AMD Radeon Pro 5500M (MPS) x86_64

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.

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 2.2.2, 21 October 2026. Result file · Self-test file

Error counters: no error counters.

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 2.2.2, 21 October 2026. Result file · Self-test file

Error counters: no error counters.

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 2.2.2, 21 October 2026. Result file · Self-test file

Error counters: no error counters.

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 2.2.2, 21 October 2026. Result file · Self-test file

Error counters: no error counters.

Memory probe

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

Tool v0.1.3, 21 October 2026. Result file · Self-test file

Error counters: no error counters.

Setup and files

The chip

The software

The test

The discrete GPU of a 2019 16-inch MacBook Pro (Intel Core i9-9980HK), driven through Apple's Metal Performance Shaders by PyTorch 2.2.2, the last PyTorch built for Intel Macs; 8 GB of GPU memory as system_profiler reports it. On this backend bf16, int8 and FP8 matrix multiplies do not exist and were skipped. The results outside the bound are all repeatable, and for RoPE the probe measured the backend's own sin and cos at 1.22e-4 absolute error against the model's 9.5e-7 allowance, so there the cause is the function library; see the finding on Metal through PyTorch 2.2.2 for what was measured about the other steps.

Result files

• radeon5500m_abft.jsonl result
• radeon5500m_abft_inject.jsonl self-test
• radeon5500m_clean.jsonl result
• radeon5500m_inject.jsonl self-test
• radeon5500m_kernels.jsonl result
• radeon5500m_kernels_inject.jsonl self-test
• radeon5500m_memcheck.jsonl result
• radeon5500m_memcheck_inject.jsonl self-test
• radeon5500m_patterns.jsonl result
• radeon5500m_patterns_inject.jsonl self-test

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 amd-radeon-pro-5500m-mps-x86-64

Cite the tool as: Yasha Khandelwal, ai-chip-integrity, version 0.4.3, 2026, https://doi.org/10.5281/zenodo.23505677

Other chips measured