GPU memory can be physically installed in a package while applications see only a restricted portion. Researcher Kate's November 10 account of Nvidia's CMP 170HX mining accelerator makes that distinction concrete: the eight-gigabyte variant reportedly contains about eighty gigabytes of HBM. The useful developer question is how much of that memory becomes reliably usable on a particular card.

TL;DR

  • Kate reports changing the effective configuration of owned CMP 170HX cards, exposing hardware already installed in the package.
  • The eight-gigabyte and ten-gigabyte variants have different physical memory packages. Their stock labels do not tell the whole hardware story.
  • Earlier software work already exposed sixty-four gigabytes on the eight-gigabyte variant. Kate's hardware result can reach eighty when the memory is intact.
  • The reported performance advantage measures FP16 half2 FMA arithmetic in OpenCL. It supplies no controlled inference-speed comparison.
  • One card with only sixty-four gigabytes intact gained no framebuffer advantage from the hardware unlock. Physical defects remained.

The primary article is an account of the author's research and measurements. We did not modify or independently test a GPU. This explanation stays conceptual and excludes the operational electrical and board procedures in the source.

GPU memory has several different meanings

An application's available memory, the amount physically installed and the amount that can be trusted under a workload answer different questions. A specification label usually describes the product's exposed configuration. A package inspection concerns physical components. A successful workload concerns the combination of hardware, configuration and software actually used.

Kate describes the CMP 170HX as an Ethereum mining card built around Nvidia's GA100 GPU die, the same family associated with the A100 accelerator. High bandwidth memory, or HBM, is co-packaged beside that die. Much of the expensive hardware is therefore already present before the product exposes it to an application.

Nvidia's Ampere architecture description independently distinguishes a full GA100 implementation from the configuration sold as an A100. It describes a full design with 128 streaming multiprocessors and an A100 implementation with 108. That is architectural context for related hardware becoming different products. The older description discusses the forty-gigabyte generation and cannot establish the reliability of Kate's particular eighty-gigabyte mining card.

Keeping these meanings separate prevents a very ordinary chart-reading mistake. An inventory screen showing a larger capacity establishes what the driver reports. It leaves the actual memory behaviour and workload stability to be checked. A model fitting in memory also leaves its useful inference speed to be measured.

CMP 170HX variants and manufacturing defects

Kate identifies an eight-gigabyte variant with about eighty gigabytes physically installed, and a ten-gigabyte variant with about forty. The packages differ. A larger stock label can consequently describe a card containing less physical memory underneath. Shopping by model name has acquired a side quest.

The explanation includes both product segmentation and manufacturing yield. Manufacturing can leave portions of a die or its associated memory defective. A product configuration can also disable resources that remain functional. Kate distinguishes the settings disabling resources from the records identifying defective hardware.

That distinction changes how an unlock should be interpreted. Exposing an intact disabled resource may make it useful. Exposing damaged hardware does not repair the damage. The physical package stays the same throughout the configuration change.

The Hacker News discussion contains competing theories about why particular resources were disabled, along with questions about factory testing and reliability. Those comments provide useful questions. They supply no independent reproduction of Kate's reported card measurements, and they cannot establish Nvidia's intent for every individual sample.

The source's observation that many of the author's eight-gigabyte cards had intact eighty-gigabyte memory is explicitly attached to a limited sample. It should stay attached. Turning that observation into a population success rate would add a claim that the research account does not establish.

One-way fuses and effective configuration records

Electronic fuses physically record settings. A blown fuse stays blown. Kate's account describes an interpretation layer in which a sequence of records determines effective configuration values. A later record can replace the effective value derived from an earlier record while the physical history remains.

An original document and a later amendment make a useful analogy. The original ink stays on the page. The amendment can change the instruction that applies. Physical permanence and the interpretation of recorded information concern different parts of the system.

The source-linked Nvidia patent provides conceptual context for a fuseless-fuse structure and overrideable configuration records. A patent is a design description. It cannot independently audit Kate's GA100 implementation or qualify a modified card for a workload.

Kate reports a brief electrical disturbance disrupting controller checks and permitting permanent configuration changes. The security consequence is that additional permanent fuse changes can relax some effective restrictions. Explaining that consequence requires no waveform, address, programming sequence or board modification recipe, so those details are excluded here.

The same permanence also creates a practical risk. An incorrect permanent change can strand the hardware. Reading a research result gives a developer a better mental model; it gives a prospective buyer no warranty or supported recovery path.

