Google paper says AI hardware-diagnosis system cut repair times 18.3% on H100, 16.1% on TPU

GOOGL

·

Google researchers said in a new arXiv preprint that an internally deployed AI hardware-diagnosis system cut repair times in production by 18.3% on an H100 GPU platform and 16.1% on a TPU platform. The results come from the paper itself, not from a separate Google announcement or independent public verification.

The claim matters because large AI training clusters are expensive and tightly coupled: when one machine fails, synchronized workloads can slow or stop, leaving many accelerators idle. In that setting, faster and more accurate hardware diagnosis can have an outsized operational effect by reducing downtime and avoiding unnecessary part swaps.

The paper, titled “HardwareDx: A Foundational Graph Model System for Datacenter Machine Repairs,” was submitted to arXiv on Nov. 17 and is listed as arXiv:2611.23398v1. It is authored by Google researchers including Yuanlin Wen, Elan Markowitz, Bryan Perozzi, Vahab Mirrokni and Amin Vahdat, with the affiliation listed as Google in Mountain View, California. Its front matter lists NeurIPS 2026.

In the abstract, the authors describe HardwareDx as “an AI-driven hardware diagnosis system that applies a graph foundation model approach to automated datacenter maintenance.” In practice, that means the system models a machine as a graph of connected hardware components and tries to identify which part is actually faulty using signals such as machine topology, telemetry, prior repair actions and text features.

That is a practical shift from hand-maintained rules systems, which typically need separate diagnosis logic for each hardware platform. Google’s researchers said HardwareDx is meant to work across multiple machine types, including newly introduced ones. The paper says each server can contain more than 100 field-replaceable units, or FRUs, making incorrect swaps costly in technician time, spare parts and added downtime.

The authors wrote that the system has already been running internally at scale. “We have a full working implementation of HARDWAREDX running for 6 months at Google across 50+ ML clusters,” the paper’s introduction said. The paper also says the system is currently deployed for six AI accelerator platforms at Google, including a new TPU platform.

The reported A/B tests compared HardwareDx with Google’s existing rules-based diagnosis system. According to the paper, the H100 trial covered 225 repair windows, with 50% of H100 machines routed to HardwareDx and the rest handled by the baseline system. On the TPU platform, the trial covered 2,567 repair windows, with 20% of machines routed to HardwareDx.

Beyond mean time to repair, the authors said HardwareDx also reduced quality-assurance time by 28.3% on H100 and 25.1% on TPU. Automated diagnostic coverage, they wrote, improved from 83.1% to 93.9% on H100 and from 72.9% to 82.8% on the TPU platform. The paper also reported more than 95% precision at more than 95% recall in identifying faulty versus non-faulty components, with fault-detection accuracy 12.6% above the production rules baseline.

In the abstract, the authors summarize the operational result this way: “Deployed at production scale at Google, HARDWAREDX yields substantial efficiency gains across diverse compute architectures.”

For now, the paper is the only public source for those deployment and performance claims. No separate Google press release or independent public corroboration was available at the time of reporting.

Tags: #google, #ai, #datacenter, #hardware

Stocks: GOOGL