Building my AI Computer
In my experiments to become a better investor, I am constantly experimenting with new AI models, and much of the art in my home results from these learnings turning into hobbies. Although many are content with using cloud-based solutions to learn about AI, I take pride in running most, if not all, of these projects locally on my own hardware (much like how I self-host many of my appliances). Not only is it more fun, but I also find I learn a lot more about implementation than I could by using online tools, and I don't feel guilty wasting credits with increasingly "out there" experiments.
My self-built gaming PC can run inference on it, but despite my upgrades, the GTX 980ti doesn't have enough RAM to train or fine-tune modern models on, and the 4790k’s four cores are a limiting factor for CPU-based training. Although this PC is a fantastic workstation for working from home (its primary function, especially with NVidia Broadcast) and general AI inference, it is not well suited for my new AI training work. I needed to build a new computer, and this time, I wanted it to be a beautiful work of art, as well as a powerful AI machine.
Regarding specs, I wanted everything to be top-of-the-line.
Corsair Dominator RAM
RTX 4090 GPU
Ryzen 7950X CPU
8TB nvme gen 4 SSD
This is the best (consumer-grade) hardware one can buy for AI workloads and the best chance of long-term usability as models continue to evolve and grow.
I also wanted this computer to be beautiful. Whereas my previous build in the large Corsair 900D case gave an impression of “understated overkill,” I wanted this build to give a sense of elegance and precision tuning. That meant keeping with custom hard-line water-cooling, but this time, I didn’t want any bends: Every pipe should be parallel and perfectly straight. This additional constraint meant that I needed a distribution plate (rather than a traditional reservoir) and many more fittings (including some small 5mm extenders) so that the water lines could be made parallel.
Although I considered the Corsair 1000D, the spiritual successor to the 900D from my previous build, I ultimately purchased the Lian Li 011 Dynamic Evo for its cleaner style despite having slightly lower cooling capacity and a smaller size. This may be the one major mistake of the build.
I decided on a full cover plate for the motherboard, which simultaneously cools the CPU, VRMs, chipset, and SSD. Although water-cooling the VRM, chipset, and SSD is not strictly necessary, I believe it may help extend the lifespan of these components by reducing the severity of thermal cycling, and it gives the PC a much cleaner look.
Similarly, I chose to use EK’s new dual-sided full-cover water blocks for my GPU. These water blocks not only cool the primary heat-generating components on the front of the GPU but also provide active cooling to the back side of the GPU, where some power control components are. Although I had considered mounting the GPU vertically to give a better view of the front-side cooling of the GPU, this would, in turn, block the view of the full cover plate on the motherboard. The dual-sided GPU cooling, however, still provides an interesting view of the GPU even in standard (horizontal) configuration.
I assembled a barebones system and conducted my first test to “POST” the system with the GPU factory cooler and (empty) full cover plate installed. I would not be able to run for long in this configuration due to the lack of heat dissipation from the CPU, but it would be a good test to make sure all the hardware was working. The last thing I wanted was to discover that one component was mysteriously broken in a fully assembled computer. With that test successfully completed, it was time actually to build the PC in the case.
Installing the water block on the GPU was the first step in building. This involves carefully removing the existing heat sink (saving the associated screws and recording their positions in case I need to reinstall it at some future date). Once the heat sink was removed, I had to apply a new thermal compound to the GPU chip and thermal adhesive pads to the memory and power chips. I then carefully aligned the front and back plates and screwed them together, being careful to tighten them in an X pattern to avoid applying too much pressure at any given spot.
Installing the distribution plate to the front of the case was simple, but after years of working in a super-tower case, adding the remaining water-cooling components proved difficult and cramped in a mid-tower. The upper radiator proved particularly challenging, requiring multiple carefully measured fittings to align it to the inlets and outlets on the distribution plate. Furthermore, installing the upper radiator required tightening and screwing two fittings simultaneously. This simultaneously created a water channel and mechanically held the radiator in place (which would be later reinforced after screwing into the case). Many of the pins and connectors on the motherboard were difficult to reach or inaccessible once water blocks were in place, so it was important to do everything right the first time.
The first full leak test resulted in a very unpleasant shock. As I filled the reservoir and pumped the water through the blocks, I noticed a steady stream of water dripping from the edge of the GPU block. Terrified of potential damage to the card, I immediately drained the computer and started disassembly.
I noticed the leak was not coming from my fittings (loose fittings are the usual suspect in leaks), but from the dual-sided water block. I suspected that the leak was not coming from the block itself but from the new “Direct Link” connection system that interconnects the front and back blocks. I needed to disassemble and re-build.
Disassembly and removal of the GPU proved much harder than anticipated because reaching the PCIe locking mechanism was nearly impossible. The thickness of the actively cooled back-plate blocked access to the PCIe lock with bare hands, and the full cover motherboard block was so tall and so wide that it was nearly impossible to angle a pen or other long tool to reach the mechanism. Fortunately, after around 90 minutes of frustrated trying, I was finally able to unlatch the lock and remove the GPU. Disassembly of the GPU coolers found that the factory-installed o-ring was misaligned, getting smashed and deformed by the block, and the likely source of the leak.
Fortunately, the blocks came with a spare o-ring, so after replacing and reassembling, the computer passed its leak test, and it was time to boot properly. POST passed successfully, and I could install the OS and begin working on this computer.
I built this computer as a next-generation AI machine, and it is exceedingly capable in that role. I plan on dual booting in Windows and Ubuntu to do my work and have already succeeded in fine-tuning concepts in Stable Diffusion (I will share more on these experiments later). The PC looks beautiful, but I wish the case was larger. Most of the problems I encountered with this build were related to the size of the case and how cramped that makes building in it. Furthermore, it makes it much harder to upgrade. I cannot add a second GPU because there is not enough room for it in the case (rather than any limitation on the hardware itself). Installing a second nvme SSD would require a complete rebuild, as the nvme slot is underneath the motherboard monoblock, and
I cannot reach that without disassembling the vast majority of the PC. The only “easy” upgrades I can do are adding two additional RAM sticks, two HDDs for mass storage, and a few SATA SSDs. Given how massively overbuilt this system is, this is not a huge issue, but I would have preferred the extra flexibility since it is difficult to predict future needs. I expect my next build, next decade, will be in a much larger super-tower case. Until then, I cannot wait to share all the interesting AI projects I will build and test.