Check a record
10,000+GPUs read
Every wear signal on record

What a GPU tells us about its own wear.

Voltry reads these counters straight from the card and signs every read the moment it is taken. Nothing on this page is estimated. Each signal is shown where it lives on the card, with what it means when it moves.

wear signals, read from the card and signed
See eight real cards, each read over 30 days or more →
Links: NVLink links up · resends · recoveries Board and power throttling · heat · power cap Memory errors · repairs · spare reserve Links: PCIe width · speed · resent packets GPU
  • Links: NVLinklinks up · resends · recoveries
  • Board and powerthrottling · heat · power cap
  • Memoryerrors · repairs · spare reserve
  • Links: PCIewidth · speed · resent packets
Memory · 7 signals

Where wear shows first

GPU memory holds data as tiny charges in rows of cells. Error-correcting code checks every chunk as it is read. Fixed errors, swapped rows and lost data are counted on the card itself and kept for its whole life, across every owner and reboot.

A memory stackOn HBM cards (A100, H100, H200, B200) memory sits in stacks beside the GPU. GDDR6 cards (A10, A40, L4, L40S) use separate chips. Both work the same way below.
Split into banksMemory is divided into banks: 192 on an A10, 640 on an A100, 2,560 on an H100 SXM. Each bank keeps a few spare rows of its own.
One bank, rows of cellsWhen a row keeps failing, the card retires it and a spare row takes its place. The spare rows are held back below the dashed line.

How memory damage escalates, and the counter that catches each step

    Normal wear and tearCard can no longer heal itself

    Spare reserve, bank by bank

    The card reports how many of its banks sit at each spare level. A card fails a repair when the bank holding the bad cells runs out of spares, however untouched the other banks are. A total spare count across the card would hide exactly that, so the record shows the levels as the card reports them.

    Also read: since-boot errors. The same correctable and uncorrectable counts since the driver last loaded. They separate old history from errors happening now.

    Between reads · 3 signals

    One read is a snapshot. Two reads show wear moving.

    Lifetime counts only go up. Reading the same card again shows whether anything moved, whether the GPU was reset in between, and that nothing was wiped.

    Movement between readsAny rise in a lifetime count between two signed reads. The clearest sign of wear in progress.
    GPU resets between readsThe card's energy counter restarts when its driver reloads, so a drop between reads marks a reset. A pending repair needs one to take effect.
    Counts never going downA lifetime count that drops would mean wiped history, a swapped card or a bad read. The record checks every pair of reads.
    Board and power · 4 signals

    Trouble that is not in the memory

    A card can have spotless memory and still be held back by its board or by the server around it. These are read at the same moment as the memory counters.

    Throttle reasonsWhy the card is holding its clocks down at the read. A hardware slowdown or a power brake while the card is idle and cool points at the board or the server.
    Temperature at idleGPU temperature against the point where the card slows itself. Mostly a reading of the host's cooling.
    Power limitIdle draw, the enforced limit and the factory default. A limit below the default means the host capped the card.
    Clock-limited timeTime spent clock-limited since the driver loaded, for power, heat, reliability or board reasons. Recorded but not graded: on some models it exceeds elapsed time.
    The cards

    Eight real cards, each read over 30 days or more

    Each map shows every memory bank on that card. Each line is the signal that matters most for it, across every signed read. Open a card for its full record.

    Specimens read and signed by Voltry with each card idle. Providers withheld. Data as of Oct 7, 2026.

    Holding a record of your own?