comfyui-rtx4090-nodes

ComfyUI RTX 4090 Nodes

High-performance optimization nodes specifically tuned for NVIDIA RTX 4090 GPUs.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/joe002/comfyui-rtx4090-nodes.git

6 Nodes

| Node | Purpose | |------|---------| | BatchImageProcessor | Process multiple images efficiently | | BatchLatentProcessor | Batch latent space operations | | RTX4090Optimizer | Apply GPU-specific optimizations | | MemoryManager | VRAM management and cleanup | | GPUMonitor | Real-time GPU metrics | | TensorRTAutoConverter | Model conversion for speed |

RTX 4090 Optimizer

Enables hardware-specific optimizations:

• TF32: 2x speedup on matrix operations
• cuDNN Benchmark: Auto-selects optimal kernels
• AMP: Mixed precision for memory efficiency
• Thread Tuning: Optimized for high-core-count CPUs

Default settings for RTX 4090 + Threadripper:
├─ TF32: Enabled
├─ cuDNN Benchmark: Enabled
├─ Memory Fraction: 95%
└─ CPU Threads: 48

Memory Manager

Actions:

• clear_cache: Standard VRAM cleanup
• optimize: Conditional cleanup based on threshold
• report: Detailed memory breakdown
• emergency_clear: Aggressive cleanup for OOM recovery

GPU Monitor

Real-time metrics via nvidia-smi:

• GPU Utilization (%)
• VRAM Usage (%)
• Temperature (°C)
• Power Draw (W)

TensorRT Converter

Convert models for faster inference:

• fp16: 40-60% speedup
• int8: 60-80% speedup (requires calibration)
• fp32: 20-30% speedup

Requirements

• ComfyUI
• torch >= 2.0.0
• NVIDIA RTX 4090 (or compatible GPU)
• nvidia-smi (for GPUMonitor)

License

MIT License

Author

Joseph Ibrahim - VFX Lighting TD