The author successfully unlocked PCIe Gen2 x16 support on NVIDIA CMP 90HX mining GPUs. While the hardware is based on the GA102 chip and physically supports x16, it was restricted by missing passive components and software limitations. The author manually installed missing coupling capacitors to enable all 16 lanes.
To overcome software restrictions, the author used a custom AI agent running on a local LLM server. The agent analyzed NVIDIA driver source code and internal GPU registers to develop a sequence for enabling Gen2 speeds. This process involves initializing the GPU with the standard NVIDIA 610.43.03 driver, then using a modified module to adjust privilege masks and perform a PCIe retrain.
The modification allows the CMP 90HX to achieve data transfer speeds of approximately 6.3–6.4 GB/s, effectively doubling the bandwidth compared to the Gen1 x16 configuration. This improvement significantly boosts performance in tasks involving tensor splitting, where data exchange between multiple GPUs is frequent.
The author notes that while previous projects like 'cmpunlocker' achieved Gen2 x4, this is the first publicly confirmed instance of a full Gen2 x16 implementation on the CMP 90HX. The research highlights the utility of local LLM infrastructure for autonomous engineering tasks, as the agent completed the complex analysis in about one hour.
