Unlocked CMP 170HX cards are having a moment

An unlocked CMP 170HX is the strangest good deal in local AI right now, and it is easy to oversell.

NVIDIA shipped the 170HX in 2021 as a mining card: a cut-down GA100, the same family as the A100, with no display outputs. The common card reports 8GB of HBM2e and runs its compute units far below what the silicon can do. A community unlock, the one LTT Labs measured in September 2026, makes that 8GB card report 64GB and turns the compute throttle off. A less common 10GB version unlocks to 40GB, so “64GB” is not every listing.

I have not logged a checked street price the way we do for the 3090. Asking prices on unlocked cards in mid-October 2026 were scattered around $3,000 to $3,500, and plenty of listings are already modded by the seller. Treat that as a snapshot of ads, not a price we stand behind.

What the measurements say

LTT Labs ran llama.cpp on one card, locked and unlocked, against an RTX 3090 Ti. Flash attention on, 128 generated tokens, prompts of 512 and 8,192, no context depth. These are their numbers:

Model	Unlocked 170HX, tg	RTX 3090 Ti, tg	Unlocked pp512
Granite 4.1 8B Q4_K_M	113 tok/s	136 tok/s	2,994 tok/s
Qwen 3.6 27B Q4_K_M	38 tok/s	47 tok/s	907 tok/s
Qwen 3.8 27B Q8_0	33 tok/s	31 tok/s on two 3090 Tis, layer split	897 tok/s

The locked card could not hold either 27B model. The 8B model generated at 30 tok/s locked and 113 unlocked. Prompt processing on the 8B was 351 tok/s locked and 2,994 unlocked, still behind the 3090 Ti at 5,142.

That is the whole story. Generation is close to a 3090 Ti because HBM2e is in the 1.5 TB/s class and generation is bandwidth-bound. Prompt processing still loses, because the unlocked chip is a cut-down GA100 at a 250W rating, not a 450W 3090 Ti. The 64GB is what a 3090 cannot offer on one card: a 27B model at Q8 with room left for a long context.

Why it is not the new default

The PCIe link stays narrow. The unlock LTT tested brings the card to PCIe Gen2 x4, about 1.7 GB/s from host to card. A full x16 link is a board modification, not a driver checkbox. Loading a model is slower, and splitting one model across several of these cards is a worse idea than splitting it across 3090s.

There is no monitor output. The driver path is a patch on Linux, not the stock NVIDIA install you get with a GeForce card. These are used mining boards. Some have been recapped, re-shrouded, or had the PCIe link modified, and a listing that says “64GB” is a claim you should watch the card enumerate before you trust it.

At $3,000 to $3,500 you are in RTX 4090 territory, or three used RTX 3090s. Three 3090s are 72GB, more power, more noise, and a normal CUDA stack. One unlocked 170HX is 64GB on a single 250W board, which is the case that is actually new.

So yes, these cards are dominating a specific argument: the most VRAM you can get on one used NVIDIA GPU. They are not dominating the speed chart, and I would not tell someone to start here. Start with a 3090. Look at an R9700 if you want new 32GB cards you can stack with a supported ROCm stack. Look at a 170HX only if 64GB on one CUDA card is the requirement, and you are willing to own the patch and the narrow PCIe link that come with it.