Need some tips/ideas on getting the most out of my gaming systems. Turns out the old “junkers” I didn’t want to sell have come back to be useful. I’m a software dev, but doing this as a hobbyist writing a game in Godot, more to learn about LLMs than to make a game. I learn best with a project.
Hardware:
64GB 6000Mhz 9800X3D + RTX 5090
32GB 3600Mhz 5900x + RTX 4090
32GB 3200Mhz 3900 + RTX 2080ti
Right now I’d call my configuration basic, running Qwen 3.8 27b nvfp4 for 5090 (200k Context, full KV with MTPx4 & Vision, xhigh) [Bazzite + distro container for llama.cpp]. I’m trying to move my gaming to Linux. It’s the workhorse. The 4090 Q4_K_M (256k KV q_8 mid [not sure its active though]), also my Dev IDE: Godot, Deepseek Harness (used to be OpenCode), VSCode and random python harnesses (basic stuff like RAG and other AI type side projects).
The 2080ti is unused. But feel I’m leaving compute on the table, wondering if there is anything productive it could be doing? Should I swap it to be my dev box, then use it as a small code predictor LLM. 4090 becoming dedicated LLM?
My perceived issue is a single threaded methodology, one problem at a time either the 5090 is running or the 4090 is running not both. No idea how to go about getting a multi agent setup to have the both working well together, while still feeling like I’m in control. Maybe that is something I need to give up for better utilization.
Should I switch up the harness? I’m convinced 50% of good AI output is down to a good harness. But not sure what would be better to unlock parallel tasking.
Nice problem to have re: your current gpu inventory.
If I had that hardware, I would build a linux system around the 9800X3D with both the 5090 and 4090 in the same box. Then I would run llama-server and experiment with tensor splits that load up the 5090 as much as possible with qwen3.8-27B. I’m running a 6-bit quant of that model with three older gpus (40GB combined vram) and 120K of context, plus the image processing model GGUF. I use the pi coding agent with no extensions installed.
I personally don’t bother with running parallel agents, but have experimented with sub-agents. Meaning, I use an orchestrator agent that takes a whole problem, creates a multi-step plan, hands off steps of the plan to an implementor agent, then runs a reviewer/tester agent, etc. The primary idea was to keep context requirements low for the various agents given my relative lack of vram. Honestly, I have found with pi and a good coding spec/design that just letting the top-level prompt do all the work seems to work great. Qwen3.8-27B seems very good.
The 2080ti is kind of interesting. One of my three gpus is a 2070 Super. It’s not bad with my two 16GB Ada-gen cards. I actually think other than the 8GB limitation (which is a hassle), it has enough memory bandwidth to be useful. But I have a 44 lane CPU, so I can run the 2070 PCIe3x16 and the two Ada cards PCIe3x8 (they are x8 cards). I’m guessing the 9800X3D could handle all three at PCIe x16, but I don’t know. So if my power supply could handle it, I’d have all three of your cards in the same box, but I think the 5090 and 4090 need a bunch of power . . .