Planning build: Power efficient headless steam machine, and later upgrade for AI tasks
Planning build: Power efficient headless steam machine, and later upgrade for AI tasks
Background
I'm planning on building a secondary server that can process more intense tasks than my current basic home server. Tasks such as light gaming (think "remote Steam Deck"), and later allowed to be upgraded with a Nvidia graphics card to handle AI tasks, such as LLM and SD.
The problem
While I have no problem picking parts to build this as a "desktop computer", I'm completely lost when trying to make it power efficient for idle load (if it's even possible with a power-hungry Nvidia card). I'd appreciate some guidance even if it's not a full parts list suggestion!
Watching Wolfgang's videos has unfortunately not translated knowledge into practice for me yet. At least I know TDP isn't an absolute determining factor anymore.
Planning the build
Due to a limited budget, the idea is split the build in 2 phases.
Phase 1 (gaming):
- Budget: $1000 (ideally below ~$800):
- Use for local headless gaming (with bazzite?)
- At least as powerful as a Steam Deck
- Parts:
- APU:
- Perhaps: "AMD Ryzen 7 7800X3D ($400)"
- or "AMD Ryzen 5 7600X ($300)"?
- Motherboard:
- No specific requirement, will mainly just use the PCIe x16 slot for a single GPU when upgrading in "phase 2".
- Okay with gigabit ethernet and basic I/O.
- Power supply:
- Power efficient power supply that can handle a class 4080/4090 card.
- Cooling:
- Air-cooled preferred
- Storage:
- Samsung 990 PRO NVMe M.2 SSD 1TB (~$100)
- RAM:
- Corsair Vengeance LPX Black DDR4 3200MHz 2x16GB (~$70)
- Case:
- As long as it fits a large graphics card
- APU:
Phase 2 (AI: LLM/SD):
- Budget: ~$2000
- After 1-2 years, upgrade with Nvidia graphics card.
- Ideally something with 24GB VRAM, like the 4090.
- Prefer Nvidia due to compatibility with SD.
- Open for suggestion, since wanting a low power draw with a 4090 might sound contradictory.
In case I missed any crucial information, let me know!
You can probably ignore the idle power consomuption and enable Wake On LAN and turn it on and of over the network. Is that a solution for you?
one of those pi kvms or the like could turn on any system even if it doesn't support wake on lan
Might be the case but is this really a problem to solve? To my knowledge, all modern boards support wol.
With all the helpful comments shared in this thread, I'm starting to realize that this approach is likely the only viable solution.
Previously when doing my research, I was naive enough that when people said "...30W at idle", it was specifically for their GPU, and not for their whole system. So now things makes a lot more sense.
Not much you can select for with desktop parts. Maybe get dual Ethernet now so you don't want to add a card later. And more disks, more power so one bigger drive is better than two smaller..
Might be better to suspend it, and wake on lan when you want to play.
Just so you know
R5 5700G
128GB ram
2 x 20TB disks
2 x 4tb NVMe
Idles at about 60w.
You should look into the RTX 6000 or higher to utilize vGPUs.
Not fully supported but it's possible.
And also get into proxmox. You can pass through part of your GPU into a "desktop" environment and also have another VM(s) running in the background. That way you can use your computer as normal with a type 1 hypervisor in the background.
Also get a mobo with 2 NICs. The fewer pcie cards you have the lower power draw.
My NVMe idle at 7w and my HDDs idle at about 15w I think. 45w is just for storage.
Some tips here:
- get a platinum rated power supply, if you can afford it go for a titanium. The efficiency in the power supply is half of the efficiency of the rig
- reduce the number of the modules to the minimum
- get a platinum rated power supply ;)
- get big passive coolers, you want to idle the fans
- reduce the number of usb and connectors to the minimum. Their converters are not the most efficient. Try not to connect enything on them.
- NO mechanical parts (including fans or water coolers)
- set schedulers to conservative or power efficient. You don't want to spike the power just because a task is 2ms longer than expected.
- pick a power efficient CPU/gpu (I think we can discard this one based in your choices)
- use the latest amd adaptative undervoltage technology to ensure to reduce the wattage of the cores
- try to reduce to the bareminimum the number of background tasks /services running.
And that's all. Sometimes there is a component of trial and error because sometimes the curve performance / power is not entirely linear and you don't want to hit exponential-non-linear zone.
Good luck and if you can post you build with numbers and some lessons learnt would be great
Good luck
Just to add my experience about PSU efficiency: for low power consumption (20-50W) you need PSU rated for minimum power your system needs. So if you are idling at 30W on 700W PSU your efficiency will be super bad because that PSU was made for higher loads and you are using <10%. No matter what PSU class you choose, efficiency will be better if your usage is at 40-70% of PSU max power. This is based on testing multiple desktop ATX PSUs for my small homelab
Definitely.
I forgot to add that it would be necessary not to overdimension the set up. Any extra power is something that needs to be powered.
But with the chosen cpu and GPU there is not a lot of room here.
Maybe add one of those dummy HDMI or Display dongles so you don't need to connect a monitor and you can set the display resolution who whatever you want.
Moonlight + Sunshine is best for headless gaming. It also has desktop as an app.