2x AMD R9700 beats 3x running Qwen 3.8 Flash Next locally at 35 t/s
MarcusAurelius68 · reddit · 2026-09-05
A Reddit user benchmarked Qwen 3.8 Flash Next on AMD R9700 32GB GPUs and found that a 2-GPU setup actually beats 3 GPUs. The culprit: in the 3-GPU config on an X570 board, the third card sits on a chipset x4 slow lane that drags down overall performance.
Key details:
- Hardware: X570 at x8/x8, Ryzen 9 5900XT, 64GB DDR4, Windows 11.
- Running the AD-4.27bpw Q4KM quant (33-part GGUF, 131k context) with tuning (no MTP yet), the 2-GPU setup hit 242 t/s prompt processing and 35 t/s generation.
- The full llama-cli command is shared, including --fa on, q80 KV cache quantization, --jinja, and sampling settings.
A directly reproducible reference for anyone running large-context local models on consumer hardware.
More from Infra
- PostgreSQL 19 slips to November as LLM-assisted code review floods committers with edge cases — sanjaykalra · 2026-10-21
- Decision Models' routing classifier hits AntSeed, cutting inference cost to 24% of always-Opus — airesearch12 · 2026-10-21
- Eon cuts engineer onboarding from 2 days to 2 minutes with an MCP Gateway on Bedrock AgentCore — AWS ML Blog · 2026-10-20
- a16z's Ben Horowitz says AI buildout isn't the 90s bubble — an investor pushes back — henloitsjoyce · 2026-10-20
- Micron and Nvidia drive ~36% of S&P 500's Q3 earnings growth — Beth_Kindig · 2026-10-20
- NVIDIA lands in Europe: Jensen Huang and German chancellor discuss industrial AI at GTC Berlin — nvidia · 2026-10-20