The CMP 170HX Memory Unlock Works, With Caveats

A mining card NVIDIA sold for $4,299 in 2021 has had its memory ceiling and its compute limits removed by a driver patch, and it now runs 64 GB of HBM2e for used-market money. The catch is that this silicon was binned for a reason, and the used market has worked out where the bad cards need to go.

What NVIDIA Actually Turned Off

The CMP 170HX is not a repurposed gaming die. It is GA100, the processor behind the A100, sold with four or five working HBM2e stacks instead of the full complement and then disabled further in software. A fully enabled GA100 never shipped in any product, so this part was already deep in a binning sort before anyone locked it.

What buyers received was 8 GB of a 64 GB card, floating-point throughput throttled until general-purpose compute was pointless, and four PCIe lanes at Gen1 speeds. The memory and the cores were present, just not addressable. Mining needed bandwidth and integer math, which the card kept, so the restrictions cost NVIDIA's mining customers nothing and cost everyone else the entire product.

It is reversible only because NVIDIA did the disabling without permanent fuses. That is the finding at the centre of Jon Pry's 2026 paper on defeating stack protection in the GPU's secure coprocessor, and it is what makes cmpunlocker, Amogh Munikote's driver-level patch, possible. It patches the driver stack so the card reports and addresses what it physically has. Two limits survive: PCIe Gen3 and on-die ECC stay locked, with no published way back.

The Install Is Easy. The Silicon Is Not.

The process is a Linux chore list rather than an exploit. Disable Secure Boot, purge the NVIDIA packages, blacklist nouveau, rebuild the initramfs, pin the driver to 610.43.02, install the open module, reboot. Afterwards nvidia-smi reports 65,536 MiB instead of an anonymous graphics device, which is when it stops feeling like a rumour. The unlock is per card and survives reboots; the driver patch is per system and does not survive a kernel update, so pin the kernel.

Then the part nobody wants to hear: these dies were binned because they failed something. The 8 GB variant generally unlocks all 64 GB. The 10 GB variant has been far less reliable, with about 40 GB of its theoretical 80 GB coming up stable.

Which is why the memory test is not optional. memtest_vulkan runs a five-minute pass and recommends two hours for anything going into production, and the difference is not academic: a failing card threw errors at one narrow address range and kept doing it after an hour of soaking at 85°C in case the fault was temperature-dependent. It was not. There is no clean way to quarantine bad blocks yet.

That creates an ugly market. Somewhere between 18,000 and 36,000 were sold, and the obvious play for anyone with a bench is to test, keep the good ones and resell the rest. Assume it is happening, and treat a screenshot of somebody's memory test as proof of very little.

Gen2 x16 Is Where It Earns Its Place

The one unlock that needs a soldering iron is the lane count. The traces are on the board; the capacitors are not. Twelve 0402 0.22 µF parts take nano-tweezers, a magnifier and a steady hand to go from four lanes to sixteen, and cmpunlocker then raises the link from Gen1 to Gen2. It is still slow, but roughly eight times the bandwidth the card starts with:

• RTX 3090 Ti, PCIe 4 x16: 23.04 GB/s to host, 26.27 GB/s from host
• Locked 170HX, Gen1 x4: 0.85 and 0.84 GB/s
• Unlocked 170HX, Gen2 x16: 6.70 and 6.74 GB/s

Six gigabytes a second sounds fatal until you remember the workload. Loading a 27B model at Q4 takes 11 seconds at Gen2 x4 and 6 seconds at Gen2 x16, identical to a 3090 Ti on Gen4 x16. Most local inference loads a model once and then sits in token generation, which is bound by memory bandwidth rather than by the slot.

That is where HBM2e does its work. In llama.cpp an unlocked card produces 113 tokens per second on Granite 4.1 8B against 136 for a 3090 Ti, and 38 against 47 on a 27B Q4 model — around 80% of the throughput at 250 W instead of 450 W. Prompt processing is the weak spot, as expected on a compute-bound task: 2,994 against 5,142. It is not a 3090 Ti. It is a 3090 Ti-shaped object holding 64 GB at half the power.

The power connector matters for a build: no 12VHPWR, no pair of 6+2-pin cables, just one 8-pin EPS. The card is passively cooled, so budget a shroud and a 120 mm fan to stay under the 85°C limit once compute is unlocked.

The Price Curve Is the Whole Risk

These sold for a few hundred dollars in June 2026, were fetching $1,300 to $1,600 by August, and are clearing $2,000 in September. Each step is the market discovering the same thing at once, while supply is fixed at whatever was mined and retired.

At a few hundred dollars the card was an obvious experiment. At $2,000 it competes with used 3090s and 4090s that need no patch, no kernel pinning, no capacitor work and no memory lottery. The memory-per-dollar argument holds, since 64 GB in one slot is not something consumer cards offer, but you are now paying a novelty premium on a genuine capability.

What I Think

This is the most interesting hardware story of the year and most buyers should not touch it. The unlock is legitimate work that turned e-waste into the cheapest 64 GB of VRAM on earth. It is also a card with no ECC, no Gen3, no display output, a passive cooler you have to engineer around, and a memory subsystem that failed somebody's binning test on the way out of the factory. That is a hobbyist's card, and it should be priced like one.

If you already run local models on a box with a spare slot and you enjoy this work, one validated card at a sane price is reasonable, and the 250 W draw makes it easier to live with than two 3090s. If you are buying because 64 GB looks like a shortcut to a cheap inference server, buy a used 3090 and spend the difference on the rest of the machine. The unlock removes NVIDIA's artificial limits. It cannot remove the ones the silicon shipped with.