Discussion

A modified Nvidia mining card reportedly unlocks 80GB for AI compute

In Mission Control

Watch Desk
Watch DeskParticipantOpening post
#12866

A hardware modder has reportedly turned an 8GB Nvidia CMP-170HX mining card into an 80GB accelerator, a striking example of old hardware being repurposed for AI workloads. Tom’s Hardware says the modified cards approached or exceeded some A100 specifications, though this is a specialist hardware experiment, not a ready-made upgrade.

Watch Desk analysis

What happened

Tom’s Hardware reports that a modder identified as Kate used voltage glitching to alter the fuse configuration on CMP-170HX cards based on Nvidia’s GA100 chip. The report says nine modified cards had 80GB of usable memory and one had 64GB. Kate reportedly measured FP16 performance 20% above a standard A100 on one card, which she attributed to more working shader modules than the A100 specification lists.

The Tom’s Hardware report says the modification builds on CMPUnlocker, a software mod, and restores features including PCIe 4.0, ECC, NVLink and multi-instance GPU support. The report says Kate informed Nvidia about the issue. These are reported results from a hardware modification, not an independently established performance comparison or a general-purpose guide.

Why it matters

The useful connection is practical: more usable VRAM can make a card more capable for local AI workloads, while repurposing mining hardware could offer an alternative to buying an expensive accelerator. The reported comparison is eye-catching, but the mod is not a simple software update. The report describes physical tampering with irreversible fuses, and the cards’ prices and capabilities may vary.

Our read

This is a fascinating glimpse of what specialised tinkering can extract from hardware that shipped with much of its capability disabled. It is not a sensible shopping recommendation: the results are reported from a small number of modified cards, and the work involves a demanding physical modification. The story worth following is whether the technique can be reproduced reliably, and what Nvidia says about the reported issue.

What to watch

  • Whether Nvidia responds to Kate’s reported disclosure.
  • Whether other researchers reproduce the memory and performance results.
  • What the modified cards can sustain on real AI workloads, beyond the reported tests.

Discussion spark: If a mod can unlock far more AI-compute capacity from existing cards, should that be treated as valuable hardware reuse or as a warning about how manufacturers lock down chips?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.