Quick Answer: Voltage glitching, a type of fault injection attack, reportedly let a researcher blow fuses on an NVIDIA GA100 die inside a CMP 170HX mining card. Blowing those fuses relaxed restrictions instead of adding them, which turned the card into something close to an A100-80G. The technique exposes a design flaw in NVIDIA's fuse-based product segmentation, not a general vulnerability in every A100.
Most people assume a fuse is a one-way door: once it's blown, the chip gets more locked down. That assumption is the core of how NVIDIA segments its silicon, and it's the assumption a recent write-up on the GA100 die says is broken. The author describes a fault injection technique that let them blow fuses that unlock hardware NVIDIA disabled on the CMP 170HX. Here's what the research shows, why the design is the real problem, and what it means if you own or secure this hardware.
What the researcher actually claims
The author says a CMP 170HX 8GB card can be pushed into a state where it behaves like an A100-80G. They report two results: memory capacity that matches the datacenter part, and streaming multiprocessor (SM) counts above what retail A100s ship with. The post says the retail A100 enables 108 SMs as a "yield-safe" choice, while the GA100 die has 128 physical SMs. Depending on the die, the author's unlocked cards reached roughly 120 to 126.
The performance number needs context. The author's OpenCL output puts FP16 FMA throughput at 85.858 TFLOP/s for the unlocked card against 72.488 TFLOP/s for a retail A100, about 1.18x. That's a synthetic peak-throughput benchmark. It doesn't tell you how a training job, a sparse model, or a memory-bound inference server will behave. Treat it as a hardware ceiling, not a production speedup.
The author also notes that the same vulnerability was independently found by another researcher, and that the findings went to NVIDIA's product security team before publication. That disclosure sequence matters. It means this isn't a fresh zero-day being dumped on the internet, and it gives vendors time to respond.
Why a blown fuse should never unlock anything
A properly designed fuse is a one-way state change. Blowing it should only make the chip more restrictive, because a blown eFuse can't be reset in any practical way. NVIDIA uses eFuses for yield control and product segmentation, so a GA100 that tests poorly or is destined for a mining SKU gets fuses set that disable features.
The author's central point is that NVIDIA's scheme breaks that rule in one place. Some fuse-encoded values can be overridden by programming more fuses, which means blowing a fuse can grant access instead of removing it. That's counter-intuitive, and it's the reason this matters beyond one card. Security designs often rely on the assumption that programming is monotonic and that only the vendor controls it. Here, the programming path is reachable under fault injection conditions that were never supposed to succeed.
Here's the practical consequence. If an attacker can corrupt the control logic that gates fuse programming, the fuse's protection is only as strong as that gate. Hardware defenses like redundant checks and glitch detectors exist for exactly this reason.
How the GA100 fuse controller divides its fuses
The fuse controller on the GA100 manages 16,384 eFuses organized as 512 rows of 32. According to the author, those rows and the surrounding structures split into four groups with very different properties.
| Region | Rows / location | What it stores | Can a blown fuse be reversed? |
|---|---|---|---|
| Regular fuses | Rows 0–221 | Single or paired fuses mapped to option words, such as the fuse-program enable bit | Effectively no, except through the repair mechanism |
| Fuseless (FF) fuses | Rows 222–487 | Records that replay option-word values into chains, encoding most product features | Yes, by writing a newer FF record that overrides the block |
| IFF fuses | Rows 487–512 | Instructions for on-the-fly reprogramming of option words | Usually not accessible in production cards |
| RIR repair fuses | 8 rows outside the main array | Up to 16 repair instructions overriding specific fuses | Set by the TSMC-supplied fuse macro design |
The table shows where the trade-off lives. Regular fuses are the hard, permanent layer. The fuseless layer is more flexible and, in the author's account, is where most segmentation features sit. That flexibility is useful for engineering and manufacturing, but it's also the surface a fault injection attack targets.
Why fuseless records are the weak point
Fuseless records are a clever design. Each record is a 32-bit entry that selects a chain and block and supplies a new 16-bit value. When two records address the same block, the later one wins. The author compares this to a JTAG-style replay script, which is a reasonable analogy.
