← sources

Advice on a new desktop build to take advantage of my existing R9700 cards (x2)

source: reddit (reddit:LocalLLaMA)published 2026-09-18 15:54ingested 2026-09-19 02:42

retained text

I started out my LLM / AI journey before I really knew what I was getting into.  When I fell in love with the tech, I grabbed an R9700 AI PRO card (32gb vram) and then found one a month later as an open box special, which I grabbed.  This was several months ago now.

**The issue:**  I'm currently using these two cards in a board that doesn't have proper PCI-E lane splitting.  I also don't have enough RAM (32gb DDR5).  CPU is a Ryzen 5 7600, so not great but paired with the proper chipset and mobo, it should be able to handle PCIE-8x / 8x.  However, if upgrading the CPU would help with some layers offloaded to CPU-RAM MoE model performance, I'd consider it.

I'm not willing to go threadripper; it's just too damn expensive.  But I might have an opportunity to switch out my motherboard and case, upgrade RAM and keep everything else.  I know RAM is crazy expensive right now, but I don't see any price drops in the near future; just increases.

**My thought:**  Upgrade motherboard to ASUS ProArt X870E-CREATOR WiFi AMD AM5 X870E or something equivalent by Asrock, etc.  Upgrade RAM to 128gb DDR5 (the biggest cost increase, might not be worth it; may just go with 64gb RAM in a two-dimm config).

My worry is that this won't be a huge step up from my current setup.  Right now, I'm left struggling with vulkan and tensor splitting which slows down my pre-fill and generation speeds.  P2P / tensor parallelism is possible with a custom fork of llama-cpp but I have mixed results and many issues with it and it's not true PCI-E lane 8x lane splitting.

Advice would be most welcomed.

document fetched 2026-09-19 00:40 · canonical https://reddit.com/r/LocalLLaMA/comments/1wju218/advice_on_a_new_desktop_build_to_take_advantage