Techie Uses Voltage Glitching to Turn $2,500 Mining GPU Into $20,000 Nvidia A100

On November 12, 2026, techie Kate used a voltage glitching hack to transform a $2,500 Nvidia 8GB cryptomining GPU into a full-fat $20,000 A100 with 80GB of VRAM, according to tomshardware.com.

The Tech TL;DR:

• An enthusiast modded an 8GB Nvidia CMP-170HX mining card into an 80GB VRAM unit matching high-end enterprise hardware.
• The modification bypasses factory eFuse locks using an $8 microcontroller to inject precise voltage glitches.
• Unlocked hardware achieves FP16 FMA performance 20% beyond standard A100 specs due to fully accessible shader modules.

The modification targets the onboard fuse controller of GA100 silicon, which manufacturers normally use once at the factory to disable defective components. While chips like the GA100 contain all corresponding HBM2e VRAM dies and shader modules physically on the die, factories permanently sever these pathways so units can be sold as lower-tier products.

Most cards feature a complex eFuse controller array containing 16,384 fuses, which restricts access to VRAM sections, shader modules, PCIe bandwidth, and NVLink interfaces.

Kate designed a simple circuit board using an $8 STM32G4 microcontroller to tap into the power signals on the card’s PCB. By intercepting the PWMVID signals sent into the voltage regulators, her custom board forced voltage glitches directly onto the line feeding the fuse controller. Repeated fuse-status reads in a tight loop during these voltage fluctuations forced the controller’s security bit flags to allow writes, permitting full hardware reconfiguration.

Prior soft-mod solutions like CMPunlocker enabled basic driver recognition but left severe hardware constraints intact, including upper VRAM limits capped at 40GB to 64GB, locked shader modules, and restricted PCIe 2.0 bandwidth without ECC support. Kate’s hardware voltage modification entirely removes these product segmentation barriers. Her best testing card successfully operates at PCIe 4.0 speeds with the complete enterprise feature set active.

Testing multiple workloads across ten modified cards revealed that nine successfully exposed the full 80GB of VRAM, while one unit retained 64GB of usable memory. Unlocking all shader modules pushed FP16 FMA performance 20% past a regular A100 card. This performance delta occurs because base GA100 dies frequently manufacture 128 shader modules to account for factory defects, exceeding the 108 modules specified on standard retail spec sheets.

Acquiring an Nvidia CMP-170HX 8GB variant requires an outlay of approximately $3,000, while 10GB variants hover near $2,500. In comparison, a factory PCIe-configured Nvidia A100 with 80GB of VRAM retails for $20,000 or more, though SXM server-module versions remain cheaper while demanding dedicated infrastructure. Kate published her research after another individual independently discovered the same vulnerability, and she has informed Nvidia of the security flaw.

“Nvidia’s design is insecure, as popping fuses can grant access to more features, not fewer.”

Kate told tomshardware.com that she published her research now because another person independently found the same issue. She also said that she informed Nvidia of the vulnerability, though little can be done for existing hardware already deployed in the field.