Does this AI chip still give the right answer?
A GPU can pass every vendor diagnostic and still return wrong numbers with no error raised. Chip Integrity is an open, vendor-neutral tester that checks whether a chip computed correctly, and publishes the evidence as a file anyone can check.
Three steps behind every result
The same three steps run on every chip, so a result on one chip means the same thing as a result on any other.
Run real AI maths, many times
Matrix multiplies in every precision, the kernels inside a transformer, data patterns built to reach particular bits, and a sweep of the device memory. Every step runs dozens of times.
Check every answer two ways
Each output is compared with an exact or float64 reference under a proven rounding-error bound, and with the first run bit for bit. A wrong answer has to get past both.
Prove the checks work, then publish
Next to every clean run, a self-test plants faults on purpose and every one must be caught. Everything is written in the OCP Test and Validation format and published with the result.
Every chip measured
Each chip has its own page: how close every precision came to its proven error limit, proof that every planted fault was caught, and the raw result files.
Apple GPU (MPS) arm64
Tested 10 October 2026. 1,785 compute runs, 0 silent faults, 264 of 264 planted faults caught.
NVIDIA A100-SXM4-80GB
Tested 7 October 2026. 600 compute runs, 0 silent faults, 90 of 90 planted faults caught.
NVIDIA GeForce RTX 4090
Tested 11 October 2026. 2,515 compute runs, 0 silent faults, 351 of 351 planted faults caught.
NVIDIA H100 80GB HBM3
Tested 10 October 2026. 2,515 compute runs, 0 silent faults, 351 of 351 planted faults caught.
NVIDIA H200
Tested 8 October 2026. 600 compute runs, 0 silent faults, 90 of 90 planted faults caught.
NVIDIA RTX A5000
Tested 11 October 2026. 2,055 compute runs, 0 silent faults, 297 of 297 planted faults caught.
NVIDIA's own fault patterns, caught on a real GPU
NVIDIA's gate-level study of a data-center GPU (Tung et al., DSN 2026) found that real faults are mostly zeroed values and multi-bit flips, arriving in warp-aligned groups. Since version 0.4.2 the self-test plants faults shaped that way. On a rented RTX 4090, the reference check on its own flagged all 294 of them across six precisions, which means the same faults would be caught even if they repeated identically on every run.
Pick a starting point
New to the subject
What a silent error is, why it matters for AI, and the words used on this site.
Checking the method
Every probe, the two checks, the self-test and what a PASS can and cannot tell you.
Testing your own chip
One command on an NVIDIA GPU, an Apple GPU or a CPU, and how to publish the result.