GPU servers · NVIDIA CMP 170HX · GA100
CMP 170HX server. All 64 GB switched on before you log in.
A CMP 170HX server from us arrives unlocked: 64 GB of HBM2e instead of the 8 GB NVIDIA left visible, FP32 at full rate and the link at PCIe Gen 2. We build the patched NVIDIA open driver into Linux while we install the machine, test the memory of each GPU, and hand you root.
The GPU is a GA100, the die in the A100. It left the factory with most of itself fenced off in firmware. The cmpunlocker project patches NVIDIA's open kernel driver so those fences come down each time the driver starts the GPU. Nothing is flashed. The machine is single tenant: nobody else runs on it.
- 64 GB HBM2e per GPU
- Ubuntu, Debian, AlmaLinux, Rocky
- Root access
- Monthly term
Unlocked CMP 170HX · before and after
What the patched driver changes on the same GPU
LTT Labs unlocked a CMP 170HX in September 2026 and tested it locked and unlocked on one bench. The figures below are theirs, and each row names its source. Token rates move with the model, the quantisation, the llama.cpp build and the flags, so read them as a ratio, not a promise.
One square per gigabyte of HBM2e.
- 8 GB visible as NVIDIA shipped it
- 56 GB more once the patched driver starts the GPU
Memory is the change that matters
At 8 GB the GPU holds a small model and little else. At 64 GB it holds a 27-billion-parameter model at 8-bit with room left for context. That is the line between a GPU you test on and one you serve from.
Generation follows bandwidth, prompt reading follows compute
Writing tokens is limited by how fast the weights stream out of memory, and HBM2e was always fast at that. Reading a long prompt is limited by the multiply-add units. The throttle hurt there most, so that is where the unlock shows the biggest gain.
The link matters at load time only
Gen 2 x4 doubles the host-to-GPU rate, so the bus part of a model load takes half as long. Once the weights are on the GPU, the bus sits idle.
| Measure | As NVIDIA shipped it | Unlocked | Source |
|---|---|---|---|
| GPU 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 GPU, 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 |
Sources: LTT Labs, “Actually Download More RAM” (12 September 2026) for every measured row; 170th Street for the stock FP32 figure; cmpunlocker for what the patches restore.
The install, step by step
How the unlock gets onto your CMP 170HX server
Anyone can read the cmpunlocker README. The work is everything around it, done before handover and kept working after. You run nothing.
-
NVIDIA's open driver, pinned to one release
The installer adds NVIDIA's CUDA repository and installs the open driver userspace for
610.57.04, headless, without DKMS, and holds it at that release: thenvidia-driver-pinning-610.57.04package on Ubuntu and Debian, a dnf version lock on AlmaLinux and Rocky. The unlock is built for one exact driver. The pin stops a routine upgrade from pulling the libraries out from under it. -
cmpunlocker, compiled for your kernel
We fetch cmpunlocker at a fixed commit and build its patched
nvidia.koinside the new system, against the kernel it is about to boot. The patches change how the driver brings the GPU up. Nothing is written to the GPU, so removing the driver returns it to exactly what it was. -
A cold first boot
The unlock only takes on the first driver load after the GPU has been powered down. So each install ends with the server switched off and powered back on through its BMC, and the first driver to touch the GPU is the patched one.
-
Checked and memory-tested at first login
When the new system first answers, the install asks each GPU what it came up as: memory, PCIe generation and link width. Then it writes and reads back every byte of memory the driver hands out, with fixed, address and pseudo-random patterns. Both results go to the server's setup log, where support can see them.
cmp-unlock statusandcmp-unlock memtestgive you the same answers from a shell. -
Kernel updates do not relock it
A kernel hook (
/etc/kernel/postinst.d, or/etc/kernel/install.don AlmaLinux and Rocky) rebuilds the patched modules for each new kernel from a copy of NVIDIA's source kept on the machine. An unattended upgrade will not send the GPU back to 8 GB. The package manager also refuses the packages that would install a second, stock driver beside it. Install CUDA withapt install cuda-toolkitordnf install cuda-toolkitand the conflict never comes up.
