Hardware Modification Unlocks High-Performance Capabilities in CMP 170HX
A hardware modification involving voltage fault injection has reportedly enabled a significant performance increase in the NVIDIA CMP 170HX, a specialized mining card, effectively transforming it into a functional equivalent of the high-end A100-80G enterprise GPU.
According to a technical writeup by a researcher identified as Kate, the modification addresses hardware limitations imposed by factory-set fuses. While the CMP 170HX typically ships with 8GB of memory, the researcher claims the hardware contains approximately 80GB of High Bandwidth Memory (HBM) that has been intentionally disabled. While previous software-based methods had reportedly allowed users to access 64GB of memory, this new method utilizes a custom STM32 glitcher to bypass the fuse lock entirely.
Technical Benchmarks and Implementation
The modification process is described as highly complex and time-intensive, requiring board wiring and a glitching period lasting between one and 48 hours per card. The researcher noted that the process is not universally successful; in one instance involving ten cards, only one successfully unlocked the full 80GB of HBM.
When successful, the modified hardware demonstrates significant performance gains over retail A100 units. The researcher reported achieving 85.9 TFLOP/s in FP16 performance, compared to 72.5 TFLOP/s on a standard retail A100. Additionally, the modification reportedly enables the use of all streaming multiprocessors (SMs)—specifically SMs 120 through 126, whereas retail A100s ship with 108 enabled—and restores support for NVLink and PCIe Gen4.
Market Context and Risks
The ability to repurpose low-cost mining hardware into high-performance AI inference tools has sparked discussion regarding the economics of GPU procurement. While the CMP 170HX was previously valued at approximately $200 to $300, the high cost of retail A100 units makes the risk of “bricking” a card during the glitching process a significant deterrent for many users.
This development occurs amid broader industry discussions regarding the scalability of AI infrastructure. Some observers have noted that as data centers face power throttling and rising costs, the efficiency of alternative architectures, such as TPUs, is being closely monitored by major industry players. Furthermore, market speculation continues regarding the long-term stability of the AI sector as major entities like OpenAI and Anthropic move toward potential public offerings, which could shift the demand dynamics for NVIDIA’s hardware.
While the technical results reported on platforms like Reddit suggest a substantial leap in utility for repurposed hardware, these findings remain unverified by independent third-party testing.
