RTX 5090 SD1.5 88 it/s SDXL 19 it/s Native FP16 with FP16 accumulation (Eager Mode)

[GPUGrandmaster]

Hi all, sharing the raw performance of RTX 5090 on Forge with custom extension.

Performance Video:
https://youtu.be/XEz8PuqKc-g
Video Demonstrations:

• Standard Benchmarks: Raw throughput testing for SD1.5 and SDXL.
• ControlNet Integration: High-performance generation with active ControlNet layers.
• Rapid Dynamic Resizing: Demonstrating seamless switching between different image dimensions without the typical "warm-up" or recompilation delays.

Full environment screenshot (GPU-Z + Console) attached below.
SD15 1x512x512x20

SD15 8x512x512x20

SDXL 1x1024x1024x20

SDXL 4x1024x1024x20

Specs & Metrics:
Precision: Pure FP16 with FP16 accumulation (No FP8/INT8/Quantization)
Mode: Native Eager Mode (Dynamic Model, Dynamic Dimension. No Compile, No Static CUDA Graphs)
Throughput: SDXL (1024x1024) @ 19.1 it/s | SD1.5 (512x512) @ 88.4 it/s

Verifiable Binary:
https://huggingface.co/gpugrandmaster/OTKN-Engine

GPU Grandmaster

#GPU #NVidia #RTX5090 #Benchmark #SD1.5 #SDXL #Performance #AI