CMP 170HX 64GB · sizing
What fits in 64 GB, and what fitted in 8
Weights only, rounded up. A model counts as fitting when the GPU holds it with a quarter to a half again to spare for the KV cache and the runtime. Below that it loads, then runs out of memory as the context grows. Our LLM page sizes every GPU by the same rule.
| Model size | Precision | Weights | 8 GB, as shipped | 64 GB, unlocked |
|---|---|---|---|---|
| 7-8 billion parameters | 16-bit | 16 GB | does not fit | fits |
| 7-8 billion parameters | 8-bit | 8 GB | does not fit | fits |
| 7-8 billion parameters | 4-bit | 5 GB | fits | fits |
| 13-14 billion parameters | 16-bit | 28 GB | does not fit | fits |
| 13-14 billion parameters | 8-bit | 14 GB | does not fit | fits |
| 13-14 billion parameters | 4-bit | 8 GB | does not fit | fits |
| 30-34 billion parameters | 16-bit | 68 GB | does not fit | does not fit |
| 30-34 billion parameters | 8-bit | 34 GB | does not fit | fits |
| 30-34 billion parameters | 4-bit | 19 GB | does not fit | fits |
| 70 billion parameters | 16-bit | 140 GB | does not fit | does not fit |
| 70 billion parameters | 8-bit | 70 GB | does not fit | does not fit |
| 70 billion parameters | 4-bit | 38 GB | does not fit | fits |
Two GPUs in one machine hold what one cannot: a 70-billion-parameter model at 8-bit needs about 70 GB, which fits across a pair. Splitting a model over several GPUs or machines is covered on the multi-GPU page. Give the machine at least as much system memory as the largest model file, or loading slows down. The entry build's memory is listed in the price table below, and the configurator adds more.
Read this before you order
A cheap GPU for AI inference, and where it stops
An unlocked CMP 170HX gives a lot of inference memory for its monthly price. It is not an A100, and these are the places that shows.
- No ECC
- Error correction on the HBM2e stays off; nobody has unlocked it yet. For inference a flipped bit costs a retry. For a week-long training run, or work that must be bit-exact, choose a GPU with protected memory, such as the L40S.
- PCIe Gen 2 x4
- About 1.7 GB/s from host to GPU, measured, so the bus needs over 20 seconds to move a 38 GB model. A model crosses it once and then lives in GPU memory, so serving barely notices. A training loop that streams batches from system memory notices every step.
- Ampere number formats
- FP16, BF16, TF32 and INT8 on the tensor cores. No FP8 and no FP4; both came with later generations. Quantised models in GGUF, AWQ or GPTQ run fine, because they unpack to formats the GPU has.
- Linux only
- The unlock is a Linux kernel driver. Order the machine with Ubuntu, Debian, AlmaLinux or Rocky Linux and it is built in. On Windows or Proxmox VE the GPU runs as NVIDIA shipped it, with 8 GB.
- Secure Boot off
- The patched modules are unsigned. A kernel with Secure Boot enforcing refuses them and leaves the GPU with no driver at all, so keep it off.
- Community code, not NVIDIA's
- cmpunlocker is an open-source project (GPL-2.0) by Amogh Munikote, and NVIDIA does not support it. We install a commit we have built and checked, and we move that pin by hand, never automatically.
- The 10 GB version stops at 40 GB
- The 8 GB CMP 170HX unlocks to 64 GB. The 10 GB version carries more memory on its package than stays stable under load, so the unlock holds it at 40 GB. The install reads which version is in the machine and applies the right memory layout to each GPU.
The machine behind the button
One unlocked CMP 170HX in the cheapest machine that takes it
The GPU is one line on the invoice and the machine is another; the two add up. The figures below are the least this build costs: the GPU in the cheapest complete machine that takes it, before any sales tax. In the configurator you can change the processor, memory, disks and GPU count, and it prices each change as you make it.
| GPU | NVIDIA CMP 170HX: 64 GB HBM2e unlocked (8 GB HBM2e without the patched driver) |
|---|---|
| GPU, 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 |
| GPUs 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 22.04–26.04, Debian 12/13, AlmaLinux 9/10 or Rocky Linux 9/10 |
| Memory test | Every byte the driver hands out, written and read back on each GPU before handover |
The setup charge exists because the GPU is fitted to order, which makes the machine a custom build. It is charged once, and the monthly figure stays the same at renewal. Every other GPU we fit is priced the same way on the GPU servers page.
Answers
CMP 170HX server questions
What is an unlocked CMP 170HX?
An NVIDIA CMP 170HX running a patched build of NVIDIA's open kernel driver from the cmpunlocker project. The GPU is a GA100, the die in the A100, which NVIDIA shipped with 8 GB of its HBM2e visible, FP32 multiply-add throttled and the link held at PCIe Gen 1. The patched driver brings up 64 GB, removes the throttle and trains the link at Gen 2 each time it starts the GPU. Nothing on the GPU itself is modified.
