Abstract:
A few days ago, foreign researcher Kate used voltage fault injection technology to successfully unlock the originally disabled video memory and computing unit of the NVIDIA CMP 170HX mining card, transforming the mining card with a nominal 8GB of video memory into an A100-class computing card with 80GB of video memory.
CMP 170HX is a mining product launched by NVIDIA. It uses the same GA100 GPU as the data center A100, but a large number of hardware functions are disabled at the factory.
This time it is not welding a new video memory chip, but re-enabling the HBM video memory that originally existed inside the mining card but was blocked. Kate said that the 8GB version actually packages about 80GB HBM, and the other 10GB version packages about 40GB.
There have been software tools that can increase the available video memory of some CMP 170HX to 64GB, but it still cannot completely remove restrictions on computing units, PCIe interfaces, etc.
Kate this time targeted the electronic fuse controller inside the GPU that is responsible for configuring hardware functions, bypassing write restrictions through voltage fault injection, and then modifying the relevant configurations to allow the originally blocked video memory and stream processor to work again.
Of the 10 8GB mining cards tested, 9 successfully enabled full 80GB of video memory, while the other could only use 64GB. Some modified cards also restored functions such as PCIe 4.0 and NVLink, but not all functions have been fully verified.
In terms of performance, a successfully unlocked mining card OpenCL FP16 achieved approximately 85.9 TFLOPS in the FMA test, which was approximately 18% higher than the A100 measured by the researcher, which was approximately 72.5 TFLOPS. Kate thinks this is related to the fact that the modified card can enable about 120 to 126 groups of SMs, while the standard A100 only enables 108 groups.
The second-hand market price of CMP 170HX is around several thousand US dollars, of which the 8GB version is currently about US$3,000 (Quick Technology Note: about 20,000 yuan), while some A100 80GB PCIe versions can be quoted for more than 20,000 US dollars (Quick Technology Note: about 133,350 yuan).
The researchers stated that they had disclosed relevant issues to the NVIDIA product security team before publishing the results. This modification involves hardware operations and irreversible configuration modifications, which carries the risk of damaging the graphics card and is not suitable for ordinary users to try.
Comments