As everyone is aware, the world is still going nuts attempting to establish more, more recent and better AI tools. Mainly by throwing unreasonable quantities of cash at the problem. Much of those billions go towards developing low-cost or totally free services that run at a considerable loss. The tech giants that run them all are intending to attract as numerous users as possible, so that they can record the market, and become the dominant or only celebration that can offer them. It is the classic Silicon Valley playbook. Once supremacy is reached, anticipate the enshittification to begin.

A most likely way to make back all that cash for establishing these LLMs will be by tweaking their outputs to the liking of whoever pays the a lot of. An example of what that such tweaking looks like is the rejection of DeepSeek's R1 to discuss what happened at Tiananmen Square in 1989. That one is certainly politically motivated, however ad-funded services will not exactly be enjoyable either. In the future, I completely anticipate to be able to have a frank and honest discussion about the Tiananmen events with an American AI agent, but the just one I can pay for will have assumed the personality of Father Christmas who, while holding a can of Coca-Cola, will intersperse the recounting of the awful events with a joyful "Ho ho ho ... Didn't you understand? The vacations are coming!"

Or perhaps that is too improbable. Today, dispite all that cash, the most popular service for code completion still has problem working with a couple of easy words, regardless of them existing in every dictionary. There should be a bug in the "free speech", or something.

But there is hope. Among the tricks of an approaching player to shake up the marketplace, is to undercut the incumbents by releasing their design for complimentary, under a permissive license. This is what DeepSeek simply made with their DeepSeek-R1. Google did it previously with the Gemma models, as did Meta with Llama. We can download these designs ourselves and run them on our own hardware. Even better, people can take these models and scrub the biases from them. And we can download those scrubbed models and run those on our own hardware. And after that we can lastly have some really useful LLMs.

That hardware can be an obstacle, though. There are two choices to pick from if you wish to run an LLM locally. You can get a huge, effective video card from Nvidia, or you can buy an Apple. Either is expensive. The main specification that shows how well an LLM will carry out is the amount of memory available. VRAM in the case of GPU's, normal RAM in the case of Apples. Bigger is better here. More RAM suggests bigger designs, which will considerably improve the quality of the output. Personally, I 'd say one requires a minimum of over 24GB to be able to run anything useful. That will fit a 32 billion parameter design with a little headroom to spare. Building, or buying, a workstation that is geared up to deal with that can easily cost thousands of euros.

So what to do, if you don't have that quantity of cash to spare? You purchase pre-owned! This is a viable alternative, but as constantly, there is no such thing as a totally free lunch. Memory might be the main concern, however don't ignore the value of memory bandwidth and other specifications. Older devices will have lower efficiency on those elements. But let's not stress too much about that now. I have an interest in developing something that at least can run the LLMs in a functional way. Sure, the current Nvidia card may do it quicker, but the point is to be able to do it at all. Powerful online models can be good, but one need to at least have the choice to switch to a local one, if the scenario requires it.

Below is my attempt to develop such a capable AI computer system without investing excessive. I ended up with a workstation with 48GB of VRAM that cost me around 1700 euros. I could have done it for less. For example, it was not strictly required to buy a brand name new dummy GPU (see below), or I might have found someone that would 3D print the cooling fan shroud for me, instead of delivering a ready-made one from a faraway country. I'll confess, I got a bit restless at the end when I learnt I had to buy yet another part to make this work. For me, this was an appropriate tradeoff.

Hardware

This is the full cost breakdown:

And this is what it appeared like when it first booted up with all the parts set up:

I'll provide some context on the parts listed below, and after that, I'll run a few quick tests to get some numbers on the performance.

HP Z440 Workstation

The Z440 was a simple pick because I already owned it. This was the starting point. About 2 years earlier, I wanted a computer system that might work as a host for my virtual makers. The Z440 has a Xeon processor with 12 cores, and this one sports 128GB of RAM. Many threads and a great deal of memory, that must work for hosting VMs. I bought it pre-owned and then swapped the 512GB difficult drive for a 6TB one to save those virtual devices. 6TB is not required for running LLMs, and therefore I did not include it in the breakdown. But if you plan to collect many models, 512GB may not suffice.

