That exact drama is playing out right now around a five-year-old cryptocurrency castoff: Nvidia's CMP 170HX. In November 2026, a hardware researcher known as Kate published findings showing how a sub-$10 microcontroller board and a precisely timed power glitch can unlock up to 80GB of HBM2e memory and more active compute units than a retail Nvidia A100 data-center accelerator. Yet as tempting as it sounds to declare a $20,000 enterprise GPU unlocked for pennies, the reality behind silicon binning—and this specific exploit—is far more nuanced.
Silicon Binning: Salvage Yard or Price Wall?
To understand why an "8GB" mining card might be hiding 80GB of physical high-bandwidth memory, you have to understand product binning. Designing a cutting-edge processor or GPU costs billions in research and development, while etching it onto a silicon wafer is an unforgiving physics exercise. Microscopic dust particles and lithography quirks mean that not every block on a giant die comes out working cleanly.
Rather than throwing away a die with a single flawed core or memory controller, chipmakers test the silicon, fence off the broken or underperforming sections, and sell the chip as a lower-tier SKU. That keeps wafer costs down and reduces electronic waste.
The plot twists when a manufacturing process matures. As fab yields climb above expectations, factories suddenly produce far more fully functional chips than the ultra-expensive flagship market can absorb, while budget and mid-range buyers still clamor for cheaper gear. When that happens, vendors routinely take healthy silicon and disable working cores, shaders, or memory channels to meet lower-tier demand. Usually, those locks are supposed to be permanent—lasered off or burned into one-time programmable e-fuses. Sometimes, though, engineers leave the door slightly ajar.
The CMP 170HX Case: Turning a Mining Castoff Toward an A100
When Nvidia launched the passive-cooled CMP 170HX in 2021 during the Ethereum mining boom, its pitch was straightforward: keep miners off GeForce gaming shelves by building a compute-only accelerator with no display outputs. Under the hood, however, the CMP 170HX used the exact same GA100 Ampere die (fabricated on TSMC's 7nm node with 54.2 billion transistors) that powered Nvidia's flagship A100 data-center GPU.
Out of the box, the card was heavily shackled:
- Memory: Software could only see 8GB (or 10GB on a sibling variant) of HBM2e.
- Compute: Active Streaming Multiprocessors (SMs) were restricted to 70 (out of 128 physical SMs on the die), FP32 Fused Multiply-Add (FMA) throughput was throttled down to roughly 0.39 TFLOPS, and Tensor Cores were gated.
- Interconnects: The PCIe interface was capped at Gen1 x4, with NVLink, Error-Correcting Code (ECC) memory, and Multi-Instance GPU (MIG) partitioning disabled.
From Software Exploits to Voltage Fault Injection
In a market starved of high-bandwidth memory for local AI inference, enthusiasts couldn't resist prying that GA100 die open. Earlier in 2026, community tools like cmpunlocker—building on hardware-security research such as Arizona State researcher Jon Pry's June 2026 Canary in the Crypto Mine paper—exploited a flaw in the GPU's Falcon BootROM via modified open-source Nvidia drivers. That pure software approach freed up 64GB of usable HBM on 8GB cards, lifted the FP32 and tensor-core stalls (pushing tensor throughput past 160–190 TFLOPS in community testing), and bumped the bus to PCIe Gen2 x4, though it stayed locked at 70 SMs.
Kate's November 2026 breakthrough went straight for the hardware lock itself. Using a small STM32G4 microcontroller board (whose schematic she had OpenAI's GPT-6 Astra design) wired directly to the card's PWMVID and NVVDD voltage-sensing lines after removing a tiny 0-ohm resistor, she performed voltage fault injection (VFI). By briefly disturbing the GPU's power delivery during startup sleep-wake transitions, she coaxed the fuse controller into a narrow "Goldilocks" glitch state where normally protected fuse-programming overrides became writable.
Here is the real security revelation from her work: Nvidia's GA100 uses a system called Fuseless Fuses (FF), which acts almost like an append-only configuration script stored in fuse rows. In secure hardware design, blowing an irreversible fuse should only make a device more restrictive. On GA100, appending new FF rows overrides earlier blocks—meaning once VFI bypassed the write guards, blowing additional fuses actually cleared the disable masks and relaxed the chip's restrictions.
Why "Faster Than an A100" Needs a Massive Asterisk
When Kate unlocked the non-defective blocks across ten 8GB CMP 170HX cards, nine of them exposed a full 80GB of usable HBM2e, while one topped out at 64GB due to genuine hardware defects. Even more eye-catching, because Nvidia caps retail A100 cards at 108 SMs to guarantee manufacturing yields, healthy unlocked mining cards brought 120 to 126 SMs online. In a single OpenCL FP16 half2 FMA test, her hardware-unlocked card scored 85.86 TFLOP/s compared to 72.49 TFLOP/s on a retail A100—roughly a 1.18x to 1.2x edge.
