Hardware security researcher Kate Temkin disclosed a hardware vulnerability affecting Nvidia's flagship GA100 silicon die. By applying voltage fault injection (VFI)—commonly known as voltage glitching—to the GPU power rail during critical boot sequences, researchers can reliably bypass on-die access controls and permanently blow configuration fuses to convert stripped-down CMP 170HX mining cards into 80GB A100 compute accelerators.
During the cryptocurrency mining boom, Nvidia released the Crypto Mining Processor (CMP) series. Cards such as the CMP 170HX utilized the same monolithic GA100 silicon die found in high-end A100 datacenter accelerators. However, Nvidia artificially segmented the hardware by blowing silicon eFuses at the factory, hobbling memory to 8GB, deactivating display pipelines, locking PCIe bus speeds to Gen3 x4, and disabling NVLink clustering.
Temkin discovered an architectural design flaw in how the GA100 fuse controller evaluates access permissions. In secure hardware architectures, blowing a physical fuse should strictly enforce monotonically decreasing privileges because blown fuses cannot be restored. On the GA100 die, however, several configuration bits are inverted: blowing specific additional fuses instructs the internal security processor to relax segmentation restrictions and unlock broader subsystem access.
By inducing precise millivolt-scale power rail dips during fuse verification routines, Temkin's exploit bypasses firmware validation gates, allowing targeted fuses to be programmed. Once executed, the modified CMP 170HX card enumerates under standard Linux drivers as an enterprise-grade accelerator with 81,920 MiB (80GB) of high-bandwidth memory and active NVLink interconnects.
Remarkably, the unlocked hardware outperforms production A100 accelerators. In manufacturing the standard A100, Nvidia conservatively harvested dies by deactivating 20 of the GA100's 128 physical Streaming Multiprocessors (SMs), standardizing commercial production at 108 SMs to maximize silicon wafer yields. By blowing the fuse locks, Temkin unlocked all physically functional SMs on the die—typically between 120 and 126 SMs. Benchmarks executed on Linux driver 610.43.03 recorded 85.858 TFLOP/s in FP16 FMA throughput, representing an 18% performance advantage over standard A100 hardware.
The findings were disclosed to Nvidia's product security team prior to publication. Because eFuses represent immutable physical hardware connections on physical silicon dies, the vulnerability cannot be patched via software or microcode updates on existing GA100 hardware in the field. The research provides a remarkable second life for tens of thousands of decommissioned mining cards amid ongoing global AI compute shortages.