Software mod reportedly unlocks 64GB VRAM on Nvidia CMP 170HX mining GPU

The CMP 170HX is built on the same GA100 silicon as Nvidia’s A100 accelerator, but with memory and compute capabilities artificially limited through firmware. The GA100 die physically contains six HBM2e memory stacks, though only one or two were enabled in the 8GB and 10GB variants of the CMP 170HX. The software mod bypasses Nvidia’s Falcon security microprocessor, allowing access to the full 64GB of on-package memory. This approach mirrors historical CPU binning practices, where manufacturers disabled cores or cache to create lower-tier products from the same silicon.

The exploit is purely software-based, requiring no physical modifications to the GPU. This lowers the barrier to adoption for engineers seeking high-memory accelerators at a fraction of the cost of an A100. However, the unlocked memory’s stability and functionality are not guaranteed. Nvidia disabled the additional memory stacks either for product segmentation or due to defects in the silicon. Users must accept the risk that unlocked memory may fail or cause system instability, particularly under sustained workloads.

Beyond memory, the mod restores compute performance by re-enabling disabled Streaming Multiprocessors (SMs) and upgrades the PCIe interface from 1.0 x4 to 2.0 x4. These changes improve the CMP 170HX’s suitability for AI and machine learning tasks, though it remains limited by its original hardware design. Full PCIe x16 bandwidth would require soldering missing capacitors to the PCB, a modification not yet confirmed to work. Even with these improvements, the CMP 170HX is based on Nvidia’s two-generation-old Ampere architecture, which lags behind Hopper and Blackwell in raw performance and efficiency.

The secondary market for the CMP 170HX has reacted sharply to the exploit, with prices quadrupling from $250 to over $1,000. This surge reflects demand for high-memory GPUs in AI applications, where VRAM capacity often dictates the size of models that can be trained or inferenced. The CMP 170HX’s appeal lies in its potential to match the A100’s memory capacity at a lower cost, but its viability depends on the silicon lottery. Engineers must weigh the cost savings against the risks of instability and the lack of official support or warranties.