Hardware

NVIDIA CMP 170HX can be hacked into an 80 GB A100 using voltage glitching

By One Percent LowsCovered by 4 outlets
★ Add us as a preferred source on Google

TL;DR

Researcher Kate Land used a $7 custom board to bypass eFuses on NVIDIA CMP 170HX mining cards, unlocking 80 GB of VRAM and up to 126 SMs.

A researcher named Kate Land has demonstrated a way to turn NVIDIA CMP 170HX mining cards into 80 GB A100-class accelerators by exploiting a flaw in the GA100 die. By using a voltage glitching technique, Land was able to rewrite the eFuses NVIDIA uses for product segmentation and yield control.

The 8GB version of the CMP 170HX physically contains five HBM2e memory stacks totaling 80 GB of capacity, though NVIDIA restricted it to 8GB and 70 streaming multiprocessors (SMs). While software tools could previously unlock up to 64 GB, Land's hardware hack unlocked the full 80 GB on nine out of ten 8GB cards she tested. One card only provided 64 GB of healthy memory. The hack also enabled 120 to 126 SMs. In FP16 compute benchmarks, the modified card reached 85.9 TFLOPS, which is about 18% faster than a stock A100 at 72.5 TFLOPS.

Performing the hack is not simple. It requires soldering, removing a single 0402 resistor, and using a custom STM32G4-based board to feed specific voltage waveforms to the regulator. This process can take anywhere from one to 48 hours per card. The changes to the fuses are permanent.

Additional hardware modifications can enable more features. A decoupling capacitor mod can enable PCIe Gen4 x16, and bridging a missing IC footprint can enable NVLink. However, enabling NVLink may tie together the 3.3V rails of cards sitting on different power supplies. Land noted that ECC, MIG, and video decoders may also unlock, but she has not tested them.

More on NVIDIA