Current page: Home
Check a record

Signed condition records for datacenter GPUs

What a GPU remembers.

Every datacenter GPU keeps a count of the damage it has seen. Voltry reads it, signs it, and keeps the record where anyone can check it.

See a real record Eight real cards, read and signed

No account needed. 10,000+ GPUs read.

Real cards

Newer doesn't always mean better.

Of our eight published cards, the oldest is the clean one. A build date says when a card was made, not how it was used.

Clean 14 reads · 87 days

NVIDIA A100-PCIE-40GB

Built Sep 2020 · serial ···3611

Lifetime correctable
0
Lifetime uncorrectable
0
Rows swapped for spares
0
Banks with no spares left
0 of 640

Signed at each read · Provider withheld

Built Sep 2020. 14 reads over 87 days, every counter at zero.

Failing 8 reads · 30 days

NVIDIA A10

Built Apr 2022 · serial ···3197

Lifetime correctable
5,605
Lifetime uncorrectable
373,598
Rows swapped for spares
8
Banks with no spares left
1 of 192

Signed at each read · Provider withheld

Built Apr 2022, 19 months later. 373,598 uncorrectable errors and a bank with no spare rows left.

Eight real cards, by the month they were built

Eight real cards by build month and condition Each dot is one card, placed by the month it was built, from 2020 to 2024, and by its condition at the last read. The oldest card, built in September 2020, is clean. A100 A100 A10 A10 A100 L40S L40 H100
Age is not condition. See all eight records
Clean 14 reads · 87 days

NVIDIA A100-PCIE-40GB

40 GB HBM2e · Built Sep 2020

Zero on every counter, every read.

Open the record

Hover or tab to any dot for its record.

Wear signals

19 wear signals, read from the card and signed.

Counters the card keeps about itself, in four groups. The reader only looks, and nothing on the card changes.

GPU read 1 read 2 7memory 4board and power 5links 3between reads
Line drawing, not to scale. Pick a group to light the parts it is read from.

Memory 7 signals

Counted on the card itself and kept for its whole life, across every owner and reboot. This is where wear shows first.

Lifetime correctable errors
Bit flips the card caught and fixed on the spot.
Lifetime uncorrectable errors
Errors the card could not fix, where data was lost.
Rows repaired after correctable errors
Rows retired for good after repeated fixed errors, each replaced by a spare.
Rows repaired after uncorrectable errors
Rows retired for good after an error that lost data, each replaced by a spare.
Repair pending
A row swap that only takes effect after the GPU is reset.
Repair failure
A swap the card could not make because the bank had no spare rows left.
Spare reserve, bank by bank
How many spare rows each memory bank still holds, as the card reports them.

Between reads 3 signals comes with tracking

Lifetime counts only go up. Reading the same card again shows whether anything moved.

Movement between reads
Any rise in a lifetime count between two signed reads, the clearest sign of wear in progress.
GPU resets between reads
A drop in the card's energy counter between reads marks a reset, which a pending repair needs.
Counts never going down
A lifetime count that drops would mean wiped history, a swapped card or a bad read.

Board and power 4 signals

A card can have spotless memory and still be held back by its board or the server around it.

Throttle reasons
Why the card is holding its clocks down at the moment of the read.
Temperature at idle
GPU temperature against the point where the card slows itself.
Power limit
Idle draw and the enforced limit against the factory default, so a capped card shows.
Clock-limited time
Time spent clock-limited since the driver loaded, recorded but not graded.

Links 5 signals

PCIe connects the card to the server. NVLink connects GPUs to each other on SXM boards.

PCIe width
Lanes in use against what the card supports.
PCIe generation
Link speed against the card's maximum.
PCIe resent packets
Packets the link had to send again, which rise as a connection degrades.
NVLink links up
GPU to GPU links active out of the total the card has.
NVLink resends and recoveries
Transfers sent again and link resets between GPUs.

NVLink signals apply to cards that have NVLink.

Tracking

One read is a snapshot. Tracking shows wear moving.

A GPU keeps lifetime wear counters on itself. Read it again and again and you see the wear move.

1 / 4

NVIDIA A100 SXM4 80 GB built Apr 2022

NVIDIA A100 SXM4 80 GB · serial ···3343 · 7 signed reads over 31 days, Sep 3 to Oct 4 · no GPU reset · Provider withheld.

NVIDIA A10 built Apr 2022

NVIDIA A10 · serial ···3197 · 8 signed reads over 30 days, Aug 30 to Sep 29 · 3 GPU resets between reads, marked · Provider withheld.

Two listings. Both built Apr 2022.

Step 1 of 4

A listing tells you the model and the year.

Both of these cards were built in April 2022. Same year, same kind of listing.

Nothing in a listing says what the memory has been through.

Step 2 of 4

One read is a snapshot.

The reader pulls the wear counters the GPU keeps on itself and signs the read. One read gives one number per counter.

The A100 shows 1 bit flip it fixed itself. The A10 shows 167,050 errors it could not fix. True, and frozen in one moment.

Step 3 of 4

Tracking shows wear moving.

The A100 held at 2 for weeks, then reached 755 by Oct 3. A row swap has waited for a reset since Sep 6.

The A10 went from 167,050 to 373,598 uncorrectable errors in 30 days.

A10: +67,055 in 2.3 days

Step 4 of 4

Counts never go down.

These are lifetime counters. They only rise, and a GPU reset never lowers them.

The A10 was reset 3 times between our reads, and its count kept climbing. The A100 was not reset once in 31 days, so its row swap is still waiting.

If a read ever came back lower than the last signed read, Voltry would flag it, not plot it.

Lifetime counts only rise.

The seal

Signed the moment it is read.

Every record is tamper-evident. Change one byte and the signature no longer matches.

L40S · ···0230 · one signed read

Model
L40S
Serial
···0230
Uncorrectable errors
2
Rows repaired after uncorrectable
1
Repair pending
1
Read
Sep 30, 2026, 08:41 UTC
Canonical bytes {"model":"L40S","read_at":"2026-09-30T08:41:45Z","repair_pending":1,"rows_repaired_uncorrectable":1,"serial_tail":"0230","uncorrectable":2}

Signature matches

ECDSA P-384

SHA-384 of the line above cca7d2e6…f57e6ed9

Values from a real signed read, in short form. The signature itself is not shown.

  1. 01 Read The reader takes the card's own wear counters.
  2. 02 Sign The read is put in one canonical form and signed on the machine with an ECDSA P-384 key, the moment it is taken.
  3. 03 Record It is filed in an append-only registry under the SHA-384 of its exact bytes. A root is published daily.
  4. 04 Verify Anyone can check it offline with the open verifier.

The signature proves the record has not changed since it was signed. It does not yet prove which machine held the key. Hardware-bound signing is our next step.

Where to start

Pick the door that sounds like you.

  • Buying GPUs

    Are these cards okay?

    Ask the seller for a Voltry record, then check it here in a second.

    Check a record No account needed

  • Selling GPUs

    Will a record help me sell?

    A signed record travels with each card, so the buyer sees the facts before they ask.

    See what a record shows One page, one price

  • Running a fleet

    Can I see wear coming?

    Tracking reads the same cards again and shows when a counter moves.

    See tracking 100 days included

On a Mac? Voltry for Mac prints a free, read-only health report for one machine. See Voltry for Mac

Pricing

Less risk, for a rounding error.

$1 per GB of memory for a signed record. 100 days of tracking included.

Our fee never depends on what we find.

See pricing Every NVIDIA GPU, one table
See a real record Eight real cards