The CMP 170HX carries the same GA100 die as an A100, and it left the factory with most of that die fenced off: 8 GB of its HBM2e visible, FP32 throttled to a trickle, the link held at PCIe Gen 1. A patched build of NVIDIA’s open driver takes the fences down. We compile it into Ubuntu while we install your server, so the card is unlocked before you log in the first time.
LTT Labs unlocked a CMP 170HX in September 2026 and ran it against itself, locked and unlocked, on one test bench. These are their figures, not ours, and every row names where it came from. We quote them for the size of the change: token rates move with the model, the quantisation, the llama.cpp build and the flags, so read them as a ratio rather than a promise.
| Measure | As NVIDIA shipped it | Unlocked | Source |
|---|---|---|---|
| Card memory | 8 GB HBM2e | 64 GB HBM2e | cmpunlocker; 65,536 MiB measured by LTT Labs |
| FP32 fused multiply-add | about 0.39 TFLOPS | unthrottled | 170th Street (stock); cmpunlocker (full SM throughput) |
| PCIe link | Gen 1 x4 | Gen 2 x4 | cmpunlocker |
| Host to card, measured | 0.84 GB/s | 1.68 GB/s | LTT Labs, nvbandwidth |
| 8B model at 4-bit, generating | 30 tokens/s | 113 tokens/s | LTT Labs, llama.cpp, Granite 4.1 8B Q4_K_M |
| 8B model at 4-bit, reading a prompt | 351 tokens/s | 2,994 tokens/s | LTT Labs, llama.cpp, pp512 |
| 27B model at 4-bit (~16 GB) | does not fit | 38 tokens/s | LTT Labs, llama.cpp, Qwen 3.6 27B Q4_K_M |
| 27B model at 8-bit (~29 GB) | does not fit | 33 tokens/s | LTT Labs, llama.cpp, Qwen 3.8 27B Q8_0 |
Memory is the change that matters. At 8 GB the card holds a small model and little else. At 64 GB it holds a 27-billion-parameter model at 8-bit with most of the card still free for context, which is the difference between a card you can experiment on and one you can serve from.
Token generation leans on bandwidth, prompt reading on arithmetic. Generating text is bound by how fast the weights stream out of memory, which HBM2e was always good at; reading a long prompt is bound by the multiply-add units, which is where the throttle hurt most and where the unlock shows up largest.
The link barely matters once a model is loaded. Gen 2 x4 doubles the host-to-card rate, which shortens loading. After that the weights live on the card and the bus is idle.
Anyone can read the cmpunlocker README. The part worth paying for is everything around it, done before handover and kept working after it. There is nothing for you to run.
The installer adds NVIDIA’s CUDA repository and installs the open kernel driver 610.57.04, headless and without DKMS, together with the nvidia-driver-pinning-610.57.04 package that holds every piece of it at that release. The unlock is built for one exact driver; the pin is what stops a routine apt upgrade pulling the libraries out from under it.
We fetch cmpunlocker at a fixed commit and build its patched nvidia.ko inside the new system, against the kernel it is about to boot, before the installer powers off. The patches change how the driver brings the card up. Nothing is flashed to the card, so taking the driver away returns it to exactly what it was.
The unlock only takes on the first driver load after the card has been powered down. Our Ubuntu installs end that way already: the installer switches the server off and we power it back on through its BMC, so the first thing to touch the card is the patched driver.
When the installed system first answers, the install asks each card what it came up as — memory, PCIe generation and link width — and writes the answer to the server’s setup log, where support can see it. cmp-unlock status gives you the same answer from a shell at any time.
A hook in /etc/kernel/postinst.d rebuilds the patched modules for every new kernel, from a copy of NVIDIA’s source kept on the machine, so an unattended upgrade does not quietly send the card back to 8 GB. apt is also told to refuse the packages that would put a second, stock driver beside it; install CUDA with apt install cuda-toolkit and it never comes up.
Weights only, rounded up. A model counts as fitting when the card can hold it with a quarter to a half again to spare for the KV cache and the runtime; below that it loads and then runs out the moment the context grows. It is the same rule our LLM page sizes every card by.
| Model size | Precision | Weights | 8 GB, locked | 64 GB, unlocked |
|---|---|---|---|---|
| 7-8 billion parameters | 16-bit | 16 GB | — | fits |
| 8-bit | 8 GB | — | fits | |
| 4-bit | 5 GB | fits | fits | |
| 13-14 billion parameters | 16-bit | 28 GB | — | fits |
| 8-bit | 14 GB | — | fits | |
| 4-bit | 8 GB | — | fits | |
| 30-34 billion parameters | 16-bit | 68 GB | — | — |
| 8-bit | 34 GB | — | fits | |
| 4-bit | 19 GB | — | fits | |
| 70 billion parameters | 16-bit | 140 GB | — | — |
| 8-bit | 70 GB | — | — | |
| 4-bit | 38 GB | — | fits |
An unlocked CMP is a very good inference card for its price. It is not an A100, and these are the places it shows.
The card is a line of its own on the invoice and the machine is another; the two add. The figures below are the least this build can cost: the card in the cheapest complete machine that takes it, before any sales tax. The configurator may open with more memory than that; change the processor, memory or disks there and it prices each change as you make it.
| Card | NVIDIA CMP 170HX: 64 GB HBM2e unlocked (8 GB HBM2e without the patched driver) |
|---|---|
| Card, per month | $189 |
| Machine | Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps |
| Machine, per month | $83.67 |
| Together | $272.67 a month |
| Setup | $29 once |
| Cards per machine | 1 to 4 |
| Cities | New York, Bucharest, San Francisco and Miami |
| Term | Monthly; 15% off on a twelve-month cycle |
| Driver | NVIDIA open 610.57.04, patched by cmpunlocker, on Ubuntu 22.04, 24.04 or 26.04 |
