Voltage glitch unlocks 80GB on Nvidia's CMP 170HX, but 'faster than an A100' rests on a single benchmark
Nvidia built the CMP 170HX to be a worse A100 - a power glitch on the GA100 fuse controller undid much of that, exposing a design that rewards blowing more fuses.
A researcher called Kate has demonstrated a method to unlock additional memory and computational capabilities on Nvidia's GA100 die, which powers the A100 data center accelerator and the CMP 170HX mining card. By injecting a voltage fault during boot, Kate was able to rewrite locked fuses on the chip, enabling memory and compute resources that were previously disabled at the factory.
This results in an 8GB card behaving like an 80GB A100, despite the card originally being designed to hold only 8GB of HBM2e. While the unlocked card shows a 1.2x improvement in FP16 benchmark performance compared to a retail A100, this achievement is based on a single benchmark and does not fully account for tensor performance, ECC, NVLink, and MIG features.
The CMP 170HX, which was introduced in 2021 for mining purposes, physically houses more high-bandwidth memory (HBM) than its 8GB or 10GB specification indicates. Previously, a project called cmpunlocker had exposed 64GB on the 8GB card and restored compute capabilities, but Kate's method goes beyond this by fully unlocking most of the card's potential.
Kate claims that Nvidia's fuse scheme should only restrict a chip further once a fuse is blown, but in Ampere-era chips like the GA100, this does not always hold true. The researcher disclosed the vulnerability to Nvidia's product security team before publishing her findings, and another researcher independently confirmed the same issue.
Nvidia has not yet commented publicly on the matter. This vulnerability adds to a growing body of research on Ampere-era security flaws, including rowhammer attacks and ECC bypasses.
Written by urgent.news from TechRadar's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.