Nvidia CMP 170HX mining cards can be unlocked for general computing and AI inference tasks

Nvidia CMP 170HX mining cards, originally designed as purpose-built hardware for cryptocurrency mining, can be unlocked to perform general computing and AI inference tasks. These cards, which share the same GA100 silicon as the A100, were originally crippled by Nvidia with limited PCIe bandwidth, restricted HBM2e memory, and disabled floating-point performance. Through the use of open-source software tools, users can bypass these limitations, allowing the cards to run large language models (LLMs) and other compute-heavy applications. While the cards are not as fast as consumer-grade GPUs like the RTX 3090 Ti, they offer significant VRAM capacity, making them useful for specific AI workloads. However, the process is not without risks; the cards are often sold as e-waste, and many units may have hardware defects or damaged memory modules from their previous life in mining operations. Furthermore, the unlock process is not permanent and must be repeated when moving the card to a different system. The modification requires technical expertise, including the potential need for physical hardware mods like soldering capacitors to restore full PCIe functionality.

The Nvidia CMP 170HX mining card is based on the same GA100 silicon as the A100 server accelerator. Nvidia intentionally restricted the CMP 170HX by limiting its PCIe bandwidth, HBM2e memory, and floating-point performance.

Open-source software tools can bypass these artificial limitations, enabling the cards to run AI inference tasks. The cards offer substantial VRAM, which is beneficial for running large language models despite slower compute performance.

Hardware modifications, such as soldering capacitors, may be required to unlock full PCIe x16 functionality. Many CMP 170HX units are sold as e-waste and may suffer from hardware defects or memory failures.