Hacker unlocks full A100 capabilities in $2,500 mining GPU through voltage glitching

For those assembling homegrown AI infrastructure, the secondary market for retired data center processors has become increasingly attractive. Used Nvidia GPUs designed for mining operations now represent a compelling alternative, particularly with the emergence of CMPunlocker, a software modification that expands the capabilities of CMP-170HX cards beyond their factory specifications. One researcher has now pushed this concept to its limits, employing voltage glitching techniques to circumvent Nvidia's security mechanisms and transform a $2,500 8GB mining card into hardware that matches or exceeds the performance of a $20,000 A100 GPU with 80GB of VRAM.

The modification was accomplished by Kate, who built upon research describing how to exploit the microcode of GA100 processors. Her focus centered on the card's fuse controller—a component that has long been standard in CPUs and GPUs but is rarely discussed publicly. During manufacturing, fuses serve a specific purpose: they are permanently disabled to deactivate portions of silicon that failed testing, allowing manufacturers to sell partially functional chips as lower-tier products.

The mechanism behind recovering gigabytes of hidden VRAM is straightforward in principle. The GA100 silicon contains all the necessary HBM2e memory modules and shader processing units from the outset; manufacturers simply disable them at the factory. The 8GB mining variant derives from a GA100 design that originally contained 80GB of memory, while the 10GB versions stem from the 40GB configuration.

Nvidia's fuse architecture on these cards is remarkably complex, featuring 16,384 individual fuses arranged in a 512 by 32 matrix. Each major subsystem—VRAM banks, shader modules, PCIe 4.0 connectivity, NVLink, and video encoding hardware—has corresponding fuses that determine whether those features are accessible. As Kate observed, the design contains a critical flaw: the fuse system was engineered to disable features, not enable them, yet the security model fails to account for attackers who can manipulate electrical conditions.

Under standard operation, users cannot modify fuses. Kate circumvented this limitation by enlisting GPT-6 Astra to design a circuit board built around an $8 STM32G4 microcontroller. This device tapped into the power delivery signals on the card's circuit board, intercepting the PWM signals that communicate with the voltage regulator modules. By forcing momentary voltage glitches along the power line supplying the fuse controller, Kate induced a state where the controller's security flags permitted write operations.

The process required executing multiple fuse-status reads in rapid succession while voltage glitches were active, eventually achieving a configuration where the controller accepted modifications. Once that threshold was crossed, reconfiguring the card became feasible. Kate exercised caution by ensuring that any shader modules previously marked as defective remained disabled, since fuses represent permanent physical changes with no reversal mechanism.

Testing across multiple modified cards revealed impressive results. Nine of the units achieved full 80GB VRAM capacity, while one card stabilized at 64GB. Most notably, unlocking all available shader modules delivered FP16 FMA performance approximately 20% higher than a standard A100 card. Kate attributes this to the GA100 die containing 120 or more functional shader modules in many cases, exceeding the 108 listed on Nvidia's official specifications. The GA100 design includes 128 shader modules initially to account for manufacturing defects.

The CMPunlocker software modification alone provides only partial unlocking. It caps VRAM at 40-64GB depending on the variant, leaves numerous shader modules inaccessible, and restricts PCIe bandwidth to version 2.0. The modification also lacks ECC memory protection, NVLink support, and multi-instance GPU capabilities—all features that Nvidia reserves for server-class products and uses as product segmentation tools.

Kate's voltage-glitching approach eliminates these remaining constraints. Her best-performing modified card operates at full PCIe 4.0 speeds with the complete feature set enabled. Video encoding capabilities were not tested, though she notes this omission is unlikely to concern users of these cards.

The economics of the modification are compelling. CMP-170HX 8GB cards currently trade around $3,000, with 10GB variants at approximately $2,500. A genuine Nvidia A100 80GB in PCIe form factor commands $20,000 or more, though SXM server variants cost less but demand supporting infrastructure. These prices may shift substantially as awareness of the modification techniques spreads.

Kate decided to publish her findings because another researcher independently discovered the same vulnerability. She notified Nvidia of the security issue, though the company faces limited options for addressing the problem on cards already deployed in the field.