 07
GPU HBM3 memory72 · 27%03 · 05
GPU SRAM memory19 · 7%03 · 05
GPU System Processor17 · 6%02 · 07 (Xid 119/120)
Silent data corruption6 · 2%04 at -r 3 / -r 4
Thermal interface, sensor6 · 2%07 temps, throttling

03 — Grades and warranty

What each grade promises.

GradeRequirementsWarranty
Grade A72 h burn-in, zero remapped rows, at least 98% of reference performance.6 months included. 12 months optional, paid.
Grade B24 h burn-in, a few correctable remapped rows that stayed stable, at least 95% performance.6 months.
Grade CWorks, but has a history: uncorrectable remapped rows that are now stable, or minor throttling. Sold at a discount.3 months, or sold as-is.
RejectRemap failure, a failed diagnostic, uncontained ECC errors, or under 90% of reference performance.Not sold. Returned to the seller or sent to an EPR-registered recycler, with the recycler certificate on file.

Claims

Open an RMA from the buyer portal. Attach the card’s Sentinel snapshot if you run it, and we can ship a replacement before the faulty card comes back.

Renewal

When warranty ends, Sentinel data prices the renewal. A card that has run clean costs less to cover.

04 — Reference sheet

What we check each card against.

Power, bandwidth and link speed must match the SKU’s baseline. A relabelled or emulated card can’t fake them.

SXM modules are tested inside their HGX server, so for now we buy and sell SXM as whole servers. Loose-module testing comes once volume pays for a spare baseboard.

ModelTDPMemoryBandwidth (spec)PCIe
H100 PCIe 80GB350 W80 GB HBM2e≈ 2.0 TB/sGen5
H100 SXM5 80GB700 W80 GB HBM3≈ 3.35 TB/sGen5
H200 SXM 141GB700 W141 GB HBM3e≈ 4.8 TB/sGen5
A100 80GB PCIe300 W80 GB HBM2e≈ 1.9 TB/sGen4
A100 80GB SXM4400 W80 GB HBM2e≈ 2.0 TB/sGen4
L40S 48GB350 W48 GB GDDR6≈ 0.86 TB/sGen4