The problem is that "later wins" logic is only safe if the writer is trustworthy. A physical fuse can't be unblown, but a fuseless record is just data in a structure the controller reads. If the controller accepts new records under conditions it shouldn't, then the segmentation effectively becomes a suggestion. The author is careful to say that the exact mapping of chains and blocks to option words comes from differential analysis, since NVIDIA hasn't published one. That's worth remembering: the public map is inferred, not official.
Here's where most security writeups stop, and the more useful question is what defenders should do. Authenticated, signed fuse payloads and monotonic counters would make a replayed or forged record fail verification. Redundancy in the write-enable path, which the regular fuses already partly provide, helps too. Neither is free in silicon area or validation effort, which is why segmentation schemes tend to be simpler than they should be.
The performance claim versus real-world value
It's tempting to read this as "buy a mining card, get an A100 for less." That's the wrong takeaway, and the author's own caveats show why.
First, die quality varies. The SM count the author reached depends on which SMs happen to be functional on a given die, so "~120 to 126" is a lottery result, not a SKU. Second, the fuse changes only some of what matters. Memory capacity and SM count are visible, but reliability, error-handling behavior, and vendor support are not the same as a datacenter part. Third, the author has not tested every feature. The post explicitly says NVJPG, NVDEC, and MIG were not specifically tested.
There's also a warranty and compliance question that any business buyer should weigh. Modifying a card's fuses and running it in production may void support terms, and it can complicate audits in regulated environments. If you're running inference for paying customers, the safe default is hardware NVIDIA sells and supports for that purpose. An unlocked consumer card is a research object, not a procurement decision.
What this means for owners, buyers, and vendors
If you own a CMP 170HX, the honest answer is that this research doesn't make your card a safe production A100 substitute. It shows that the segmentation you paid less for was softer than it looked. If you're a buyer, the lesson is to verify what hardware your provider actually runs, since capacity and SM counts are cheap to misreport and hard to audit remotely.
For vendors, the lesson is broader. Any product that uses fuses for segmentation should assume a determined attacker can glitch the control path. That means designing the fuse gate so that a single fault can't flip a permission, and treating fuseless or replayable configuration as a security boundary, not a convenience. This next part matters more than it looks: segmentation that depends on a single check is only as strong as its weakest clock edge.
Frequently Asked Questions
What is voltage glitching on a GPU?
Voltage glitching is a fault injection technique that briefly disturbs the power supply to a chip so that a specific operation fails or behaves unexpectedly. On a GPU, the target is usually a security-relevant step such as fuse programming or a configuration check. The attack doesn't break encryption by itself. It forces the chip to take a path its designers assumed it could never reach.
How to tell if a CMP 170HX has been unlocked?
You can check the reported memory capacity and SM count using tools like nvidia-smi and the OpenCL device info, as the source article does. Compare the reported values against what the original SKU is supposed to show. Be aware that the apparent capacity and the functional behavior can differ, so benchmark your actual workload before trusting the numbers.
Why does blowing a fuse unlock features on the GA100?
Most fuses are designed to only restrict a chip. The GA100 scheme, according to the research, includes a fuseless layer where later records override earlier option values. That means programming more records can relax a restriction. The design is only safe if the controller never accepts unauthorized records, and the research suggests that assumption can fail under fault injection.
Is it safe to use an unlocked mining GPU for production AI workloads?
For most production workloads, no. Unlocked consumer cards may lack vendor support, may not be covered by warranty, and have not been validated the way datacenter parts are. The unlocked performance also varies by die. If your workload needs predictable reliability and support, buy hardware sold for that purpose.
Trust in hardware segmentation depends on the controls behind the fuses, not just the fuses themselves. If you work with GPUs in production, pull the vendor's specs for your exact SKU and check them against what your hardware reports today. Pass this to anyone who buys GPUs on a budget and assumes a cheaper card is a discount on the same silicon.