Nvidia CMP 170HX memory: physical HBM, effective configuration and defects
Kate reports exposing memory already installed in Nvidia's CMP 170HX mining accelerator.
Kate reports exposing memory already installed in Nvidia's CMP 170HX mining accelerator. The eight-gigabyte variant contains about eighty gigabytes physically, while usable capacity depends on intact hardware and effective configuration. One card with only sixty-four gigabytes intact gained no framebuffer advantage over the earlier software unlock.
Read the written edition (English) ↗
What this video covers
- Physical memory, effective configured capacity and reliably usable memory are distinct.
- The eight-gigabyte and ten-gigabyte CMP 170HX variants contain different physical memory packages.
- Later configuration records can replace effective settings while physical fuse marks remain permanent.
- Kate's reported performance comparison measures FP16 half2 FMA arithmetic in OpenCL, with no controlled inference-speed comparison.
- Manufacturing defects remain after a configuration unlock; the author's sixty-four-gigabyte card gained no extra framebuffer capacity.
Questions and answers
Why can a mining GPU contain more memory than applications can use?
Installed memory and exposed capacity are controlled separately. Kate's research reports that the eight-gigabyte CMP 170HX variant contains about eighty gigabytes physically, while configuration restrictions expose a smaller capacity. The ten-gigabyte variant has a different package with about forty gigabytes physically. Usable memory still depends on intact hardware and a configuration the driver can handle.
How can permanent electronic fuses allow effective settings to change?
Later configuration records can replace the effective values interpreted from an irreversible physical history. Kate describes this distinction in the GA100 research, and Nvidia's source-linked patent provides conceptual context for overrideable fuse records. A blown physical fuse remains blown. The episode explains that interpretation without supplying electrical procedures, hardware addresses or a modification recipe.
Does the reported arithmetic benchmark establish faster AI inference?
The reported comparison measures a specific half-precision arithmetic operation. Kate's hardware-unlocked example performs about eighteen percent ahead of a retail A100 in the ProjectPhysX OpenCL FP16 half2 FMA microbenchmark. The article supplies no controlled tokens-per-second comparison. Model fit, arithmetic throughput and complete application performance therefore require separate evidence, and the measurements remain the author's results.
Why did one unlocked card gain no additional usable memory?
That particular card had only sixty-four gigabytes intact. Kate reports that its hardware unlock produced no framebuffer advantage over the earlier software unlock, which already reached that capacity. Changing a restriction left the damaged hardware damaged. The counterexample shows why physically installed memory and reliably usable capacity must be evaluated separately for an individual card.
What was The Daily Diff's verdict on relying on a modified mining accelerator?
The Daily Diff's verdict was NEEDS REVIEW. Niko would want independent workload and reliability testing before depending on a modified mining card as an accelerator. The episode credits Kate's published research while preserving the author's measurement boundaries and the defective-card counterexample. It reports no independent hardware modification, reproduction or application benchmark performed by the show.
Transcript
Where did the extra memory come from?
0:00 An eight gigabyte GPU sounds like eight gigabytes of hardware. Researcher Kate reports eighty physically present inside an Nvidia mining card. In this video, where was the memory hiding? Why can unlocking it still leave you short? It's Wednesday, November eleven, and this is The Daily Diff. And one of Kate's cards shows exactly where the free upgrade runs out. The CMP one seventy H X is Nvidia's Ethereum mining accelerator,
What is inside this mining card?
0:22 built around the same GPU die as the A one hundred. Kate reports that configuration restrictions hide memory already in the package. Usable capacity still depends on intact hardware and a configuration the driver can handle. Kate published the research on November ten. Today Hacker News readers argued about hidden hardware and manufacturing defects. Both product segmentation and manufacturing yield matter here. The comment section has already promoted itself to semiconductor quality
0:46 assurance. The expensive part is already there.
How can installed memory stay invisible?
0:48 High bandwidth memory, or HBM, sits beside the GPU die in the package. Nvidia's architecture description distinguishes the full chip from the configuration sold as an A one hundred. Related hardware can become different products. Kate identifies two versions. The eight gigabyte card has about eighty physically installed. The ten gigabyte card has about forty. Those are different packages, so the larger product label can come with less
1:10 memory underneath. Shopping by model name just acquired a side quest. Manufacturing can leave portions of a chip defective. Kate separates the configuration that disables a feature from the record identifying defective hardware. Some disabled resources work when exposed. A physically broken memory bank remains broken, however persuasive your GitHub readme becomes. Electronic fuses are tiny physical structures that permanently
Why do one-way fuses permit a new configuration?
