I'm looking to build a high-end pc to be an all-in-one general build (gaming, productivity, daily driver, LLM/AI tinkering). How does the list below seem for a high-end build? Is there anything I've forgotten? Haven't thought of? Am un aware of? Basically I'm asking for a sanity check or for any additional insight I should be aware of before I commit to purchasing.
Now for a bit of a rambling of my ideas. I'm not sure how much response or discussion I'll get from any of this, but at least the process of writing everything down has been helpful. :shrug:
Goals:
Migrate from Windows 10 to Linux
Game in Linux using Steam and Proton as much as possible
Use a Windows in a Virtual Machine (with VFIO) in Linux, when necessary
Experiment with LLMs and other AI things, possibly one day adding a used 3090 if I decided I want to dive deeper into LLMs/AI
Case has room for a second massive GPU
PSU calculations are showing that adding a 3090 would put me around 1285W usage
I'd probably limit the power draw on the 3090 too, at the cost of a bit of its performance, but I'll still need a very beefy PSU when I do get around to using dual GPUs
Case/Aesthetics:
Going for a white/black mix color aesthetic
Going with some budget ARCTIC case fans (not listed above) until Noctua releases their new 140mm fan (sometime within the next year, I hope?)
Intel vs AMD, and RAM:
Intel and AMD seem to be pretty much neck and neck with their current generations, so it was kind of a tough choice
I decided on Intel
I do a bit of everything, and not just gaming, otherwise I might've picked the 7950X3D or even the 7800X3D (wow! what an amazing gaming processor!)
I was ready to pull the trigger on the i9-13900k ... but I just saw that the 14th gen Raptor Lake-S refresh is dropping next month, so I guess I'm waiting for that now.
Supposedly the current line of motherboards (z790) will need a BIOS update, but should be able to use the 14th gen processors
I opted for faster/denser RAM
I'll be messing around with LLMs/AI, and I know (llama.cpp) wants as much RAM as possible, but I ended up settling for 2x48 GB
I've heard that 12th/13th Intel generation has issues running DDR5 at high speeds when running heavy RAM configurations
AMD can't run much faster than 6000 and/or also has issues with RAM amount and speed
Memory controllers are antagonized by both dual-rank DIMMs and number of slotted DIMMs, basically more RAM means more problems and the speed has to drop to maintain stability
I'm pretty sure the 48GB DIMMs are dual-rank, but the tradeoff is I'm keeping it in two slots
If I find myself desperate for more RAM, then I'll up to 192GB and drop the speed to whatever it has to go to
GPU and Motherboard:
I've read that Gigabyte GPUs have little to no coil whine, which was a major factor in selecting their hardware, in addition to the aesthetic pairing
...I've also read that Gigabyte has terrible customer support and a horrible RMA process, so I think I'll be purchasing an insurance plan for the GPU through B&H. :unamused:
I'm hoping a combination of a high-quality GPU, high-quality Motherboard, and high-quality PSU will prevent any coil whine from occurring
Also, hoping to avoid any buzz in my head phones -- I seem to have stumbled upon a ground loop or something with my current setup and the buzzing increases in intensity when my GPU increases its power consumption :frown:
Worst case scenario, I buy a dedicated sound card or SPDIF DAC or something
The Z790 Aero motherboard has 1 PCIe 5.0 x16 channel (which I don't plan on splitting with the SSD) and 2 PCIe 4.0 x4 channels (in x16 slots)
The 4090 can take the main slot,
When I slot a second GPU, it'll take the x4 slot. I'll take the performance hit on the bandwidth, which should be okay since the LLM workload is split and contained within separate GPUs and there isn't any transfer between the two
NVLink isn't necessary AFAIK, but the AI field seems so crazy right now that it'll probably change in the future
PSU:
Most troublesome part to choose, I want something that was future proof with the following features
Dual native GPU support via 12VHPWR (12+4 pin) adapters (even though I'm tentatively planning on a 4090 + 3090 combo)
ATX 3.0 support
Is ATX 3.0 support required at this time? Probably not, but I wanted to future proof if the chance provided itself
At least 1300W, but preferably 1600W
I've been a SeaSonic stan for ages and I saw they had a 1600W PSU, the Prime ATX3.0 PX-1600 Platinum (SSR-1600PD2), but...
