A used mining GPU (cryptocurrency mining card, specifically the CMP 170HX) plus a free open-source tool (cmpunlocker v0.5) outperforms Jensen Huang's $5,000 new release DGX Spark this week: 1.5TB/s memory bandwidth versus 273GB/s, at less than half the price. We believe this deserves a second look from every enterprise manager currently evaluating the cost of "going AI."

What this is

cmpunlocker is a GitHub open-source tool purpose-built to unlock features NVIDIA disabled on mining cards. v0.5 dropped this week, newly unlocking ECC (Error-Correcting Code memory) and 4 additional SMs (Streaming Multiprocessors — think more compute units). Developers hint a third-generation unlock is already on the way. The CMP 170HX being unlocked is NVIDIA's mining-dedicated card from a few years back, now flooding the secondary market post-mining-boom, currently priced around $2,000. DGX Spark, by contrast, is NVIDIA's new "desktop AI workstation" aimed at developers — the 64GB version at $5,000, and the 128GB has already climbed from $4,700 to $7,000.

Industry view

The LocalLLaMA community is celebrating: "Jensen Huang's IQ tax" and "Blessing for those who can't afford DGX Spark." But cautionary voices deserve equal airtime:

  • Mining cards are a product line NVIDIA actively discontinued — driver updates and enterprise support are absent. Unlocking ECC doesn't equal stable long-context production workloads; the hardware carries no NVIDIA backing.
  • CMP 170HX's advantage concentrates on memory bandwidth, but it lacks unified memory architecture (CPU and GPU sharing the same memory pool), making local Agent-style code with frequent data exchange less efficient than DGX Spark.
  • Deploying second-hand mining cards as production hardware saves a few thousand dollars but gambles with stability and compliance.

Impact on regular people

  • For enterprise IT: the assumption that "AI compute cost is rigid" needs reassessing — but stability and enterprise support are a separate matter. Don't shoulder production workloads with consumer-grade hardware.
  • For individual careers: the barrier to running AI locally keeps dropping; tech enthusiasts with a ~¥4,000 budget now have an entry path, though commercial use remains a stretch.
  • For consumer markets: AI services priced on "token cost" rest on hardware floor prices less rigid than advertised. The room for future price cuts may be larger than expected.