The mod described on Reddit starts with an Nvidia CMP170hx, a mining-focused card with 8GB of memory, and turns it, at least on paper, into a GPU comparable to an 80GB A100. The trick is not software: the author uses voltage glitching to intervene on the fuses Nvidia uses to disable VRAM and shader modules. The source reports the card reaches 120 SMs.

This story is more than a modding exercise. Physical fuses are not a minor detail: they are how manufacturers segment the same silicon into different products. Part of a die can be switched off to create a mining card or to isolate manufacturing defects. When an external intervention manages to reactivate those blocks, the latent value of the silicon emerges in a blunt way. The gap between the $2,500 CMP170hx and the $20,000 indicated for an 80GB A100 does not depend only on memory or cores: it also depends on market positioning, testing, warranty, and the enterprise channel.

For teams running on-prem inference workloads and looking at cost per gigabyte of VRAM, an operation like this can look like an opportunity. But there are three layers of consequences. First, reliability: a GPU modified outside its official specification has no guaranteed binning, can degrade, lose memory channels, or fail in production. Second, compliance and security: in environments that must meet data sovereignty or audit requirements, unsupported hardware introduces maintenance and liability risk. Third, the market: if these procedures become repeatable, the price of used cards and trust in the secondary channel change, and vendors may move segmentation toward deeper controls.

It is no coincidence that this arrives while interest in self-hosted and local AI is growing. But choosing a documented, supported configuration for inference is one thing; building a pipeline on a component that can lose unlocked features after a firmware update or an electrical fault is another. Voltage glitching, by its nature, operates at the edge of electrical specifications: it does not produce a new product, it produces a prototype.

AI-RADAR readers evaluating on-premise deployment know there are trade-offs between acquisition cost, TCO, and warranty: the analytical frameworks at /llm-onpremise help lay out these constraints without shortcuts. The source does not report stability tests or benchmarks for inference workloads. Until those appear, the distance between an unlocked card and a certified A100 remains wide.