MYBIGGAMING
LIVE
Hardware 13 November 2026 • 3 min read •

NVIDIA Mining GPU Hack: Unlocking 80 GB VRAM and 80% More Cores

A hardware hack targeting the GA100 chip allows NVIDIA CMP 170HX mining cards to bypass manufacturer locks, unlocking full 80 GB HBM2e capacity and significantly more streaming multiprocessors.

Author: Гика • PC •
NVIDIA Mining GPU Hack: Unlocking 80 GB VRAM and 80% More Cores

A hardware exploit has fundamentally changed the utility of NVIDIA's CMP 170HX mining accelerators. By targeting the eFuse controller of the GA100 GPU, a developer has successfully bypassed manufacturer restrictions to unlock the full 80 GB of HBM2e memory and up to 80% more computing cores, transforming a specialized mining tool into a potent asset for local AI workloads.

The CMP 170HX, originally released in 2021 specifically for Ethereum mining, was designed around the GA100 GPU—the same silicon used in NVIDIA's professional A100 accelerators. While the physical hardware contains five HBM2e memory stacks providing a total capacity of 80 GB, the mining-specific version was shipped with only 8 GB accessible to the user. The remaining 72 GB resides on the same chip package but remains locked away by NVIDIA's firmware and hardware configurations.

Unlocking the Full 80 GB Capacity

While previous software-based methods like the open-source 'cmpunlocker' allowed users to access up to 64 GB, a new hardware hack published by developer Kate on November 10 has pushed the limit to the chip's theoretical maximum. In a test of ten CMP 170HX units, nine successfully unlocked the full 80 GB of HBM2e memory. Only one unit was limited to 64 GB. The developer verified these results using intensive memory access tests and real-world AI applications, including vLLM and SGLang.

This massive increase in VRAM capacity gives the aging mining card a new lease on life in the era of Large Language Models (LLMs). With 80 GB of high-bandwidth memory, the card can run massive models locally that simply cannot fit into the memory of high-end consumer hardware, such as the GeForce RTX 5090, which is expected to feature 32 GB of GDDR7.

NVIDIA Mining GPU Hack: Unlocking 80 GB VRAM and 80% More Cores

Bypassing Hardware Locks via Voltage Manipulation

The method used to bypass these restrictions is far more invasive than a simple BIOS or driver update. It targets the eFuse controller of the GA100, which NVIDIA uses to permanently set hardware configurations—such as which parts of the GPU are active—during manufacturing. By using a custom add-on board equipped with an STM32 microcontroller, the developer can induce precise voltage drops, temporarily putting the controller into a state where these 'permanent' settings can be modified.

This process is highly technical and carries significant risks. It requires soldering and can take anywhere from one to 48 hours of repeated attempts to succeed. Because the changes to the eFuses are permanent, any error during the process can render the graphics card completely unusable.

NVIDIA Mining GPU Hack: Unlocking 80 GB VRAM and 80% More Cores

Massive Gains in Compute Performance

The impact on raw compute performance is even more dramatic than the memory expansion. While the standard CMP 170HX ships with only 70 active Streaming Multiprocessors (SMs), the hack allowed the developer to activate between 120 and 126 SMs—an increase of up to 80% in available computing units. In an FP16 compute test using a specialized OpenCL benchmark, the modified card achieved 85.9 TFLOPS, surpassing the standard NVIDIA A100 (72.5 TFLOPS) by approximately 18%. This represents a 71% performance jump compared to the previous software-only unlocked version (50.1 TFLOPS).

Beyond memory and cores, the hack also enables other previously disabled features, including PCIe 4.0 with 16 lanes and NVLink support. While the developer has noted that functions like ECC (Error Correction Code), Multi-Instance GPU (MIG), and video decoders are theoretically unlockable, they have not yet been practically verified in a real-world environment.

Article author

Гика

Cookies

We use cookies

The site uses technically necessary cookies for sign-in, security, storing preferences, and correct page operation. Analytics and other non-essential technologies should only be enabled after your consent. You can read more about data categories, purposes, and retention in the privacy policy.

Cookie policy
Quick actions