What motherboard and CPU make sense for a four-RTX 3090 Ti local LLM server?

0
8
Asked By MellowOrbit42 On

I'm putting together a dedicated local LLM server with four RTX 3090 Ti cards. I already have a 1000W Platinum PSU and plan to add a second PSU for the remaining GPUs. The system will be built in a Phanteks Enthoo Elite Server case. I originally considered a Threadripper Pro 3945WX with a WRX80 motherboard, but those boards have become extremely expensive. Another suggestion was an AMD EPYC 7452 with a ROMED8-2T board. I'm looking for practical advice on CPU, motherboard, memory, PCIe layout, and power requirements. I currently own one 3090 Ti and am moving it from my gaming PC into this server; cost is flexible as long as the parts are reasonably justified.

2 Answers

Answered By ServerSparrow19 On

EPYC can provide a lot of PCIe lanes and proper eight-channel memory, but server platforms tend to be more restrictive about supported RAM and can make upgrades expensive. Threadripper Pro is generally easier to live with as a workstation-style build, while EPYC makes more sense if you specifically need its lanes, memory bandwidth, or server features. Since this will be a separate dedicated server, the usual server-platform drawbacks may not bother you. Before choosing, confirm the board’s slot spacing and whether your very thick 3090 Ti cards can be installed with risers. Power is the bigger issue: four cards at their stock limits could approach 1,800W before accounting for the CPU and the rest of the system, so power limiting is likely essential.

MellowOrbit42 -

I already have one 3090 Ti and a 1000W PSU, and I’m moving both into this dedicated build. My gaming system will use an 850W supply with a separate RTX 5080, so this server can be designed around the GPUs and won’t need to function like a normal desktop.

Answered By CopperViolet7 On

A proven configuration is the Gigabyte MC62-G40 with a Threadripper Pro 3945WX, 192GB of DDR4, powered riser cables, and a 1000W plus 1500W PSU setup. The four-GPU system works well for local inference when the cards are power-limited to around 250W each. I’d use a high-quality PSU synchronizer, make sure every GPU has the required PCIe power cables, and consider a single 1600W unit if your electrical setup supports it. The 3945WX also appears to effectively operate as a four-channel memory platform, despite the platform offering eight channels, so a higher-end CPU may be worthwhile if memory bandwidth matters. With four 3090-class cards, the GPUs usually dominate performance anyway.

MellowOrbit42 -

That’s very helpful. I’m using 3090 Ti Zotac Extreme Amp cards, which are about 3.5 slots wide and can draw roughly 450W each. I’m concerned about both physical clearance and whether I’ll need something closer to a 2000W supply unless I limit them heavily. I’d also like to know how useful the 192GB of RAM is for your workloads and which quantization you use for DeepSeek.

Related Questions

LEAVE A REPLY

Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.