Testing Flux 2 quants on a 9070 XT: BF16 beats FP8 in speed and quality
Sofa-Sleuth · reddit · 2026-09-25
A ComfyUI newcomer systematically tested Flux.2 9B variants and quantizations on an AMD 9070 XT (16GB VRAM), with wildly inconsistent results:
- Quants (Q4/Q6/Q8): Q6K is visually acceptable and Q8 very good, but all are painfully slow as diffusion models—only useful as text encoders (UDQ8XL helps a lot).
- INT8Conv: as slow as Q8 with regular nodes; faster with custom nodes but quality degrades badly (deformed lines), suggesting ROCm custom-node issues.
- FP8: finally fast and good, but wildly unstable—50 steps of Base sometimes finishes in under 3 minutes, sometimes 20-40 minutes; 2-reference runs range 8-35 minutes.
- BF16 surprise: the 17-18GB BF16 model sometimes runs faster than the 9GB FP8 on a 16GB card, with better proportions and fewer odd shapes—but slows to a crawl or crashes after a few generations, needing restarts. The author now drafts on FP8 and does final renders on BF16.
The post questions AMD's compatibility and memory management: identical tasks vary 10x in runtime depending on precision, nodes and run order, and asks the community for explanations and speedups.
More from Infra
- More Budget 4-GPU Inference Tricks: x8 Splitters and m.2-to-x4 Adapters — TheZachMueller · 2026-09-25
- Skip ECC RAM: ASUS ProArt X870E-Creator Suggested for 4-GPU LLM Inference — TheZachMueller · 2026-09-25
- Google's Antigravity SDK now runs open models like Gemma 4 fully offline on local GPUs — HankYeomans · 2026-09-25
- Running a 27B model locally on 2x RTX 5090 with vLLM — piddlefaffle12 · 2026-09-25
- CLion 2026.2.3 adds NVIDIA CUDA Tile C++ support with dedicated inspections — blelbach · 2026-09-25
- AMD claims edge box runs 2.3x more workloads than Jetson T5000-class hardware — shashib · 2026-09-25