A CMP 170HX 8GB graphics card originally built for cryptocurrency mining has been converted into an A100 80G with 80GB of usable memory: tested FP16 performance reaches 85.9 TFLOP/s, above the retail A100’s 72.5 TFLOP/s. The result shows that large-memory AI hardware does not have to come exclusively through NVIDIA’s official supply chain, but the modification cost and risk of scrapping the card are far beyond what ordinary procurement can absorb.
What this is
Kate used voltage fault injection to bypass the fuse lockout on the CMP 170HX board, enabling previously disabled compute units and memory before restoring interfaces such as NVLink and PCIe Gen4. The goal was to put the mining card to work on large-model inference: the full 80GB of memory can reduce model splitting and cross-card communication, while a 64GB version was built for testing.
This was not a matter of simply flashing software. The conversion required modifying the board’s circuitry and building a custom STM32 fault injector, with each card requiring 1 to 48 hours of continuous work. Only one out of ten cards retained a fully intact 64GB; the others suffered memory damage. The author herself also characterized the project as a technical experiment rather than a stable product.
Industry view
We believe the real value of this modification lies in two points. First, mining cards of the same model may contain large amounts of unused high-bandwidth memory (HBM, the high-speed data link between an AI chip and its memory). Second, the inference market is highly sensitive to memory capacity, giving second-hand hardware capable of holding larger models a new use.
But the risks must carry equal weight. Repeated voltage fault injection may permanently damage the memory or the entire card, and test results do not predict the consistency, thermal performance, power consumption, or long-term stability of mass production. The more realistic constraint is that NVIDIA’s software ecosystem and support infrastructure have not disappeared. An organization willing to accept downtime and asset write-off risk for experimental hardware is not necessarily able to rebuild its compute plans around it.
In the short term, this looks more like an enthusiast-driven reuse of existing hardware. It could influence the second-hand AI accelerator market only after the modification process becomes standardized and the defective-card rate and performance variance are proven manageable.
Impact on regular people
For enterprise IT: Businesses with limited budgets and well-defined inference needs can include modified mining cards in a small-scale validation list. They should first assess cooling, power delivery, failure rates, and software compatibility, rather than deploy them as direct replacements for mainstream servers.
For individual careers: It reminds technical teams to look beyond the “hardware performance sheet” and examine the real bottlenecks. The cost of deploying large models is sometimes determined first by memory capacity and communication efficiency—not just chip compute.
For the consumer market: This graphics card will not soon become an upgrade option for ordinary players. The more likely impact is that as second-hand, high-memory compute supply increases, rental prices and hardware choices for local-model inference will be disrupted.