NRGUMN AI CREATION
ComfyUI RTX 5080 Benchmark Results
RTX 5080 16GB • Validated Testing
ComfyUI Performance: 2.3x Faster
Complete benchmark results from RTX 5080 16GB testing. SageAttention 2.2.0 + TorchCompile + Triton 3.5.1 Windows + LOW_VRAM configuration achieves 130% performance improvement (2.3x speedup) in real AI video generation workflow.
10m 12s
Optimized Time
23m 28s
Baseline Time
+130%
Improvement
13.8 GB
VRAM Used
📊 Test Environment
Hardware
• GPU: NVIDIA GeForce RTX 5080 16GB
• RAM: 32 GB
• OS: Windows 11
• GPU: NVIDIA GeForce RTX 5080 16GB
• RAM: 32 GB
• OS: Windows 11
Software
• ComfyUI v0.3.59
• PyTorch 2.8.0+cu129
• SageAttention 2.2.0
• Python 3.12.10
• Triton 3.5.1 Windows
• ComfyUI v0.3.59
• PyTorch 2.8.0+cu129
• SageAttention 2.2.0
• Python 3.12.10
• Triton 3.5.1 Windows
Test Workflow
• Type: Video generation (audio-driven)
• Frames: 452 total
• Resolution: 832x464
• Windows: 7 overlapping
• Steps: 6 per window
• Type: Video generation (audio-driven)
• Frames: 452 total
• Resolution: 832x464
• Windows: 7 overlapping
• Steps: 6 per window
📈 Performance Comparison
Four different configurations tested to identify the optimal setup for RTX 5080 16GB.
Total Execution Time (seconds) - Lower is Better
| Configuration | Time | vs Baseline | VRAM | Status |
|---|---|---|---|---|
| PyTorch + LOW_VRAM | 23m 28s | Baseline | 13.8 GB | ✓ Stable |
| PyTorch + TorchCompile + LOW | 22m 48s | +2.9% | 13.8 GB | ✓ Stable |
| SageAttention + LOW | 11m 28s | +104.7% | 13.8 GB | ✓ Stable |
| SageAttention + TorchCompile + LOW | 10m 12s | +130.2% | 13.8 GB | ✓ OPTIMAL |
| PyTorch + HIGH_VRAM | CRASH | N/A | OOM | ✗ Failed |
| SageAttention + HIGH_VRAM | FREEZE | N/A | OOM | ✗ Failed |
⚡ Speedup Analysis
Performance Improvement (%) vs Baseline
Key Finding: SageAttention provides the primary speedup (~2x), while TorchCompile adds an additional 12% when combined with SageAttention, creating a synergistic effect.
🔬 Per-Window Performance
The optimized configuration shows highly consistent performance across all 7 windows, with minimal variance compared to the baseline.
Per-Window Execution Time (Optimal Config)
Consistency Analysis:
• Average time per window: ~1m 17s (77 seconds)
• Standard deviation: ±6.5 seconds
• Most consistent iteration time: 12.00-12.67 s/it
• Total reduction: 13 minutes saved vs baseline
• Average time per window: ~1m 17s (77 seconds)
• Standard deviation: ±6.5 seconds
• Most consistent iteration time: 12.00-12.67 s/it
• Total reduction: 13 minutes saved vs baseline
💡 Why This Configuration Works
SageAttention (+105% boost)
• FP8 precision (8-bit vs 32-bit)
• Custom CUDA kernels for RTX
• 50% memory bandwidth reduction
• <0.5% quality loss (imperceptible)
• Attention is memory-bound, not compute-bound
• FP8 precision (8-bit vs 32-bit)
• Custom CUDA kernels for RTX
• 50% memory bandwidth reduction
• <0.5% quality loss (imperceptible)
• Attention is memory-bound, not compute-bound
TorchCompile (+12% additional)
• Optimizes non-attention operations
• Reduces Python overhead
• Fuses operations for better memory access
• Synergizes with SageAttention
• Multiplicative effect, not additive
• Optimizes non-attention operations
• Reduces Python overhead
• Fuses operations for better memory access
• Synergizes with SageAttention
• Multiplicative effect, not additive
LOW_VRAM Mode (stability + speed)
• Smart CPU offloading
• Prevents memory fragmentation
• Clean state between windows
• Counter-intuitively faster
• Essential for 16GB cards
• Smart CPU offloading
• Prevents memory fragmentation
• Clean state between windows
• Counter-intuitively faster
