Reddit asks: building a 64GB VRAM local inference rig on a £7k budget
Repulsive-Juice6676 · reddit · 2026-10-05
A Reddit user with a £6-7k budget is building a PC from scratch for local coding models, liking DeepSeek V4.1 Flash and Qwen3.8 27B and needing reasonable tok/s. They want 64GB VRAM but suspect it's a stretch without dual R9700s or Intel equivalents, and are unsure whether those setups actually work well, so they're asking for build guidance.
More from Infra
- 691K H100-hours: community estimates compute behind from-scratch checkpoint trained on just 320 H100s — teortaxesTex · 2026-10-06
- Dev reports Clef-Flash runs fast locally even on memory-bandwidth-limited Jetson Orin — gregmushen · 2026-10-06
- Running LLMs Fully Client-Side: llama.cpp Compiled to WASM with WebGPU Proof of Concept — Numerous-Fan8138 · 2026-10-06
- Llama.wasm: llama.cpp Compiled to WASM with WebGPU Runs LLMs Fully in the Browser — Numerous-Fan8138 · 2026-10-06
- NVIDIA Dynamo lets coding agents point at self-hosted endpoints with native tracing — TheZachMueller · 2026-10-06
- Schmidhuber: compute gets 10x cheaper every 5 years, 100,000x in 25 years — SchmidhuberAI · 2026-10-06