← sources

Upgrade advice: 2× RTX 3090 + 512GB DDR5 UDIMM—can I reach 10 t/s with large models?

source: reddit (reddit:LocalLLaMA)published 2026-09-10 17:57ingested 2026-09-10 22:44

retained text

I’m looking for advice on building a better local AI system while reusing as much hardware as possible.

**What I already own:**

* 8×64GB Crucial DDR5-5600 **UDIMMs**, purchased as individual sticks—512GB total ([CT64G56C46U5](https://www.crucial.com/memory/ddr5/ct64g56c46u5)).
* Two systems, each with an Intel Core Ultra 7 265K and MSI MAG Z890 Tomahawk WiFi motherboard.
* 2× RTX 3090 in the AI system—48GB total VRAM. The other system handles homelab duties.

The largest model I’ve run is [DeepSeek-R1-0528 Q3](https://huggingface.co/ubergarm/DeepSeek-R1-0528-GGUF), inspired by [this Level1Techs video](https://www.youtube.com/watch?v=KQDpE2SLzbA&t=167s). I got roughly **2 tokens/sec**.

My goal is **at least 15 tokens/sec during generation** with a **"**very large model", potentially using much of my available 512GB RAM. I realize capacity in this context hurts speed.

Before saving up for a platform upgrade, I’d appreciate advice on:

1. Could Threadripper, EPYC, Xeon, or a similar platform realistically achieve this with my two 3090s?
2. Which CPU/motherboard combinations should I consider?

I have no enterprise hardware experience or firm budget yet—I’m trying to establish what’s feasible and what I should save toward. Specific hardware suggestions and firsthand benchmarks are helpful.

document fetched 2026-09-11 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1wcqljl/upgrade_advice_2_rtx_3090_512gb_ddr5_udimmcan_i