Watch the video, transcript and sources
https://www.youtube.com/watch?v=dRG6q19Wj1w
+ Kate reports exposing HBM already installed beside the GPU
- Physical defects still limit usable capacity
+ The FP16 result measures one arithmetic microbenchmarkAn eight-gigabyte GPU sounds like eight gigabytes of hardware. Researcher Kate reports about eighty physically present inside the eight-gigabyte version of Nvidia's CMP 170HX mining accelerator. The post was published on November 10; fresh Hacker News attention arrived on November 11. The hardware itself predates this research.
The useful first answer is that physical memory, effective configuration and reliably usable capacity are different things. The mining card shares the GA100 GPU family with the A100. HBM sits beside the GPU die in the package. Restrictions can hide hardware already installed, while defects can make parts of that hardware unusable.
Check which package the number describes
Kate distinguishes the eight-gigabyte variant, with about eighty gigabytes physically installed, from the ten-gigabyte variant, with about forty. Those are different packages. The larger stock product label can therefore come with less physical memory underneath. Shopping by model name has acquired a side quest.
Both product segmentation and manufacturing yield belong in the explanation. Kate separates disabled resources from defective resources. Exposing an intact disabled portion can make it useful. A damaged memory bank stays damaged.
Nvidia's architecture description independently distinguishes a full GA100 implementation from the configuration sold as an A100. That supports the architectural context. It supplies no independent qualification of Kate's particular modified cards.
Permanent marks, changeable interpretation
An electronic fuse physically records a setting. A blown fuse stays blown. Kate describes configuration records whose later entries can replace earlier effective values. Think of an original document and a later amendment: the original ink remains, while the instruction that applies changes.
The research reports a brief electrical disturbance permitting permanent configuration changes. This edition stays conceptual. We performed no hardware modification or card test, and the article's operational procedures are outside the video and companion explanation. Permanent errors can strand the hardware.
What the evidence measures
The earlier CMPunlocker software approach already exposed sixty-four usable gigabytes on the eight-gigabyte variant, according to Kate's comparison. The hardware result can reach eighty when the memory is intact. The author reports full-memory write/read tests after cache invalidation and functioning inference workloads.
Those are author-reported tests. We haven't independently reproduced the modified-card result. The third-party benchmark program is method context; its name on a chart does not turn the operator's measurement into an independent test.
Kate's FP16 half2 FMA OpenCL microbenchmark puts the hardware-unlocked example about eighteen percent ahead of a retail A100. This measures a specific repeated arithmetic operation. Model fit, arithmetic throughput and application speed still require separate evidence. The post supplies no controlled tokens-per-second comparison.
Some feature claims also retain explicit boundaries. Kate says image/video decoding and multi-instance GPU features weren't specifically tested. The interconnect and faster PCI Express connection require additional changes. A shared die does not qualify the entire board as a supported accelerator.
The card that gained no extra memory
Kate had one card with only sixty-four gigabytes intact. For that card, the hardware unlock produced no framebuffer advantage over the software unlock. The restriction changed; the damaged hardware stayed damaged.
That counterexample is the useful takeaway. Physically installed capacity and reliably usable capacity can differ dramatically. A seller's capacity screenshot provides a starting point for validation; it cannot establish long-term workload reliability by itself.
Verdict: NEEDS REVIEW
I'd want independent workload and reliability testing before depending on a modified mining card as an accelerator. Kate's research makes the hidden hardware understandable. Purchasing a dependable result takes more evidence.
Which matters more for your workload: usable memory or measured speed?
Sources
Artwork and brief reaction credits
Jensen Huang portrait at Stanford, April 30, 2026, by Anderseidesvik, source, CC BY-SA 4.0. Normal crop/resize and typography; natural unflipped expression preserved. Complete portrait-based A/B/C artwork, covers and adaptations are offered under the same licence. Jensen identifies Nvidia as vendor, with no authorship/current-reaction/endorsement claim.
Brief silent fictional process commentary, under two seconds per use: Toy Story/Pixar-Disney via Imgflip, and Futurama via Gifsee/Giphy, source. Hosting grants no blanket reuse licence; original media remain copyrighted.
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.