Voltage Glitch Unlocks Full Memory on Nvidia CMP 170HX Mining Card
A hardware researcher has demonstrated that a voltage fault injection attack can bypass factory-imposed restrictions on Nvidia’s CMP 170HX mining card, effectively unlocking its full 80GB of high-bandwidth memory and additional compute resources. The exploit reveals that the chip’s secure fuse system contains a fundamental design flaw allowing users to relax hardware limitations rather than just enforce them.
A new hardware security analysis indicates that Nvidia’s silicon, long marketed for its ability to restrict functionality through immutable fuses, contains a critical design flaw that allows attackers to reverse those restrictions. The research, published by a researcher known as Kate on her personal blog, demonstrates that the CMP 170HX, a mining accelerator introduced in 2021, can be modified to behave like an Nvidia A100 data center GPU by exploiting a voltage glitch during the boot process.
The CMP 170HX was built using the GA100 die, the same silicon architecture that powers Nvidia’s high-end A100 accelerators. However, Nvidia intentionally disabled significant portions of this hardware at the factory to differentiate the mining card from its professional computing counterparts. The card shipped with only 8GB or 10GB of High Bandwidth Memory (HBM) exposed to software, 70 active streaming multiprocessors (SMs), and a limited PCIe Gen1 interface. Despite these restrictions, the physical chip contains closer to 80GB of HBM and up to 128 SMs, with the remaining capacity hidden by firmware and locked fuses.
Kate’s method involves voltage fault injection, a technique used in hardware security to disrupt a chip’s operation at precise moments. By tapping the signal line between the GPU and its voltage regulator with a microcontroller board, she was able to disturb the power supply during startup. This disturbance nudged the fuse controller into a state where normally unwritable fuses could be burned. Contrary to the standard behavior of secure fuse schemes, which are designed to permanently restrict functionality, the GA100’s architecture allowed Kate to "relax restrictions by blowing more fuses." This effectively re-enabled the memory and compute units that Nvidia had switched off.
The implications of this exploit are significant given the current state of the hardware market. The industry is currently experiencing a severe shortage of HBM, making high-bandwidth memory effectively sold out. Consequently, hardware that conceals usable memory behind software locks has become highly desirable. While community projects like cmpunlocker had previously used patched drivers to expose 64GB of memory on the CMP 170HX, Kate’s hardware-level approach goes further by unlocking the underlying silicon capabilities that software mods cannot access.
In testing, an unlocked CMP 170HX achieved approximately 85.9 TFLOP/s of FP16 fused-multiply-add throughput, compared to roughly 72.5 TFLOP/s for a retail A100. This performance gain stems from the fact that production A100s typically enable only 108 of the die’s 128 SMs for yield reasons, whereas Kate’s unlocked cards reported between 120 and 126 active SMs. However, these figures represent non-tensor FP16 throughput from a single OpenCL benchmark and do not account for tensor-core performance, Error Correction Code (ECC) functionality, NVLink connectivity, or Multi-Instance GPU (MIG) partitioning, which are critical for data center workloads.
Kate noted that ECC, MIG, and video engines remained untested in her initial findings. Furthermore, NVLink is not fully functional without additional board surgery, and pushing the PCIe link to Gen4 requires further hardware modifications. The researcher emphasized that while the card can reach higher raw throughput in specific synthetic tests, it does not function as a drop-in replacement for an A100 in a production environment. She described the modified card as a "GA100 running close to its native potential" rather than a cloned accelerator.
The vulnerability highlights a broader issue with Nvidia’s Ampere-era security architecture. Kate’s work builds on previous research, including a June 2026 paper by Arizona State researcher Jon Pry titled "A Canary in the Crypto Mine," which defeated stack protection in the GPU’s security coprocessor. Academic teams have also spent recent years demonstrating Rowhammer bit flips on Nvidia GPUs and defeating recommended ECC fixes. The common thread across these findings is that the silicon protections often prove less robust than marketing materials suggest.
Kate disclosed the vulnerability to Nvidia’s product security team before publishing her findings. She went public only after another researcher independently discovered the same issue. Nvidia has not issued a public comment on the matter. Because the flaw resides in the physical silicon of chips that have already been manufactured and shipped, there is little Nvidia can do to mitigate the issue for cards already in the field.
For users considering this exploit as a cost-effective route to AI computing power, significant practical caveats remain. Unlocking a single card requires between one and 48 hours of physical rework and per-card tuning. The fuse burns are permanent, meaning any error cannot be reversed. Of the ten cards Kate examined, one had only 64GB of healthy memory rather than the full 80GB, illustrating the variability in yield.
Ken Ashe, an analyst who commented on the work, described the process as "silicon archaeology with a soldering iron." He noted that while one successful hack is impressive, creating ten repeatable cards would constitute infrastructure. The economics of the exploit are also complicated by market volatility. Before software unlocks became available, CMP 170HX cards traded for as little as $250. After word spread about their potential, prices surged past $1,000. In contrast, retail A100s cost near $20,000, but those prices are based on stable supply chains rather than the volatile secondhand market where mining cards are traded.
The research serves as a warning regarding hardware security in an era of component scarcity. It suggests that the locks keeping hidden resources concealed are not as solid as vendors imply, particularly for last-generation hardware that repurposed higher-end die and memory configurations into lower-end packages. The exploit demonstrates that hardware people already own may conceal far more than their vendor cares to admit.
A.Rastelli--INP