What does a CMP unlock actually give back?
Speed. Every NVIDIA CMP mining card left the factory with its compute throttled, and a CMP unlock switches it back to full rate. On a CMP 90HX that is FP32 going from 0.72 TFLOPS to 18.24, about 25 times, with no firmware flash and no fuse burned. What it does not give back is memory, and that distinction is the whole article.
The memory story belongs to one card. The CMP 170HX was sold as an 8 GB card with far more HBM2e on the package, and cmpunlocker opened 64 GB of it in software. No other CMP card has hidden memory to open. What the rest of the line shares with the 170HX is the throttle on its compute, and that is what this page covers, card by card, for people who want to run local LLMs on them.
Nvidia throttled every CMP card the same way
The SMs were never switched off. What NVIDIA limited was how fast they may issue instructions, through two registers the projects call speed selectors. frolovser-boop's CMP 90HX research puts it in one line: the SMs were not blocked, the instruction issue rate was.
The unlock writes those selectors back to full rate at every boot (SS0=0x88888888, SS1=0x8), before or while the NVIDIA driver binds the card. Wh1stle05's CMP 90HX repository is explicit about what it does not do: no VBIOS flash, no OTP or fuse writes, only runtime register writes and a patched kernel module. Nothing is written to the card itself: remove the module and the next boot is stock.
iatethelogs, who wrote the three-part Habr series on the 90HX, summed up what "mining card" meant: NVIDIA handed you almost a whole GA102 and did not let you use all of it.
The same two selectors exist on GeForce cards, where they hold FP16 tensor math with FP32 accumulation at half rate. That is the GPU unlock for RTX 20, 30 and 40 cards: up to 80% faster BF16 matrix multiplication on an RTX 3090 Ti, and an RTX 4090 going from 165.6 to 331.6 TFLOPS of FP16 tensor throughput with its fork, by TechPowerUp's count. On a GeForce the limit is a half. On a CMP card it was a factor of about 25.
What a CMP 90HX runs after the unlock
@pirosap bought a used CMP 90HX on Xianyu for 1,250 yuan. The seller had re-pasted it and a technician had soldered capacitors onto the board. It has 10 GB of GDDR6X at 760 GB/s on a GA102. He unlocked it (FP32 from 0.72 to 14.31 TFLOPS, PCIe from Gen1 x4 to Gen2 x16), then asked the obvious question: is a 10 GB card with a narrow bus any use?
His answer is Strata's standard benchmark, run in full, 21 of 21 requests, 12 of 12 sanity checks:
| Qwen 3.8 Flash Next, IQ3_S, 131,072 context | Prompt reading | Writing |
|---|---|---|
| Short prompt | n/a | 52.1 tok/s |
| 4K prompt | 639.9 tok/s | 47.5 tok/s |
| 16K prompt | 706.4 tok/s | 52.0 tok/s |
That is a 180B model, 6B active, on one 10 GB mining card, a Core i9-7960X and 78 GB of RAM, on Ubuntu 24.04.5 with Strata engine 0.1.61. Strata keeps the most-used experts on the card and the rest in system RAM, which is how a 10 GB card runs a model that size at all. His own verdict: not bad for under 30,000 yen. You can size the same machine in the Strata calculator, which lists the card as CMP 90HX 10GB.
Smaller models fit whole. Wh1stle05's unlocked 90HX ran Qwen3.8-27B at 2 bits at 35.5 tok/s writing and 840 tok/s reading, at 32K of context in 9.0 GB of VRAM.
Prompt reading is what the unlock fixes
Writing a token is mostly a memory read, and 760 GB/s was never locked, so writing speed was always set by the memory. Reading a prompt is compute, and that is where the throttle hurt. frolovser-boop's before and after on llama-bench: prompt processing at 512 tokens went from 224 to 1,770 tok/s, about 7.9 times, with generation at 55.5 tok/s.
iatethelogs' five-card server shows the same thing at scale. Five unlocked CMP 90HX, 49,383 MiB visible to llama.cpp, Qwen3.8-27B-Heretic, every layer on the cards: layer split read prompts at 1,992.67 tok/s and wrote at 23.76. Tensor split wrote faster, 31.93, but its prompt reading fell to 268.46, because on PCIe Gen1 the cards spent their time waiting on each other.
PCIe: Gen1 out of the box, Gen2 after the unlock
The CMP cards also ship with PCIe held at Gen1. Wh1stle05 measured host-to-card copies at 3.08 GB/s on Gen1 x16 and 6.29 GB/s on Gen2 x16 after the unlock. For one card running a model that fits, the link mostly costs load time. For Strata streaming experts from RAM, and for several cards splitting a model, it costs speed.
Width is a separate problem. On some boards the 220 nF capacitors on the upper lanes were simply never fitted, so the slot runs at x8 or x4 electrically. Soldering them on restores x16. iatethelogs did exactly that on his five cards and reports the first confirmed PCIe Gen2 x16 on a CMP 90HX.
