Dual 7900 XTX Hits 82 tok/s With RDNA3-Optimized llama.cpp Fork
deathcom65 · reddit · 2026-09-18
A Reddit user shared a community-optimized llama.cpp fork (llama.cpp-RDNA3-7900xtx-opt) targeting dual AMD 7900 XTX consumer GPUs.
- Running Qwen 3.8 Q8, decode speed jumped from 28 tokens/s on vanilla llama.cpp + Vulkan to 82 tokens/s at 60k context load — nearly 3x faster.
- The poster says an AI assistant set up the whole Linux environment in one go.
- Shared in hopes that more people push RDNA3 local inference further.
More from Infra
- Fire at Taiwan Panasonic PCB materials plant may tighten CCL supply further — zephyr_z9 · 2026-11-04
- Same Claude model served by Bedrock one day and Anthropic the next, dev finds — georgos_pappas · 2026-11-04
- Self-improving software is inevitable: closing the agent-observability loop — bibryam · 2026-11-04
- Gigabyte's AI TOP motherboard packs 2TB RAM and 9 GPUs for home AI labs — vincent_koc · 2026-11-04
- Building a 7-GPU 132GB VRAM local inference rig: consolidate on EPYC or split across two systems? — infectiousstupidity · 2026-11-04
- imec neural probe integrates photonics directly into 130nm CMOS backend — jwt0625 · 2026-11-04