# Is this overkill or viable? (8x gpus, 256gb ddr4) — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/0dd61aeb5a32c863

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-25 00:57 ingested 2026-10-26 02:40

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xeouhq/is_this_overkill_or_viable_8x_gpus_256gb_ddr4/)

## retained text



```
4x 5060ti16gb  
4x 3060 12gb  
256gb ddr4 3200 ram (4-channel speeds)  
On a mobo with 128 pcie4 lanes

I’ve been slowly collecting gpus whenever seeing deals over time and ended up with this. Ended up comparable in cost to current prices of dgx spark

I was thinking:  
TP4 5060s -> pipeline -> TP4 3060s -> offload/overflow less used experts to RAM -> keep any engram on nvme ssd

Is this viable to run inference with big MoE models at high speed and precision, or are cheaper multigpu setups extremely suboptimal compared to unified vram/ram
```



document fetched 2026-10-26 00:38 · canonical https://reddit.com/r/LocalLLaMA/comments/1xeouhq/is_this_overkill_or_viable_8x_gpus_256gb_ddr4
