NVIDIA B200

Error budget used

Floating-point answers are never exact, so each check allows a rounding error and proves how large it can honestly be. The red line is that proven limit. A bar shows the largest error this chip produced, as a share of the limit, across every run. Anything past the line would be a wrong answer.

Checked exactly, where any difference at all fails: FP8 matrix multiply, FP8 fast-accumulation matrix multiply, INT8 matrix multiply, FP8 data patterns, FP8 fast-accumulation data patterns, INT8 data patterns. Every one matched.

Bars close to the line at FP16 and BF16 are expected: there the limit is dominated by the single rounding of the output, so a correct chip has almost nothing to spare. A bar past the line is a failed run.

Every planted fault was caught

A clean result is only worth something if the tester would have seen a fault. So next to every clean run, the self-test plants faults on purpose, including the patterns NVIDIA measured at gate level, and every one must be caught. How the self-test works.

Compute probe

Matrix multiplies at several shapes and precisions. Every output element is checked against an exact or float64 reference under a proven rounding-error bound, and every run is compared bit for bit with the first run.

Tool v0.3.4, PyTorch 2.8.0+cu128, 19 October 2026. Result file · Self-test file

Error counters: error counters quiet.

Transformer kernel probe

Softmax, layer norm, GELU, rotary embeddings and fused attention, each checked against a float64 reference under its own proven bound and against the first run bit for bit.

Tool v0.2.3, PyTorch 2.8.0+cu128, 19 October 2026. Result file · Self-test file

Error counters: error counters quiet.

Checksum probe

A matrix multiply protected by row and column checksums (Huang and Abraham) with proven thresholds, so a wrong element is located to its exact row and column.

Tool v0.1.0, PyTorch 2.8.0+cu128, 19 October 2026. Result file · Self-test file

Error counters: error counters quiet.

Data pattern probe

Inputs built to reach particular bits (wide exponents, long mantissas, cancellation, alternating signs, sparse inputs, subnormals and values near the top of the range), fed through the same matrix multiply checks.

Tool v0.1.4, PyTorch 2.8.0+cu128, 19 October 2026. Result file · Self-test file

Error counters: error counters quiet.

Memory probe

Most of the device memory written with six patterns, left to dwell, and read back word by word.

Tool v0.1.1, 19 October 2026. Result file · Self-test file

Error counters: error counters quiet.

Hours under load probe

One fixed matrix multiply run back to back for as long as asked, one precision at a time, every result compared bit for bit with a first result that was checked against the proven bound. Each one-minute window records the chip's temperature, clocks, power and throttling reasons beside its verdict.

Tool v0.1.0, PyTorch 2.8.0+cu128, 19 October 2026. Result file

Error counters: error counters quiet.

Setup and files

The chip

The software

The test

NVIDIA driver 580.126.09, read with nvidia-smi on the pod after the run; the files were written by v0.4.2, which did not record the driver (v0.4.3 and later do). Power limit 1000 W, 183,359 MiB total memory, as nvidia-smi reported them.

Result files

• b200_abft.jsonl result
• b200_abft_inject.jsonl self-test
• b200_clean.jsonl result
• b200_inject.jsonl self-test
• b200_kernels.jsonl result
• b200_kernels_inject.jsonl self-test
• b200_memcheck.jsonl result
• b200_memcheck_inject.jsonl self-test
• b200_patterns.jsonl result
• b200_patterns_inject.jsonl self-test
• b200_soak.jsonl result

Run it on your chip

The same tests on your own hardware, in one command. Every result file it writes can be published here as a new row.

pip install ai-chip-integrity
chip-integrity run --name nvidia-b200

Cite the tool as: Yasha Khandelwal, ai-chip-integrity, version 0.4.3, 2026, https://doi.org/10.5281/zenodo.23393790

Other chips measured