• Essential for 16GB cards
📋 Test Methodology
Controlled Variables:
• Same workflow file for all tests
• Same input files (audio, prompts, seed)
• No other GPU processes running
• Fresh system state between tests
• Multiple runs to validate consistency
• Same workflow file for all tests
• Same input files (audio, prompts, seed)
• No other GPU processes running
• Fresh system state between tests
• Multiple runs to validate consistency
Measured Metrics:
• Total execution time (end-to-end)
• Per-window timing breakdown
• Seconds per iteration (s/it)
• Peak VRAM usage
• Stability (crashes, freezes)
• Total execution time (end-to-end)
• Per-window timing breakdown
• Seconds per iteration (s/it)
• Peak VRAM usage
• Stability (crashes, freezes)
Reproducibility:
• All test scripts included in package
• Complete workflow configuration documented
• Installation steps validated
• Community can reproduce and validate
• Results within ±10% considered successful reproduction
• All test scripts included in package
• Complete workflow configuration documented
• Installation steps validated
• Community can reproduce and validate
• Results within ±10% considered successful reproduction
⚠️ Important Notes & Disclaimers
What Was Tested:
• ONE GPU: RTX 5080 16GB only
• ONE workflow: 452-frame video generation
• ONE configuration: Specific settings and models
• Windows 11 environment
• ONE GPU: RTX 5080 16GB only
• ONE workflow: 452-frame video generation
• ONE configuration: Specific settings and models
• Windows 11 environment
What This Means:
• Your GPU may perform differently
• Your workflow may have different results
• Results WILL vary based on many factors
• No guarantees or warranties provided
• Use these results as a reference guide
• Your GPU may perform differently
• Your workflow may have different results
• Results WILL vary based on many factors
• No guarantees or warranties provided
• Use these results as a reference guide
Expected Results by GPU (estimated):
• RTX 5080 16GB: 10-12 min (tested ✓)
• RTX 4090 24GB: ~9-10 min (similar or better)
• RTX 4080 16GB: ~10-12 min (similar to 5080)
• RTX 4070 Ti 12GB: ~12-14 min (slightly slower)
Community validation on other GPUs welcome!
• RTX 5080 16GB: 10-12 min (tested ✓)
• RTX 4090 24GB: ~9-10 min (similar or better)
• RTX 4080 16GB: ~10-12 min (similar to 5080)
• RTX 4070 Ti 12GB: ~12-14 min (slightly slower)
Community validation on other GPUs welcome!
🎯 Recommendations
For RTX 5080/4080 16GB
Launch:
Workflow: TorchCompile ON, attention_mode "sageattn"
Expected: 10-12 minutes (our test case)
Launch:
--lowvram --use-sage-attentionWorkflow: TorchCompile ON, attention_mode "sageattn"
Expected: 10-12 minutes (our test case)
For RTX 4090 24GB
Launch:
Workflow: TorchCompile ON, attention_mode "sageattn"
Expected: 9-10 minutes (slightly better)
Launch:
--use-sage-attentionWorkflow: TorchCompile ON, attention_mode "sageattn"
Expected: 9-10 minutes (slightly better)
For RTX 4070 Ti 12GB
Launch:
Workflow: TorchCompile OFF (may OOM), attention_mode "sageattn"
Expected: 12-14 minutes
Launch:
--lowvram --use-sage-attentionWorkflow: TorchCompile OFF (may OOM), attention_mode "sageattn"
Expected: 12-14 minutes
📚 Related Resources
- 📖 Complete Beginner Guide to AI Video Generation – Learn how to create AI videos from scratch
- ⚡ ComfyUI RTX Optimization Guide – Complete installation guide
- 🎬 WAN 2.2 for ComfyUI Workflow – Production video workflow
- 🏠 All ComfyUI Workflows – Browse our catalog
- 💬 NRGUMN Discord – Share your benchmark results
←
Get the Optimization Package
Free download • Complete installation guide • Reproduce these results yourself