The PX-1600 nearly costs as much as the CPU, and the Titanium TX-1600 variant cost more than the CPU. I can't consciously spend more on a PSU than a top of the line CPU :frown:
...maybe this 12V-2x6 pin terminal/cable is part of the reason why there aren't many ATX 3.0 PSUs yet?
So there aren't many PSUs that meet my criteria and I decided to settle on the bare minimum until I need a dual-GPU PSU and have set a rigid $200 price limit, which these PSUs should I choose?
Gaming on Linux with a NVIDIA GPU is bad. You'll be limited to shitty drivers. I recommend using an AMD card if you truly want to game on Linux at full potential.
The PSU could use a little more overhead. I saw you are looking for 1300-1600 watts but you still have a 1000 W gold PSU in there.
If you're adding a 3090 for LLM stuff, get a 1600 W PSU minimum so you have enough overhead.
Yeah, the PSU is a holdover until something a more favorable 1600W PSU appears as an option. I don't really like buying a stop gap, but I don't even have a spare 1000W PSU laying around.
AMD cards didn't really have much of a chance when I'm planning on working with AI stuff, though. I've read they're working on some things to make them more competitive, but without CUDA they've always been lagging behind in performance, implementation, and favor.
I'll try researching shitty driver issues and see what sort of pain I'm in for.
I agree. Nvidia is the easy choice if you're interested in ML, regardless of driver issues. A lot of projects do support non-CUDA backends these days but even so, Nvidia performs better. Choose your distro carefully. I strongly recommend an LTS distro that is officially supported by Nvidia.
That said, if you're dealing with medium-sized LLMs leveraging that 96GB of RAM, that means you'll be running them on the CPU anyway. 24GB of VRAM will only support smaller LLMs. Consumer hardware can't run the largest networks at all (e.g. Bloom needs 360GB, and consumer mobos typically only support up to 192GB). Just something to keep in mind.
I have a similar build for similar reasons. It works great, though I use Windows, so no driver issues (VR introduces too much jank with Linux). Notes below.
CUDA is essential. Definitely the right call paying the Nvidia tax.
My Gigabyte 4090 works for LLM stuff without a second card, and has no coil whine I can hear. I use Alpaca 4-bit entirely in VRAM, and SDXL runs like a dream. I only have 48GB of RAM total, but VRAM is pretty much always the limiting factor (if I understand correctly, it works best when you have at least enough spare RAM as you have VRAM when you're loading the model, but after that the computation is on the GPU if you have a 4090. Moving layers to the CPU/RAM drops performance fast). I have an A4000 in another machine that I was planning to add with a riser cable, and I just haven't bothered because I didn't end up needing it.
Leaving the upgrade path open is a solid choice. The space is so volatile that it's impossible to predict what the requirements will be like in six months. They could even go down like they did when 4-bit happened.
I use an external DAC, so can't speak to the whine there. They're not that expensive though.
are you fully committing to the white aesthetic? cause black fans cpu cooler and power supply. Anyway id remove the extra thermal paste because you dont need it. i'd switch the power supply too because it has bad coil whine
I was aiming towards a primarily white aesthetic with black accents since the motherboard has both colors and the heatsink is black. Something like this build here, though I realize that particular build has a lot of help balancing the colors because they vinyl wrapped the backplate of the case. The PSU doesn't matter because it's hidden behind the backplate out of sight.
Thermal paste was an interesting bit of research -- I was going to go with Kryonaut, like with my last build, but I read that it starts to break down at 85C and is susceptible to being heat pumped off the processor (which is a concern since this generation of high end processors run at their thermal limit of 95C if your heatsink isn't able to keep up with it). I found some people saying that the Thermalright TFX paste was very resistant to heat pumping because of its make up, which is to say that it's thick as mud and a pain to apply because of it. I figure I configure a small script+cron job to log the temps to a file for a year and keep an eye on the temps, if they start to drift upward, I get something new and reapply.
The Corsair has bad coil whine? Ugh. I hadn't thought to do coil whine research on PSUs. I'll have to do that now. Have you heard anything about Super Flower PSUs? I had already done a bit more research previously and couldn't decide between that Corsair or this Super Flower LEADEX VII XG 1300 W. Supposedly, Super Flower is similar to SeaSonic, so I could go with that.