Nvidia crypto mining GPUs hacked to restore locked-away VRAM — software mod unlocks 64GB of VRAM on $250 CMP 170HX

The GA100 silicon features a total of six HBM2e memory stacks. Depending on the SKU Nvidia is aiming for, it could have five or six memory stacks enabled. The chipmaker carves the CMP 170HX from defective or binned GA100 silicon that did not meet the stringent quality requirements for the enterprise-grade A100. Nvidia officially sold the CMP 170HX in 8GB and 10GB variants. All the chipmaker did was disable the additional memory through firmware, along with other limitations, such as nerfing compute performance and PCIe interface speeds.

Nvidia's approach is somewhat reminiscent of the old AMD Phenom II and Athlon II days, where some models shipped with disabled cores that could sometimes be unlocked through BIOS tweaks. Similarly, the CMP 170HX 8GB and 10GB models can potentially be unlocked to access their full memory capacity of 64GB and 80GB, respectively, though only 40GB is confirmed to be working from user feedback. As detailed in "A Canary in the Crypto Mine: Defeating Stack Protection in a GPU Secure Coprocessor," Jon Pry's research paper provides the technical blueprint for bypassing Nvidia’s Falcon security microprocessor, which tools like CMP Unlocked are based on.

One of the most remarkable aspects of this exploit is that it is entirely software-based and requires no physical modifications to the CMP 170HX. It sounds like a joke, but you are literally downloading more memory for your graphics card. It’s important to note, however, that the amount of memory you can successfully unlock and maintain stability will vary. The issue is that there is no way to know whether the locked-away memory is fully functional because Nvidia did so for product segmentation, or whether the chipmaker deactivated it due to physical defects. It's a similar situation to another repurposed crypto device, AMD's BC-250, that we recently tested.

Beyond unlocking the hidden memory, the software exploit can also restore processing power to the CMP 170HX’s Streaming Multiprocessors (SMs), which increases its computational performance. Another notable benefit is the upgrade to a faster PCIe interface speed. The CMP 170HX can now operate at PCIe 2.0 x4 speeds, a significant upgrade from its original restriction to PCIe 1.0 x4. There is still untapped performance, though. Nvidia implemented hard restrictions on the CMP 170HX’s connectivity by limiting the accelerator to just four PCIe lanes and physically omitting 12 capacitors from the PCB. If you solder the missing capacitors to the PCB, it could unlock full PCIe x16 bandwidth and enable the accelerator's maximum throughput, though we haven't seen that in action.

The CMP 170HX launched with a hefty price tag of $4,300 at the height of the cryptocurrency mining boom. Nowadays, they used to sell for around $250 on eBay. However, when word of the exploit got out, they immediately jumped to over $1,000, a 4X increase in market value. Nvidia’s A100, offered in 40GB and 80GB PCIe variants, starts at around $3,500 and $11,500, respectively. The CMP 170HX is attractive for AI users because it costs a fraction of the A100 but can offer the same memory capacity, though the silicon lottery is not always generous. Regardless, Nvidia’s Ampere architecture is now two generations old. Even with added memory capacity, it cannot match the raw computational performance or efficiency of Hopper or Blackwell.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.