FLUX 3 Klein INT8 tested on RTX 3090/5090: 2x speedup, better accuracy than FP8
CeFurkan · reddit · 2026-10-17
CeFurkan benchmarked FLUX 3 Klein's ConvRot Int8 HQ quantization on an RTX 3090 and 5090 in one PC, with ComfyUI screenshots:
- Speed: 3090 text-to-image 0.384s → 0.175s (2.19×), image edit 2.29×; 5090 113ms → 52ms (2.18×), consistently over 2x
- Accuracy: 2.14% deviation from BF16, beating official INT8 same-attention (2.26%) and default (3.04%); FP8 Scaled was slower than BF16
More from Infra
- Custom VBIOS loader lets you edit RTX 50/40/30 power limits and voltages without flashing — panchovix · 2026-10-17
- Perplexity Flash matches Wald's accuracy at nearly 2x speed in local agent-decision benchmark — WolframRvnwlf · 2026-10-17
- Mac LLM Calculator tells you which Apple Silicon runs which model, and how fast — shauntrennery · 2026-10-17
- llama.cpp PR #30517 Delivers Big CUDA Speedup for Low-Bit Q2_K/Q3_K Quantization — jacek2023 · 2026-10-17
- India has only ~30,000 GPUs nationwide, 10-20x short of frontier model needs — CatAstro_Piyush · 2026-10-17
- MiniCPM-V 4.7 runs locally at 100 tok/s using only ~3 GB of memory — NielsRogge · 2026-10-17