Nvidia P40s with XCP-ng 8.3 for inference and light training

gskger

Being curious about Large Language Models (LLMs) and machine learning, I wanted to add GPUs to my XCP-ng homelab. Finding the right GPU in June 2024 was not easy, given the numerous options and constraints, including a limited budget. My primary setup consists of two HP ProDesk 600 G6 running my 24/7 XCP-ng cluster with shared storage. My secondary setup features a set of Dell R210 II and Dell R720 servers, which I use for memory-intensive tasks. The Dell R720 can hold up to two full-sized GPU which lead to my first requirement: a GPU must fit into the Dell R720.

With R720 compatibility as a requirement, gaming GPUs (RTX 2080, 3080/3090, 4080/4090) were not an option. Additionally, they are expensive and do not all come with a lot of VRAM memory. But I admit that those are more powerful compared to what I came up with.

Since I want to test different LLMs with various parameter sizes, my minimum memory size requirement was 24GB VRAM. That clearly reduced the GPU options again, even for compatible low power (~70W) GPUs like the Nvidia RTX A2000 (12GB) or the Nvidia Tesla T4 (16GB).

My budget limit for getting started was around €300 for one GPU. That narrowed down my search to the Nvidia Tesla P40, a Pascal architecture GPU that, when released in 2016, cost around $5,699. In Europe, the P40 is available on Ebay for around €300 to €500, and I was very lucky to get two P40s for around €510 in total. Seeing two P40s on Ebay in the US for $299 with no delivery option for Europe was a painful experience though.

However, I had to make two compromises with the P40, which may be a problem in the future. First, the P40 is lacking Tensor Cores, which are essential for deep learning training compared to FP32 training. Additionally, the P40 is limited by its CUDA compute capability of 6.1, which is lower than that of newer GPUs like the H100 (9.0). At some point, software tools might stop supporting the P40.

To install a second GPU I had to swap the Dell R720 riser card #3 from 2 PCIe x8 slots with a 150W power connector to a 1 PCIe x16 slot with a 225W power connector. Like the K80/M40/M60/P100 the P40 has a 8-pin EPS connector, so you need a special power cable that can be sourced from Ebay. Using the standard Dell general-purpose GPU cable risks damaging the GPU or motherboard. Yesterday, the last part arrived so today is install day.

The process of swapping riser #3, installing the two P40 GPUs, and connecting the power cables was straightforward.. During boot, the server checks PCI devices and updates the inventory, which might take some minutes. After the initial fan ramp-up, the fan speed dropped back to normal and the Dell R720 idles at about 126W with both GPUs installed.

Next step was installing and updating XCP-ng 8.3 beta, which was as easy as installing the GPUs. Adding the host to XO from source and activating the PCI pass-through in the hosts advanced view required a reboot, but after that I could setup an Ollama VM to run LLMs and another Open WebUI VM to chat with the LLMs. With 48 GB of VRAM, I can run llama3-70b with some headroom and about 6 tokens/sec while llama3-8b is much smaller and answers with 23 tokens/sec on this setup.

So what are the next steps? On one hand, I want to setup a development environment for Phyton and API based usage of LLMs (not only local LLMs, but also cloud based LLMs like ChaGPT or Claude). That will be fun, since I have zero experience with that. On the other hand, I will setup more GPU supported services like Perplexica or AUTOMATIC1111 or whisper. Apart from that, I will also try to improve my prompt engineering skills and learn about LLM multi agent frameworks.

The best thing on this setup is that XCP-ng 8.3 beta provides a robust foundation for running Large Language Models (LLMs) and other AI workloads on one machine. Looking forward to the release candidate!

wilsonqanda

@gskger Nice. I am curious too as there are some many opensource LLM available nowadays to test out.

planedrop

Love this, planning to eventually do something similar, mostly running LLMs on my desktop right now but would much prefer to properly host them instead.

olivierlambert

Nice setup! I played with Ollaama myself. I might try to use the UI and make it run for our company internally

JamesG

I just added a P4 to one of my hosts for exactly this. My servers can only handle low-profile cards, so the P4 fits. Not the most powerful of GPU's but I can get my feet wet.