I have actually pertained to like this workstation. It feels all extremely strong, and I haven't had any issues with it. A minimum of, till I began this task. It ends up that HP does not like competition, and I came across some difficulties when switching components.

2 x NVIDIA Tesla P40

This is the magic active ingredient. GPUs are pricey. But, as with the HP Z440, often one can discover older equipment, that used to be top of the line and is still extremely capable, pre-owned, for fairly little money. These Teslas were indicated to run in server farms, for things like 3D making and other graphic processing. They come geared up with 24GB of VRAM. Nice. They fit in a PCI-Express 3.0 x16 slot. The Z440 has 2 of those, so we purchase 2. Now we have 48GB of VRAM. Double nice.

The catch is the part about that they were meant for servers. They will work great in the PCIe slots of a typical workstation, but in servers the cooling is managed in a different way. Beefy GPUs take in a lot of power and can run very hot. That is the factor customer GPUs always come equipped with huge fans. The cards require to take care of their own cooling. The Teslas, however, have no fans whatsoever. They get just as hot, however expect the server to provide a stable flow of air to cool them. The enclosure of the card is somewhat formed like a pipeline, and you have two choices: blow in air from one side or blow it in from the other side. How is that for flexibility? You absolutely need to blow some air into it, though, or you will damage it as soon as you put it to work.

The option is simple: simply mount a fan on one end of the pipe. And certainly, it seems an entire cottage market has actually grown of people that sell 3D-printed shrouds that hold a basic 60mm fan in just the best place. The problem is, the cards themselves are currently quite large, and it is challenging to find a configuration that fits two cards and two fan installs in the computer system case. The seller who offered me my two Teslas was kind enough to include 2 fans with shrouds, however there was no chance I could fit all of those into the case. So what do we do? We buy more parts.

NZXT C850 Gold

This is where things got annoying. The HP Z440 had a 700 Watt PSU, which might have been enough. But I wasn't sure, and I needed to purchase a brand-new PSU anyway since it did not have the right connectors to power the Teslas. Using this useful website, I deduced that 850 Watt would be sufficient, and I purchased the NZXT C850. It is a modular PSU, indicating that you only require to plug in the cables that you in fact require. It included a cool bag to save the extra cables. One day, I might offer it a great cleansing and use it as a toiletry bag.

Unfortunately, HP does not like things that are not HP, so they made it hard to swap the PSU. It does not fit physically, and they also changed the main board and CPU adapters. All PSU's I have ever seen in my life are . The HP PSU also is a rectangle-shaped box, however with a cutout, making certain that none of the normal PSUs will fit. For no technical factor at all. This is simply to mess with you.

The installing was ultimately resolved by using 2 random holes in the grill that I in some way managed to line up with the screw holes on the NZXT. It sort of hangs stable now, and I feel lucky that this worked. I have seen Youtube videos where people turned to double-sided tape.

The adapter required ... another purchase.

Not cool HP.

Gainward GT 1030

There is another issue with utilizing server GPUs in this customer workstation. The Teslas are meant to crunch numbers, not to play computer game with. Consequently, they do not have any ports to connect a display to. The BIOS of the HP Z440 does not like this. It refuses to boot if there is no other way to output a video signal. This computer will run headless, however we have no other choice. We need to get a third video card, that we do not to intent to utilize ever, just to keep the BIOS pleased.

This can be the most scrappy card that you can discover, obviously, but there is a requirement: we must make it fit on the main board. The Teslas are large and fill the two PCIe 3.0 x16 slots. The only slots left that can physically hold a card are one PCIe x4 slot and one PCIe x8 slot. See this site for some background on what those names imply. One can not purchase any x8 card, however, because typically even when a GPU is advertised as x8, the real adapter on it might be simply as wide as an x16. Electronically it is an x8, physically it is an x16. That won't deal with this main board, we actually require the small port.

Nvidia Tesla Cooling Fan Kit

As said, the challenge is to find a fan shroud that suits the case. After some browsing, I found this set on Ebay a purchased 2 of them. They came delivered total with a 40mm fan, and it all fits completely.