What the software unlock and author tests establish

The earlier CMPunlocker project already existed. In Kate's comparison, its software approach exposes sixty-four usable gigabytes on the eight-gigabyte card. The reported hardware unlock can expose eighty when the memory is intact. Repurposing mining hardware therefore has a longer history than this week's publication.

Kate reports synthetic tests that write pseudorandom data throughout the full framebuffer and read it back after invalidating the GPU cache. The article also reports functioning inference workloads. Such checks address actual access to the reported memory more directly than a capacity readout alone.

Their ownership remains important. These are the author's tests. The article does not supply an independently controlled application comparison, a full qualification campaign or a population-wide reliability guarantee. We did not operate the modified hardware ourselves.

The existence of a public research repository supplies provenance for some of the author's work. Its code licence does not automatically license every photograph or graphic in the article, and repository availability does not replace an independently executed test. No source code or hardware procedure was compiled or executed for this episode.

The FP16 microbenchmark is a specific operation

Kate's compute example uses 124 functional streaming multiprocessors, compared with 108 in a retail A100. Which groups function depends on the individual die. A schematic map of groups can explain that relationship, but the example does not establish the actual factory defect map of every mining card.

The reported ProjectPhysX OpenCL benchmark comparison measures FP16 half2 fused multiply-add arithmetic. The software-unlocked CMP result is 50.129 TFLOP/s, the retail A100 result is 72.488 and the hardware-unlocked CMP result is 85.861. That is approximately an eighteen-percent advantage over the retail reference for the measured operation.

Those numbers have a precise job. They describe repeated arithmetic throughput under the reported benchmark conditions. The primary article's introductory raw output uses 85.858, a small discrepancy from the later results figure. Our chart uses the latter figure and retains the operation and author labels.

A third-party benchmark program is a measurement method. Kate remains the operator reporting these results. The program's origin does not make the reported modified-card run an independent replication.

For a developer, capacity may decide whether a model fits at all. Useful speed also depends on the workload and data movement through the system. The post supplies no controlled tokens-per-second comparison. Reading the arithmetic chart as a universal AI uplift would give it a promotion it hasn't earned.

The sixty-four-gigabyte card is the useful counterexample

The author had one card with only sixty-four gigabytes intact. Kate writes that, in this case, the hardware unlock produced no framebuffer advantage. The earlier software approach already reached the intact capacity.

That detail is why I prefer this story's memory explanation to a bargain-accelerator headline. The effective restriction changed while the damaged hardware stayed damaged. Physically installed capacity and reliably usable capacity can be very different numbers on the same card.

Additional features retain their own boundaries. Kate explicitly says that image and video decoding features and multi-instance GPU mode were not specifically tested. The interconnect and faster PCI Express connection require additional changes. The board, firmware and functional hardware still decide which capabilities become usable.

Kate also reports disclosure to Nvidia's product security team before publication and credits another researcher with independently discovering the vulnerability. We verified no public Nvidia acknowledgement, approval or tested fix from this account. Independent discovery of a vulnerability is also a different claim from independent replication of a performance result.

Verdict: NEEDS REVIEW

I'd want independent workload and reliability tests 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 about the particular sample and the job it will run.

FAQ

Does an eighty-gigabyte physical package guarantee eighty usable gigabytes?

Usable capacity remains conditional on intact hardware and effective configuration. Kate's sixty-four-gigabyte-intact card gained no extra framebuffer capacity from the hardware unlock.

Did the researcher add more memory chips?

The reported memory was already co-packaged with the GPU die. The result changes access to existing hardware; physical defects remain.

Does the eighteen-percent comparison predict model inference speed?

It measures an author-reported FP16 half2 FMA OpenCL microbenchmark. The article supplies no controlled tokens-per-second comparison, so application speed needs its own evidence.

Sources and artwork credits

Cover portrait: Jensen Huang at Stanford, April 30, 2026, by Anderseidesvik, source, CC BY-SA 4.0. Normal crop, resize and typography; unflipped natural expression. Complete portrait-based A/B/C artwork, covers and adaptations are offered under CC BY-SA 4.0. Jensen identifies Nvidia as vendor; no research authorship, current reaction or endorsement is implied.

Brief silent fictional process commentary sources: Toy Story/Pixar-Disney via Imgflip and Futurama via Gifsee/Giphy, linked source, under two seconds per use. Hosting grants no blanket reuse licence; original media remain copyrighted.

This article expands on an episode of **The Daily Diff, a five-minute daily video on what shipped and what broke in tech.
Watch the episode · Subscribe on YouTube · the written diff lands in your inbox every morning at thedailydiff.dev.

Video transcript and primary sources