Electricity per AI image measured across RTX GPUs: ~3-10 cents per 1,000 images with FLUX.2 klein
strata2signal · reddit · 2026-11-04
strata2signal measured board electricity per 1,000 images (FLUX.2 [klein] 4B, 1024x1024, 8 steps, fixed ComfyUI stack, at 18.31¢/kWh):
- RTX 5090 Laptop 24GB: 3.37¢ (4.48 s/img)
- RTX PRO 6000 Blackwell 96GB @300W: 3.39¢ (2.30 s/img)
- RTX 5080 16GB: 5.04¢; RTX 3090 Ti: 8.47¢; RTX 3090: 9.13¢; RTX 3080 Ti 12GB: 9.71¢
Key levers: 8→32 steps costs 3.5-3.9x; 512²→1024² costs 3.0-3.3x; Z-Image-Turbo runs 1.5-1.6x klein and FLUX.1 [dev] fp8 2.1-2.5x. Cheapest measured: 0.79¢/1,000 (512², 8 steps, laptop GPU capped at 95W); dearest: 150¢/1,000 (Z-Image-Turbo 2048x1280, 32 steps on a 3080 Ti with PCIe streaming). Capping 3090 Ti/3090 at 300W saves 9% energy for 4.5% more time. Wall power reads 1.47x board power in the one dual measurement.
More from Infra
- Marvell to Acquire Lumentum for AI Optical Connectivity, Closing Expected H1 2028 — BenBajarin · 2026-11-04
- Analyst: AMD now best positioned for server CPU growth as market heads to $80-100B by 2027 — BenBajarin · 2026-11-04
- AMD data center revenue doubles to $8.5B as Anthropic, OpenAI, Meta run models on first Helios racks — AccBalanced · 2026-11-04
- Nina Schick: we are industrializing superintelligence — intelligence-per-watt is the new power law — NinaDSchick · 2026-11-04
- Cerebras CTO: our plate-sized chip drew more power than an entire 5 kW rack — AccBalanced · 2026-11-04
- AMD guides 67% revenue growth but Nvidia's growth floor next year is 70% — BenBajarin · 2026-11-04