<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom"><title>Chip Integrity</title><subtitle>Findings and releases from an open tester for silent computation errors in AI chips</subtitle><link href="https://chipintegrity.org/"/><link rel="self" href="https://chipintegrity.org/feed.xml"/><id>https://chipintegrity.org/</id><updated>2026-10-15T00:00:00Z</updated><author><name>Yasha Khandelwal</name></author><entry><title>NVIDIA's own fault patterns, caught on a real GPU</title><link href="https://chipintegrity.org/findings/nvidia-fault-patterns.html"/><id>https://chipintegrity.org/findings/nvidia-fault-patterns.html</id><updated>2026-10-15T00:00:00Z</updated><summary>The self-test now plants faults shaped like NVIDIA's gate-level measurements. On an RTX 4090 the reference check alone flagged all 294.</summary></entry><entry><title>Release 0.4.2: Self-test faults shaped like NVIDIA's gate-level measurements</title><link href="https://github.com/tech4biz-yasha/ai-chip-integrity/releases/tag/v0.4.2"/><id>https://github.com/tech4biz-yasha/ai-chip-integrity/releases/tag/v0.4.2</id><updated>2026-10-15T00:00:00Z</updated><summary>Self-test faults shaped like NVIDIA's gate-level measurements</summary></entry><entry><title>When ECC is off, the memory error counters cannot see anything</title><link href="https://chipintegrity.org/findings/ecc-off.html"/><id>https://chipintegrity.org/findings/ecc-off.html</id><updated>2026-10-11T00:00:00Z</updated><summary>Two rented cards ran with ECC switched off, so their counters could not move. The tester now says so instead of calling them quiet.</summary></entry><entry><title>Release 0.4.1: The ecc-off verdict, faster exact references, and the RTX 4090 and RTX A5000 rows</title><link href="https://github.com/tech4biz-yasha/ai-chip-integrity/releases/tag/v0.4.1"/><id>https://github.com/tech4biz-yasha/ai-chip-integrity/releases/tag/v0.4.1</id><updated>2026-10-11T00:00:00Z</updated><summary>The ecc-off verdict, faster exact references, and the RTX 4090 and RTX A5000 rows</summary></entry><entry><title>BF16 matrix multiplies on Intel AMX treat tiny inputs as zero</title><link href="https://chipintegrity.org/findings/intel-amx-bf16.html"/><id>https://chipintegrity.org/findings/intel-amx-bf16.html</id><updated>2026-10-10T00:00:00Z</updated><summary>Above 16 × 16 × 16, BF16 matrix multiplies on AMX CPUs treat subnormal inputs as zero, whatever set_flush_denormal says.</summary></entry><entry><title>Release 0.4.0: Six probes, one command, OCP Test and Validation output</title><link href="https://github.com/tech4biz-yasha/ai-chip-integrity/releases/tag/v0.4.0"/><id>https://github.com/tech4biz-yasha/ai-chip-integrity/releases/tag/v0.4.0</id><updated>2026-10-10T00:00:00Z</updated><summary>Six probes, one command, OCP Test and Validation output</summary></entry></feed>
