← Latest articles

Engineer Converts $2,500 Mining GPU Into $20,000 AI Accelerator Using Voltage Glitching

A complete XFX AMD Radeon RX 570 graphics card with two red cooling fans.
Photo credits
Photo: Chaddy · CC BY-SA 4.0 · Cropped within the original photograph to 16:9, resized, and converted to WebP; no padding, extension or fabricated pixels.
Related GPU file photo: AMD Radeon RX 570 (file photo).

An engineer has bypassed physical hardware restrictions on an older Nvidia cryptomining graphics card, transforming a budget mining device into a high-powered processor capable of artificial intelligence workloads. The modification unlocks memory capacity and processing modules normally reserved for expensive data-center hardware.

Tech enthusiast Kate used a custom circuit board designed with the help of an AI tool and an $8 microcontroller to target an Nvidia CMP-170HX mining card. These enterprise-derived processors are built with large amounts of internal memory and processing cores, but manufacturers intentionally disable parts of the silicon during production to sell them at lower prices.

The modification targets internal security switches on the hardware known as fuses. Under normal conditions, these safety controls operate as read-only systems that cannot be altered by users. By tapping into the board's power signals and intercepting pulse-width modulation signals, the engineer forced the voltage regulator to produce precise voltage interruptions. This glitching technique caused the fuse controller to enter a state where its security flags permitted writes.

The process successfully activated the full amount of working memory and extra processing sections on the board. Testing across multiple workloads showed that nine modified cards operated with the full 80 gigabytes of video memory, while one card retained a usable capacity of 64 gigabytes. Furthermore, unlocking the full set of processing units delivered floating-point performance 20% higher than a standard Nvidia A100 processor, a variance attributed to factories shipping silicon with extra functional cores to account for manufacturing defects.

Used mining cards typically cost between $2,500 and $3,000, whereas factory-standard enterprise processors with comparable capabilities can reach $20,000 or more. However, the modification process requires extreme precision. Because the changes permanently alter physical connections on the circuit board, any previously defective processing sections must be manually avoided to prevent catastrophic system failures.

The researcher published her findings after another individual independently discovered the same vulnerability, and she has informed Nvidia of the issue. There are limited options for the manufacturer to address hardware that is already deployed in the field.

Summary

This hardware modification demonstrates that lower-cost cryptomining equipment can be altered to match and exceed the performance of expensive data-center processors. However, the process involves permanent physical changes to the hardware and carries significant risks of damaging the device.