Built a home server from an old PC with GPU upgrade. Qwen3.8 27B runs at ~30 tokens per second.

I needed a relatively simple but acceptable level of AI for working on one project. I didn't have any heavy requests, I just needed to give the AI access to the project files so it could search through them for bugs and stuff. I already had an old computer that I decided not to throw away and instead give it a new life as a git server (and sometimes a minecraft server).

Well, my idea stopped at maintaining the computer, because the gpu, a GTX 1070, was overheating. It needed a complete repaste, but the cooler screws were completely stripped, so while trying to remove the cooler I accidentally knocked off several important smd components with a screwdriver. R.I.P. GPU.

Without a GPU, inference was running entirely on the CPU, and only MoE models were kind of usable, giving around 10–20 tokens per second, while dense models couldn't get past 3 tokens per second. I didn't even get to test it with the GTX 1070, because I decided to service it lol.

I started looking for a replacement on the secondary market, but quickly realized that this would cost too much for the minimum entry point I wanted for experimenting with AI. Until my eyes fell on mining cards. There were plenty of cmp40hx, cmp50hx, cmp70hx and cmp90 cards for sale, and the prices were pretty reasonable(it was a month ago), considering that I was originally looking for a cheap replacement for my dead one.

Getting closer to the actual build, I started calculating how much vram I would need for ± acceptable AI with tolerable speed, and after getting inspired by this sub I decided to take a step further and went with a modified cmp50hx with 20gb of memory and pcie modded to 16 lanes. Very quickly after that I bought another one, this time unmodified (10 gb, only 4 pcie lanes). So, together that's 30gb of vram. Both cards cost me $250 in total (it was also a month ago, right now they've suddenly doubled the price).

Luckily for me, around the same time new driver patches appeared that almost completely remove the limits on their compute performance, and even add pcie 2.0 support (these things have pcie 1.1).

I initially tried one of the newly released cmp50hx driver patches, but it ended badly and I had to spend a lot of time trying to get the drivers working. The patches were new and didn't account for the 20gb version. Later the author fixed that, but even then the driver didn't work for me because of some other problem that I don't want to get into.

I went digging through the driver's github issues and quickly found a guide posted there by another user.

And yes, now the drivers work, the cards are detected and even pcie 2.0 works, but not without problems. The author of the guide said that pcie 2.0 support was only confirmed on the X99 chipset. Well, it also works on Z97, however after waking the computer from sleep the driver crashes completely. After looking into it a bit, I quickly came to the conclusion that the problem was specifically with the pcie patch. Disabling sleep completely solves the only problem I had while using them. :)

Without the specially patched drivers, Qwen 3.6 27B did around 15–20 tokens per second without mtp. MoE models were faster, giving 45–50 tokens per second. With the new drivers, performance doubled. With Qwen 3.8 27B(mtp on) I get around 30–35 tokens per second now, and prompt processing is around 300–400 tokens (including degradation as the token count increases). Ornith 1.5 35BA3B(Heretic-MTP-APEX-I-Balanced) gives around 80–100 tokens per second. I capped GPUs at 180w power limit due to the psu I have, I don't want it to work at its limits, but running them at their 225w surely boosts speeds.

In general, I ended up making a lot of presets for different quantizations with different quality and KV cache sizes (I still need to test all of this in real work), but if we take the better options, I managed to get a Q6K model with 130k context (K – Q8_0 and V – Q5_1).

I also followed a guide for running 27B Qwen with large context on limited vram. Using the same general approach, I managed to get 256k context with K and V Q8_0. The speed is slower though, around 10–14 tokens per second, and prompt processing is around 40–50 tokens. Maybe I can tune it even further. I needed this preset for tasks that I can leave generating overnight :3

- Not enough VRAM. The ideal option would be having 2 identical cards with the same amount of memory. You can run dense models in tensor mode, split the weights evenly and get increased generation speed for basically free + fit the full context. I tried many variations of Qwen 3.8 27B quantization, but the only one I could properly run in 1,1 tensor split was Q4KM (the Unsloth one) together with mtp = ~40 tokens per second. However, there is critically little space left for the KV cache, because it gets distributed together with the model weights, and the second 10gb card simply became the bottleneck. Without mtp, running models with a 1,1 split basically loses its purpose. Pcie 2.0 and the number of lanes probably also play a significant role here. That could in principle be solved by adding more pcie lanes to the second GPU and perhaps buying an nvlink cable(who even does that?), but I decided it wasn't worth it just to get another 5–6 tokens per second.

- No NVMe SSD. Yeah, all models are loaded from a sata ssd so the speed is around 500mb. It's terrible. The motherboard actually has an m2 sata slot with a pcie 2.0 x2 interface, but even its 1gb per second would be too slow for fast model loading. This could be solved by installing an expansion card into one of the pcie slots (there is one free pcie 3.0 x4 slot), but the current price of those things including the ssd is too high, considering that I'm building a cheap system from what I already have with minimal additional spending for an acceptable result. Model switching takes 1–2 minutes. But whatever. (Not whatever, i'm buying a cheap used 256gb nvme ssd :D)

- Amount of RAM and Linux (Ubuntu Server 24.04). Apart from AI, I also run gitlab on the server. And here is the problem: after loading a model, all available ram gets cached by the system for the model files. I'm talking about file cache, not KV. The system was leaving around 200–300mb of free ram for everything else. As a result, openwebui and gitlab started acting laggy (after the model was loaded), as well as the kde plasma interface I installed. I don't completely understand why linux decided to keep this cache until the very last moment instead of freeing it for other programs. I tried adding the no-mmap parameter to the model presets, but nothing helped, and I had to manually clear the cache after loading models, which obviously wasn't acceptable. Together with chatgpt (who else?), I made a command that launched the model and then cleared the cache. It turned out that this broke llama-server, causing model switching to stop unloading the previously loaded model.

- Llama-server flexibility. The list of presets is defined in the models.ini file, where each parameter is in key-value format. For running Q6K with 256k context according to the guide, I needed to set GGML_CUDA_DISABLE_GRAPHS=1, which applies to the entire cuda environment and remains active even after unloading the model. That's undesirable, because with it enabled I lose 1–2 tokens per second on other presets.

So for one specific preset I need to enable cuda graphs, while for the other presets I need to disable them. The llama-server parser does not support things like this, and doing it manually is not an option either.

Together with the other problem with ram cache getting stuck, this led me to making an alternative way to launch the models and proxy requests to llama-server.

To solve problems 3 and 4, I made a launcher (well, chatgpt did, because I'm not a server/python specialist) that proxies requests to llama-server but takes over the functionality of collecting model presets from .sh files and launching them. It also clears ram page cache after loading a model.

In case someone needs that launcher, I can leave it in the comments, along with any other links to the drivers, fixes, build params, etc. Just ask. (Reddit removes the post when I include them, not enough karma, I guess.)

Overall, I’m pretty happy with how this setup turned out. The performance is much better than I expected from these cards, especially considering how cheap they were. I had a hard time getting the patched drivers to work and linux didn't make my life any easier, and sometimes I even regretted buying these GPUs, but in the end, it was worth it. Qwen 3.8 27B really works like Opus 4.5