VoltageGPU hardware check: real runs of 17 October 2026
=======================================================

hwcheck.sh (MIT, https://github.com/VoltageGPU/hwcheck) compares a machine with its
listing: GPU model, GPU count, VRAM, and a CUDA matmul on every GPU. It is read-only and
sends nothing anywhere.

Rented standard machines, one per model, each checked over SSH right after launch.
Some models were only available as part of a multi-GPU node: the H100 and B200 runs
rented one GPU of an 8-GPU node, the A100 run rented a whole 8-GPU node.

  rtx-3090.txt    1x RTX 3090                          RESULT: MATCHES THE LISTING (exit 0)
  h100.txt        1x H100 (one GPU of an 8-GPU node)   RESULT: MATCHES THE LISTING (exit 0)
  b200.txt        1x B200 (one GPU of an 8-GPU node)   RESULT: MATCHES THE LISTING (exit 0)
  a100.txt        8x A100 (whole 8-GPU node)           RESULT: MATCHES THE LISTING (exit 0)
  a6000.txt       1x RTX A6000                         RESULT: MATCHES THE LISTING (exit 0)

Earlier runs (RTX 4090, RTX 5090, H200, B300, and a mismatch caught on purpose):
https://voltagegpu.com/hwcheck-proofs/2026-10-16/README.txt

Not checked yet: L40S, RTX A4000 and RTX A5000. None was available to rent on
17 October 2026.

Run it yourself on any machine:

  curl -fsSL https://voltagegpu.com/hwcheck.sh | bash -s -- --gpu "<model>" --gpus <n>

Memory is printed as the driver reports it, which can differ from the round figure on a
spec sheet.