Kate has documented an intervention that looks less like an upgrade and more like a reverse-engineering experiment. The CMP 170HX mining card ships with 8 GB of usable HBM, but the silicon contains about 80 GB, disabled by fuses. With software-only unlocks, some people had already exposed 64 GB. Kate used voltage fault injection to bypass the fuse lock and re-enabled between 120 and 126 streaming multiprocessors, compared with the 108 that NVIDIA enables on retail A100 cards. The result is an FP16 improvement from 72.5 to 85.9 TFLOPS, with a full 80 GB available for inference. NVLink and PCIe Gen4 also come back, after being disabled in the mining version.

For teams running LLM inference on-premise, the number to watch is not just the benchmark: it is confirmation that VRAM is the most artificially constrained resource. With 80 GB of HBM, you can host larger models without aggressive quantization and benefit from bandwidth that makes token generation smoother. But that is where the deployment interest ends: the procedure is not replicable in production. It requires board wiring, a custom STM32 glitcher, and between 1 and 48 hours of attempts per unit; out of ten cards treated, one still had only 64 GB working. The risk of bricking hardware is real, and the recovered capacity is not uniform.

The structural angle matters more than the mod itself. NVIDIA segments the same commercial die with fuses, not with different silicon, to separate mining, data center, and workstation products. That makes manufacturing efficient, but it also creates an incentive to dig into the disabled part. Kate's work shows that the line between a mining card and an A100 is a decision, not a physical boundary. For teams designing on-premise compute infrastructure, the message is not to start voltage glitching: the TCO of a fleet of modified cards is incalculable because of success rates and lack of support. The message is that the market has a clear demand for abundant VRAM for local inference, and fuse policies are coming under scrutiny.

There is also the homogeneity problem: if one card in ten remains at 64 GB, any cluster built on these foundations would have variable capacity that is hard to plan around. Teams evaluating self-hosted LLM deployments do not need extreme solutions; they need cards with predictable specifications and measurable power draw. At AI-RADAR we track frameworks for comparing these architectures under /llm-onpremise. The fact remains that Kate's operation is both an electronics lesson and a symptom: beneath commercial GPU labels, there is a reserve of silicon that someone, rightly or wrongly, wants to make usable.