Be cautioned that they make an awful great deal of noise. You do not wish to keep a computer system with these fans under your desk.

To keep an eye on the temperature level, I whipped up this quick script and put it in a cron job. It occasionally reads out the temperature level on the GPUs and sends that to my Homeassistant server:

In Homeassistant I added a chart to the control panel that shows the values over time:

As one can see, the fans were noisy, but not particularly reliable. 90 degrees is far too hot. I searched the web for a reasonable ceiling but might not discover anything particular. The paperwork on the Nvidia website discusses a temperature of 47 degrees Celsius. But, what they mean by that is the temperature level of the ambient air surrounding the GPU, not the determined value on the chip. You understand, the number that really is reported. Thanks, Nvidia. That was practical.

After some further browsing and reading the opinions of my fellow internet citizens, my guess is that things will be great, provided that we keep it in the lower 70s. But do not estimate me on that.

My first attempt to treat the situation was by setting an optimum to the power usage of the GPUs. According to this Reddit thread, one can decrease the power usage of the cards by 45% at the cost of just 15% of the performance. I attempted it and ... did not discover any distinction at all. I wasn't sure about the drop in efficiency, having only a number of minutes of experience with this configuration at that point, however the temperature level characteristics were certainly unchanged.

And then a light bulb flashed on in my head. You see, prior to the GPU fans, there is a fan in the HP Z440 case. In the picture above, it remains in the right corner, inside the black box. This is a fan that draws air into the case, and I figured this would operate in tandem with the GPU fans that blow air into the Teslas. But this case fan was not spinning at all, due to the fact that the remainder of the computer system did not require any cooling. Looking into the BIOS, I discovered a setting for the minimum idle speed of the case fans. It varied from 0 to 6 stars and was presently set to 0. Putting it at a greater setting did marvels for the temperature level. It also made more noise.

I'll unwillingly admit that the 3rd video card was useful when adjusting the BIOS setting.

MODDIY Main Power Adaptor Cable and Akasa Multifan Adaptor

Fortunately, in some cases things simply work. These 2 products were plug and play. The MODDIY adaptor cable connected the PSU to the main board and CPU power sockets.

I utilized the Akasa to power the GPU fans from a 4-pin Molex. It has the good function that it can power 2 fans with 12V and two with 5V. The latter certainly decreases the speed and therefore the cooling power of the fan. But it likewise minimizes sound. Fiddling a bit with this and the case fan setting, I discovered an appropriate tradeoff in between noise and temperature level. For now a minimum of. Maybe I will need to review this in the summertime.

Some numbers

Inference speed. I collected these numbers by running ollama with the-- verbose flag and asking it five times to write a story and averaging the result:

Performancewise, ollama is configured with:

All designs have the default quantization that ollama will pull for you if you don't specify anything.

Another crucial finding: Terry is without a doubt the most popular name for a tortoise, followed by Turbo and Toby. Harry is a favorite for hares. All LLMs are caring alliteration.

Power usage

Over the days I kept an eye on the power consumption of the workstation:

Note that these numbers were taken with the 140W power cap active.

As one can see, there is another tradeoff to be made. Keeping the model on the card enhances latency, however consumes more power. My current setup is to have actually two models loaded, one for coding, the other for generic text processing, and keep them on the GPU for approximately an hour after last use.

After all that, am I happy that I started this job? Yes, I think I am.

I invested a bit more money than prepared, wifidb.science however I got what I desired: a way of locally running medium-sized models, completely under my own control.

It was an excellent choice to begin with the workstation I currently owned, and see how far I might include that. If I had actually begun with a brand-new device from scratch, it certainly would have cost me more. It would have taken me a lot longer too, as there would have been much more options to select from. I would likewise have actually been extremely lured to follow the hype and purchase the latest and greatest of everything. New and glossy toys are fun. But if I buy something new, I desire it to last for many years. Confidently anticipating where AI will go in 5 years time is impossible today, so having a cheaper maker, that will last at least some while, feels satisfying to me.

I want you all the best on your own AI journey. I'll report back if I discover something new or interesting.