How a Voltage Glitch Turns Dead NVIDIA Mining Cards into 80GB A100 Beasts

Deep inside the NVIDIA CMP 170HX mining card beats the exact same GA100 silicon die found in the elite A100 accelerator, complete with around 80GB of HBM memory physically soldered to the board while software is forced to see a crippled 8GB. These artificial limitations are enforced by microscopic eFuse hardware on the chip itself. By introducing a precise voltage fault injection attack using a custom STM32G4 controller board attached to the MP2888B voltage regulator, it becomes possible to bypass security checks and rewrite arbitrary fuses on the fly.

This hardware override disables the factory floorsweeping masks that deliberately crippled functional components during manufacturing. As a result, the unlocked card unleashes up to 128 streaming multiprocessors instead of the stock 70, running about 1.2 times faster than a retail A100 in specific compute workloads. The modification also brings back disabled features including PCIe Gen4 speeds, ECC memory, Multi-Instance GPU capabilities, and even NVLink with a minor 3.3V trace bridge workaround on the board layout.

While corporate product segmentation relies on the comforting assumption that burnt silicon cannot be unburnt, modern hardware hacking proves that silicon often just needs a very aggressive persuasion technique. NVIDIA built these chips to survive the crypto crash, but nobody planned for engineers treating enterprise accelerators like an open-source Lego set.

Source: sourcecraft.dev