Unlocking Dormant Power in Budget Nvidia Chips

2026-11-13 companies

Santa Clara, Thursday, 12 November 2026.
Hardware engineers unlocked dormant features in 2,500-dollar Nvidia crypto processors, achieving 80 gigabytes of memory and outperforming twenty-thousand-dollar enterprise chips in initial tests.

The Hardware Glitch That Unlocked Hidden Capacity

A hardware modder identified as Kate demonstrated a voltage-glitching technique that bypasses Nvidia’s eFuse-based segmentation on the CMP-170HX crypto-mining card, unlocking up to 80 GB of HBM2e memory from the card’s GA100 silicon [1][3][5]. The method exploits a fault-injection flaw in the fuse controller, allowing the 8 GB retail card—originally locked to 8 GB and 70 streaming multiprocessors (SMs)—to expose its full 80 GB VRAM and 120–126 active SMs [1][3]. In comparative testing, one unlocked card delivered 18.483 ≈ 18 % higher FP16 FMA throughput than a standard Nvidia A100 80 GB while retaining full memory capacity [1][3].

From Mining Hardware to AI Accelerator

The CMP-170HX was purpose-built by Nvidia (NASDAQ: NVDA) in 2021 using down-binned GA100 dies for Ethereum mining, with eFuses permanently disabling all but 8 GB of the 80 GB HBM2e on the board [3][5]. Kate’s technique—implemented with an STM32G4 microcontroller and a custom PWMVID waveform generator—cycles the host PC through ACPI S0→S3→S0 sleep states for 1–48 hours per card, injecting precise voltage dips that trick the MP2888B VRM controller into revealing latent silicon [1][3]. Out of ten 8 GB CMP-170HX units tested, nine fully unlocked to 80 GB and eleven additional features including PCIe Gen4 x16, NVLink, ECC, and multi-instance GPU (MIG) [3][5].

Performance vs. Enterprise Cost Benchmarks

Retail pricing for a genuine Nvidia A100 80 GB PCIe card currently exceeds $20,000, whereas second-hand CMP-170HX units trade between $2,500 and $3,000 depending on memory configuration [1][3][5]. After unlocking, the modified cards match or exceed A100-class performance ceilings: independent OpenCL benchmarks recorded 18.449 ≈ 18 % higher FP16 FMA throughput on the glitched CMP-170HX, coupled with the full 80 GB VRAM allocation [3][5]. Industry observers note that reseller markets have already priced in the unlock potential, with premiums reflecting the latent silicon value [8][9].

Security Implications and Future Research

On 11 November 2026, Kate disclosed the vulnerability to Nvidia’s Product Security team after an independent researcher reported the same flaw, citing a security risk in the fuse-controller logic that permits arbitrary fuse manipulation rather than restriction [1][3][5]. The hardware modification does not restore video decoders or NVJPG blocks and leaves ECC, MIG, and NVLink partially untested across all silicon samples [3][5]. Kate’s next research phase targets deeper glitch states aimed at debug-level execution within the die, potentially enabling access to previously inaccessible security-critical regions [3][alert! ‘research stage not confirmed by 12 November 2026’].

Sources

AI Hardware Nvidia GPU