Every CMP card, where the CMP unlock stands
| Card | Stock memory | Compute unlock | Measured LLM runs we found |
|---|---|---|---|
| CMP 30HX | 6 GB, 336 GB/s | Listed by one Windows toolbox, no measured unlock found | none |
| CMP 40HX | 8 GB, 448 GB/s | A UEFI unlock existed (CMP40HX-Unlock v3.0.0); its repository is gone | none |
| CMP 50HX | 10 GB, 560 GB/s | Proven: FP32 31.7x, plus Gen2 PCIe, one-click on Linux and Windows | two 20 GB cards (memory-modded), 75 tok/s in Strata |
| CMP 70HX | not in our list | Listed by one Windows toolbox, no measured unlock found | none |
| CMP 90HX | 10 GB, 760 GB/s | Proven by several projects, including VBIOS .07 and a Windows EFI loader | 52.1 tok/s on a 180B model in Strata |
| CMP 170HX | 8 GB, 1,493 GB/s | Proven, and its hidden memory too | the CMP 170HX page |
The CMP 50HX is the other proven card. UzixLS's unlock, ported from the 40HX's UEFI application, measured FP32 going from 0.427 to 13.512 TFLOPS, 31.7 times, against NVIDIA's own stock driver. The same project's patch also handles 50HX cards rebuilt to 20 GB, validated on four cards reporting 20,480 MiB each.
More memory is a soldering job, not an unlock
Those 20 GB CMP 50HX cards did not unlock anything. Their owners replaced the memory chips with larger ones, the same rework that produces a 22 GB RTX 2080 Ti or a 48 GB RTX 4090. It needs a rework station and the hands to use it, and any consumer card with room for bigger chips can take it. xrip runs two of them in a dual Xeon E5-2670 v2 box with 64 GB of DDR3, and Qwen 3.8 Flash Next answered at 75 tok/s in Strata after a 172K-token prompt.
For the CMP 90HX the same swap is still a plan: frolovser-boop lists the 20 GB mod as theory, proven only on the 50HX.
That is why the CMP 170HX stands apart. Its 64 GB was on the package from the day it shipped, and a script opened it. Every other CMP card has exactly the memory it was sold with, unless someone solders on more.
Before you unlock a CMP card
- Linux first. The proven paths run on Linux with NVIDIA's open kernel module 610.43.03 and Secure Boot off, because the patched module is unsigned. Windows has an EFI loader for the 90HX and a one-click bundle for the 50HX.
- Check your VBIOS. The 90HX paths differ by VBIOS version: .01 and .05 on one route, .07 on another. cmpunlocker refuses to run on a driver, kernel or VBIOS it does not list.
- Nothing is written to the card. None of these compute unlocks flashes the VBIOS or burns a fuse; the installed module redoes the state at every boot, and removing it returns the card to stock.
- L2 stays half. The 90HX's L2 cache is 2.5 MiB, half of a full 320-bit GA102, and no unlock changes it.
- Size it before you buy. A 10 GB card runs a 180B model only through Strata's expert streaming, so system RAM matters as much as the card. The GPU checker and the Strata calculator answer that for your machine.
FAQ
What is a CMP unlock?
A boot-time patch that sets an NVIDIA CMP mining card's speed selectors back to full rate, so its SMs issue instructions as fast as the chip allows. NVIDIA had throttled them. It flashes no VBIOS and burns no fuse; uninstall it and the card boots stock again.
Does a CMP unlock add VRAM?
Only on the CMP 170HX, which carries far more HBM2e than it was sold with; cmpunlocker opens 64 GB on the 8 GB card. Every other CMP card keeps its stock memory. A 20 GB CMP 50HX is a chip swap done with a soldering station, not an unlock.
How much faster is a CMP 90HX after the unlock?
FP32 rises from 0.72 to 18.24 TFLOPS in Wh1stle05's measurement, and prompt processing at 512 tokens from 224 to 1,770 tok/s in frolovser-boop's. Writing speed is set by the 760 GB/s memory, which was never locked.
Can a CMP 90HX run a large model?
Yes, through Strata. @pirosap's unlocked 10 GB CMP 90HX ran Qwen 3.8 Flash Next, a 180B model, at 52.1 tok/s with a short prompt and 52.0 at a 16K prompt, at 131,072 tokens of context, with 78 GB of system RAM holding the experts that do not fit.
Which CMP cards can be unlocked?
The CMP 50HX, 90HX and 170HX have proven unlocks with published measurements. The 40HX had a UEFI unlock whose repository is now gone. The 30HX and 70HX appear only in one Windows toolbox's list, with no measured unlock we could find.
Is the CMP unlock the same as the GeForce GPU unlock?
Same registers, different severity. On GeForce RTX 20, 30 and 40 cards the speed selectors hold FP16 tensor math with FP32 accumulation at half rate, and the GPU unlock doubles it. On a CMP card the throttle covered ordinary FP32 too, at about a twenty-fifth of the chip.