Before anyone rushes to eBay thinking they've found a $20,000 AI workstation cheat code, look at what the unlock doesn't give you:
| Configuration | Usable HBM2e | Active SMs | PCIe Link | Tensor & FP16 Status | Enterprise Features (ECC / MIG / NVLink) |
|---|---|---|---|---|---|
| Stock CMP 170HX (8GB SKU) | 8 GB | 70 SMs | Gen1 x4 | Heavily throttled (0.39 TFLOPS FP32 FMA) | Disabled |
Software Unlock (cmpunlocker) | 64 GB | 70 SMs | Gen2 x4 | ~50.1 TFLOP/s OpenCL FP16 FMA; Tensor cores unthrottled | ECC unavailable; no NVLink or MIG |
| Hardware VFI Unlock (Kate) | Up to 80 GB (sample-dependent) | 120–126 SMs | Gen4 (requires cap mod) | ~85.9 TFLOP/s OpenCL FP16 FMA; Tensor untested in VFI report | ECC/MIG/video untested; NVLink needs PCB bridge mod |
| Retail Nvidia A100 80GB PCIe | 80 GB | 108 SMs | Gen4 x16 | 312 TFLOPS FP16 Tensor (624 sparse) | ECC on by default, MIG (up to 7 instances), NVLink |
TechRadar rightly emphasized that Kate's headline speed win rests on a single non-tensor OpenCL benchmark, leaving Tensor Core throughput, ECC memory protection, MIG partitioning, and media engines untested or incomplete. Running large AI models for hours without ECC increases the risk of silent bit errors, and getting NVLink working requires physically bridging pins where an IC (U11) is missing from the mining PCB—a hack that dangerously ties +3.3V power rails together if GPUs use separate power supplies. Add in that the VFI process requires soldering, die-specific fuse math, and 1 to 48 hours of trial-and-error per card (with irreversible bricking risk), and this is clearly hardware-security research, not a consumer upgrade path.
Meanwhile, the economics have already shifted. Secondhand CMP 170HX cards that once lingered around $200 to $250 surged past $1,000—and even higher on overseas listings—once the unlocks went viral. And buyers chasing the 10GB CMP 170HX variant hoping for 80GB should be wary: Kate notes those boards typically carry 40GB of physical HBM, and claims of 80GB unlocks on 10GB cards often stem from memory aliasing rather than genuine capacity.
Ghosts of the Silicon Lottery Past
If the CMP 170HX feels like an extreme datacenter echo of classic PC modding, that's because Windows PC builders have played—and occasionally won—this game for nearly two decades. Three historical milestones show both the thrill and the limits of unlocking hidden hardware:
- AMD Phenom II X3 Core Unlocking (2009): During the early Phenom II era, AMD's triple-core X3 710 and X3 720 processors were built on four-core Deneb dies with one core disabled. In February 2009, enthusiasts discovered that on compatible AM2+/AM3 motherboards (such as Biostar's 790GX boards), toggling Advanced Clock Calibration (ACC) to "Auto" in the BIOS could reawaken the dormant fourth core right inside Windows. As Tom's Hardware documented in its April 2009 investigation Phenom II: Unlocking Cores, Cache, And A Free Lunch, it really was a literal silicon lottery: some lucky buyers got a rock-solid quad-core CPU for triple-core money, while others discovered the fourth core was disabled for a reason when Prime95 crashed on contact.
- AMD Radeon HD 6950 to HD 6970 (2010): When AMD launched its Cayman-based Radeon HD 6950 and 6970 in December 2010, TechPowerUp discovered that early 2GB reference HD 6950 cards locked out their extra shaders purely in the VGA BIOS rather than via on-die fuses. Thanks to a physical dual-BIOS recovery switch on reference boards, PC gamers could safely flash an HD 6970 BIOS using ATIFlash, unlocking the full 1,536 shaders. Still, some cards couldn't sustain the 6970's higher 880 MHz clock speeds without artifacting, proving binning still mattered.
- AMD Radeon RX 480 4GB to 8GB (2016): In July 2016, rumors flew that launch-day 4GB Radeon RX 480 cards actually packed 8GB of GDDR5 memory. TechPowerUp bought an off-the-shelf Sapphire 4GB card, peeled off the cooler to verify eight 1GB Samsung memory chips on the PCB, and flashed it with an 8GB BIOS. The modified card recognized all 8GB and matched a retail 8GB model (24.7 fps vs. 24.6 fps in testing). Why did AMD do it? Early 4GB retail batches likely reused 8GB reference boards with a capacity-limiting BIOS originally prepared so reviewers could test both SKUs on one sample. Later retail waves physically omitted the extra memory chips, leaving anyone who blindly flashed an 8GB BIOS with an unbootable brick.
What This Means for PC Builders and IT Teams Today
Can you still buy a mid-range GeForce or Radeon card today and softmod it into a flagship? Almost certainly not. Chipmakers learned hard lessons from the Phenom II and Cayman eras. Modern consumer GPUs and CPUs routinely use laser-cut interconnects, cryptographically signed firmware, and hardware-enforced monotonic fuse trees precisely to prevent BIOS swaps and register hacks from turning an RTX 5080 into an RTX 5090.
Instead, the CMP 170HX saga highlights two very modern realities for Windows enthusiasts and enterprise IT administrators:
- Be skeptical of "unlocked" secondhand hardware: If you are building a budget local-AI rig and spot a modded mining card advertised as an "80GB A100 equivalent," remember what is missing. Without ECC validation, signed enterprise firmware, or guaranteed long-term memory health, a rescued mining card is a tinkerer's project—not drop-in production infrastructure.
- Hardware fuses aren't infallible: For security teams, Kate's GA100 work—alongside recent Rowhammer and BootROM research on Ampere silicon—is a reminder that physical access and voltage glitching remain potent threats against hardware-enforced feature and security boundaries. Once flawed silicon is in the field, no Windows Update or driver patch can re-wire a fuse controller.
Hardware makers don't always lock down silicon out of pure malice; yields, thermals, and warranty margins dictate most product tiers. Still, every few years, a research paper or a clever BIOS tweak pulls back the curtain—reminding us that inside the box, our hardware sometimes knows how to do far more than its price tag lets on.
References
- Voltage glitch unlocks 80GB on Nvidia's CMP 170HX, but 'faster than an A100' rests on a single benchmark | TechRadar techradar.com
- This is how PC hardware makers intentionally limit performance to sell you more expensive gear How-To Geek · 2026-11-22T16:30:15+00:00
- github.com github.com