How much VRAM does a CMP 170HX have when unlocked?
64 GB of HBM2e on the 8 GB version, which is what nvidia-smi reports once the patched driver runs. The 10 GB version is held at 40 GB, the size that stays stable on it. Without the patched driver, either one shows what NVIDIA shipped: 8 or 10 GB.
Is the CMP 170HX good for AI inference?
Yes, for models that fit in 64 GB and work that does not need ECC. In LTT Labs' tests an 8B model at 4-bit generated 113 tokens per second unlocked against 30 locked, and a 27B model at 8-bit ran at 33 tokens per second. The limits are a PCIe Gen 2 x4 link that slows model loading, no FP8 or FP4, and no ECC on the memory.
Can a CMP 170HX run a 70B model?
At 4-bit, yes: the weights come to about 38 GB, which leaves the head room our sizing rule asks for on a 64 GB GPU. At 8-bit they come to about 70 GB, which needs two GPUs; one machine takes up to four.
Does the unlock survive reboots and kernel updates?
Yes. The patched modules are installed like any other kernel module and load on every boot. When a new kernel arrives, a hook rebuilds them for it from the NVIDIA source kept on the machine, before you boot into it. If a GPU still comes up at 8 GB, run cmp-unlock build with your kernel version and power-cycle the server from the control panel.
Which operating systems get the unlocked driver?
Ubuntu 22.04, 24.04 and 26.04, Debian 12 and 13, AlmaLinux 9 and 10, and Rocky Linux 9 and 10, installed by us on any machine with the GPU in it. It is part of the install, not an option to tick. Windows and Proxmox VE get the GPU as NVIDIA shipped it: the unlock is a Linux kernel driver, and on Proxmox the GPUs usually go to virtual machines, which a host driver does not reach.
Can I use CUDA, PyTorch, vLLM and llama.cpp on it?
Yes. The driver is NVIDIA's own 610.57.04 with the patches applied, so software built for an Ampere GPU with a current driver runs. Install the CUDA toolkit on its own, with apt install cuda-toolkit or dnf install cuda-toolkit. The cuda and cuda-drivers meta-packages would bring a second, stock driver with them, so the package manager is set to refuse them.
How many CMP 170HX GPUs can one server take?
Up to four in the machine this page links to. Choose the count in the configurator. For more than four, or a model split across machines, see the multi-GPU page.
Is the unlocked memory reliable?
The 8 GB version at 64 GB is the combination that has held up in public testing, and the 10 GB version stops at 40 GB for the same reason. It is still memory NVIDIA binned, and one GPU can carry a weak patch: LTT Labs found one among those they tested. So every install tests it. After the first cold boot, every byte the driver will hand out is written and read back on each GPU, with fixed, address and pseudo-random patterns, each inverted in place, and the result goes to the setup log. That is a check at handover, not a guarantee for years. If a GPU misbehaves later, run cmp-unlock memtest and send us the output. A GPU that fails is replaced like any other part.
Why is there a setup charge on this build?
Because the GPU is fitted to order. A machine with a GPU added is a custom build, and a custom build carries the one-time setup charge shown in the price table. It is charged once, and the monthly figure does not change at renewal.
Where are CMP 170HX servers located?
In the cities listed in the price table. The machine line is also sold in Amsterdam, but the configurator cannot fit a GPU there, so an unlocked CMP 170HX is not available in that city.
How do I get help with the server?
Open the chat on this site. An AI assistant answers in seconds and can check your account and your server. Ask for a person and one takes over.
GPU marketplace
CMP 170HX machines listed by their owners
Also rentable by the month: CMP 170HX machines that owners list on our marketplace, each at a price its owner sets. Some are listed through our sister company, Primcast, and those are rented on primcast.com.
Elsewhere
If the CMP 170HX is not the right fit
- GPU serversEvery GPU we fit, each priced as its own monthly line beside the machine that hosts it.
- LLM serversWhich machine for which model, from CPU-only boxes up to data-centre GPUs.
- Multi-GPU serversSplitting a model across GPUs and machines, and what the bus and the port cost you.
- L40S servers48 GB with error correction, for jobs that cannot tolerate a flipped bit.
Configure your CMP 170HX server
Pick the processor, memory, disks and up to four GPUs. The unlocked driver goes in with Linux, and each GPU's memory is tested before you get the login. Questions first? The chat answers in seconds, and a person takes over when you ask.