1:31 record settings. A blown fuse stays blown. Effective settings can be encoded through configuration records. Kate describes later records replacing earlier effective values. The physical history remains, while the configuration interpreted from it changes. Imagine a document with an original instruction and a later amendment. Adding the amendment changes which instruction applies, while the original ink stays.
1:51 The installed memory is unchanged. Nobody repaired damaged HBM by convincing a fuse to grow back. Kate points out the awkward consequence. Burning more fuses can relax some restrictions. Somewhere, a perfectly reasonable design review is discovering that the word permanent needed a second column in the spreadsheet. The research reports a brief electrical disturbance disrupting the controller's checks and permitting permanent configuration changes.
2:13 We'll keep that explanation conceptual. Reproducing the hardware procedures can destroy a card. Reading this result gives you a mental model; it gives you no warranty. A software unlock called CMP unlocker already existed.
What could the earlier software unlock already do?
2:24 Kate's comparison puts the eight gigabyte version at sixty four usable gigabytes that way. The hardware unlock can reach eighty when the memory is intact. Kate reports these hardware results; repurposing mining hardware has a longer history. Kate reports checking the full memory by writing data throughout it and reading it back after clearing the GPU cache. The article also reports working inference workloads. Those are the author's tests, stronger than a capacity readout. We haven't independently tested a modified card.
2:49 The retail A one hundred has a hundred and eight streaming multiprocessors,
What does the faster-than-A100 result measure?
2:53 the groups doing the work. Kate reports enabling more functional groups. The measured card has a hundred and twenty four. Which groups work depends on the individual die. Kate's half precision arithmetic microbenchmark puts the hardware unlocked card about eighteen percent ahead of a retail A one hundred. It measures a specific repeated operation in OpenCL.
3:10 Reading that as a boost to every AI model gives the chart a promotion it hasn't earned. Hacker News readers raised factory testing and reliability. Their competing theories supply no independent reproduction. I can appreciate the engineering and still want workload measurements and long stability runs before trusting the hardware with a job I care about. Extra usable memory can decide whether a model fits on a card.
What still separates it from a supported accelerator?
3:30 Speed also depends on the workload and moving data through the system. This article supplies no controlled tokens per second comparison. Evaluate model capacity, arithmetic throughput and application performance separately. Kate says the image and video decoding features and multi instance GPU mode weren't specifically tested. The interconnect and faster PCI Express connection require additional changes. The board, firmware and functional hardware still decide what becomes usable. Kate says the findings went to Nvidia's product security team before
3:58 publication. The post credits another researcher with independently finding the vulnerability. We've verified no public Nvidia response or tested fix here. Permanent configuration changes can permanently strand the hardware. A successful unlock still leaves cooling, power and reliability to evaluate. A seller's spectacular headline is a claim about one sample. A capacity screenshot gives you a starting point for checking it.
Why did one card gain no extra memory?
4:19 Now the card that ruins the free upgrade fantasy. Kate had one with only sixty four gigabytes intact. The hardware unlock gave it no memory advantage over the software unlock. The restriction changed; the damaged hardware stayed damaged. That is the useful buying lesson. Physically installed capacity and reliably usable capacity can be very different numbers. If you'd rather read this than hear me say it, the diff lands in your inbox every morning, free at the daily diff dot dev,
4:42 link below. So, today's verdict, NEEDS REVIEW.
What's the verdict on the unlock?
4:44 I'd want independent workload and reliability testing before treating a modified mining card as a dependable accelerator. Kate's research makes the hidden hardware understandable. Purchasing a dependable result takes more evidence. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
Sources
- Using voltage glitching to unleash the A100 in my mining cardKate's blog
- NVIDIA Ampere Architecture In-DepthNVIDIA Developer
- OpenCL-BenchmarkProjectPhysX
- GA100VFI research provenanceKate
- Earlier CMPunlocker projectAmogh Munikote
- Integrated circuit having an enhanced fuseless fuse structureNVIDIA patent
- November 11 discussion of Kate's researchHacker News
- Jensen Huang at Stanford, April 30, 2026 — AnderseidesvikWikimedia Commons
- Portrait and complete artwork licence: CC BY-SA 4.0Creative Commons
- Toy Story process-commentary still sourceImgflip / Pixar-Disney media
- Futurama fictional sceptical reaction sourceGiphy / Gifsee
- Futurama GIF linked sourceGifsee
