An NVIDIA CMP 170HX originally equipped with 8GB of VRAM and sold for about $2,500 has been unlocked to 80GB by community developers through voltage and fuse modifications; we believe the most significant aspect of this case is not “$2,500 for the capabilities of a $20,000 A100,” but the possibility that second-hand AI accelerators will see a wave of non-standard modifications.
What this is
This is a hardware modification to NVIDIA’s professional mining card, the CMP 170HX. The developers used specific voltage states to activate previously disabled memory and shader modules, allowing the 8GB version to access the full 80GB of VRAM. Reports have directly compared its performance with the $20,000 A100 80GB, but the test conditions, cooling, power consumption, and long-term stability still require further verification.
Industry view
Supporters argue that the many mining cards in circulation, despite their constrained utility, could provide an inexpensive resource for lowering the barriers to large-model inference and fine-tuning, particularly for workloads that require substantial VRAM but do not need top-tier general-purpose computing power. Products calling the CMP 170HX the “poor person’s A100” have even appeared on the market, indicating that modification solutions are becoming commercialized.
We do not, however, think this should be interpreted as a low-cost substitute for the A100. Fuses are safety-protection mechanisms, and forcing them to trigger may cause irreversible damage; unofficial modifications can also affect warranties, driver compatibility, and deployment reliability. Companies evaluating the hardware should focus on stress-test results, power consumption, cooling, and residual value rather than looking only at VRAM capacity and benchmark scores.
Impact on regular people
For enterprise IT: VRAM capacity, rather than per-card compute performance, is becoming an important cost factor in large-model deployment. If low-cost modification schemes pass stability validation, small and medium-sized teams may be able to run larger models on smaller budgets, but procurement should prioritize legitimate products with after-sales support.
For individual careers: The hardware barriers to local deployment and private fine-tuning may fall, giving model engineers easier access to real compute bottlenecks. Still, “it runs” and “it can be maintained” are two different things; teams will need to add monitoring, cooling, and operations expertise.
For the consumer market: Prices for second-hand professional cards and related services may rise, but modification outcomes depend heavily on individual execution. For ordinary consumers, on-demand cloud usage is generally safer than buying and modifying a mining card.