Intel Arc B580 runs INT8 ConvRot acceleration, Flux 2 Klein 9B in 3.9s
Valuable-Subject-274 · reddit · 2026-09-24
A Reddit user got INT8 ConvRot acceleration working on an Intel Arc B580 using Intel LLM Scaler inside ComfyUI, with the integration itself done by Codex: they handed it the path to their portable ComfyUI build and asked it to wire in LLM Scaler.
Post-warmup average generation times:
- Krea 2 INT8 ConvRot: 768×768 in 10.27s, 1024×1024 in 15.71s, 1920×1088 in 30.21s
- Flux 2 Klein 9B INT8 ConvRot: 768×768 in 3.90s, 1024×1024 in 6.11s, 1920×1088 in 12.58s
The author reports stable generation and notably fast results with Flux 2 Klein 9B — a useful reference for running local image generation on budget consumer GPUs.
More from Infra
- Ant Group open-sources MECT voiceprint model: 9.57M params rivals 587M at 100ms latency — aigclink · 2026-09-24
- Chose rustpython-parser over tree-sitter after measuring both on real LLM output — SprayPuzzleheaded533 · 2026-09-24
- More compute made one agent 13x faster, another just 4%: agents' bottlenecks are task-dependent — alex_verem · 2026-09-24
- beffjezos claims thermodynamic compute will hit 1 billion pbits next year, mocking quantum's ~100 qubits in a decade — mjdramstead · 2026-09-24
- DeepSeek tests running smaller models on cheap gaming GPUs, per The Information — rohanpaul_ai · 2026-09-24
- NVIDIA DGX Spark sold out at retail as the last Micro Center units go — TheZachMueller · 2026-09-24