Reddit user finds single R9700 beats every multi-GPU config in local LLM inference tests
SomeoneInHisHouse · reddit · 2026-10-23
A Reddit user tested an R9700 (x16 link) plus RX 7800 XT (x4 link) on an old Ryzen 5800X system for local LLM inference. Across all configurations, both in Strata QFN and llama.cpp Q27B, the single R9700 outperformed every multi-GPU setup. Others report faster multi-GPU results even on x4 links, but he cannot reproduce them and is asking for help.
More from Infra
- After Mozilla compromise, Rust devs eye "remediation speed" as the next hillclimb — charliermarsh · 2026-10-23
- Migrating AI workloads often means 70%+ redevelopment, warns cloud architect — DavidLinthicum · 2026-10-23
- Perplexity Computer now runs fully local on Mac — model and agent both on-device — heyneighbor · 2026-10-23
- Underdog partners with NVIDIA: Saluki 27B claims Opus 4.6 parity in under 8 GB — heyneighbor · 2026-10-23
- Google's Agent Compute Engine gives each Gemini agent a hardware-isolated micro VM — The Decoder · 2026-10-23
- The math on Meta's $825B AI capex: 20GW of compute, but DeepSeek's architecture could rewrite demand — bookwormengr · 2026-10-23