Know a GPU's history before you buy it.

Every data center GPU gets a health score from 0 to 100 %, read from its own logs: memory errors, Xid events, link faults, throttling and failed diagnostics. Every deduction is shown, so buyer and seller look at the same evidence.

Get a free pilot report See the code

Sample reportnode-00307 · 5 of 8 GPUs

H1000000:1b:00.0

97.6 %

Age −2.4

excellent

H2000000:43:00.0

88.1 %

Link trained at x8: −10 · age −1.9

good

H1000000:25:00.0

74 %

Xid 48 and its row remap: one uncorrectable incident

caution

H1000000:9a:00.0

30 %

Memory banks with no spare rows left

do not sell

H2000000:c1:00.0

10 %

Failed matrix-math test: possible silent data corruption

do not sell

Example from a synthetic fleet. Bands: 90+ excellent · 75+ good · 50+ caution · below 50 do not sell.

How it works

Five sources, one language

The logs already sit on the servers being sold. We read them before decommissioning, translate every vendor's wording into one set of concepts, and tie each event to a GPU by its PCIe address. Nothing is installed on the servers; the tool reads copied files, offline.

IPMI SELBMC events, binary and text, several vendor dialects

Kernel logNVIDIA Xid codes, PCIe AER, MCE, EDAC

DCGM telemetryECC counters, row remapping, NVLink CRC, throttling

Diagnosticsdcgmi diag results, including matrix-math tests

PCIe treeMissing and degraded links, switch versus GPU

What moves the score

Rules from fleet practice

The score starts at 100 and loses points in six areas, each with a ceiling. A few findings cap the score outright, whatever else the card's history shows.

Finding	Effect	Score
Failed matrix-math diagnostic	Possible silent data corruption: keep out of any cluster	≤ 10 %
Row remapping failed, spare rows exhausted	The next bad memory row cannot be repaired	≤ 10 %
Memory banks with no spare rows	No room left for repairs	≤ 30 %
Repeated uncorrectable memory errors	Two or more Xid 48 or row-remap incidents	≤ 49 %
One uncorrectable memory error	Xid 48 and the remap it triggers count as one incident	≤ 74 %
Severe Xid: bus loss, GSP, init failure	8 points each, up to 25 for the area	−8
Link trained below spec	From the PCIe tree, e.g. x8 instead of x16	−10
Fleet-wide PCIe storm	One cause, not thousands of bad cards: not charged to the GPUs	0

Who it is for

Where a card changes hands

Brokers and marketplaces

Both sides of the deal price the same card from the same evidence. Fewer disputes after delivery.

Buyback and ITAD

Grade a fleet before pickup, while its logs still exist. Quote with fewer surprises at intake testing.

Operators and lenders

Know which cards to sell, which to keep, and what collateral is really worth.

What it is not

Stated limits

History, not a forecast. The score describes what happened to a card. It is not a probability that the card will survive a given time.
Tested on synthetic fleets. Real logs differ; each new fleet starts with tuning its vocabulary.
Starting weights. Set from operations practice and to be calibrated on real failure data.
No logs, no history. A card whose logs were wiped can only be graded by testing.

Free pilot

One fleet, one report, no cost

Send logs from a batch you are buying, selling or decommissioning. You get a health card for every GPU and the fleet summary. Logs stay confidential and are used only for your report.

de.mille@outlook.com