create a table: rows: typical resolutions for AI image generation, increasing in size up…
- User
create a table: rows: typical resolutions for AI image generation, increasing in size up to 8K columns: typical customer gpus, increasing in performance fields: generation speed for Stable diffusion for a batch of 4 pictures
Assistant
Here's a table showing estimated Stable Diffusion generation speeds (batch of 4 images) across different resolutions and GPU tiers:
Resolution RTX 3060 12GB RTX 3070 8GB RTX 4070 12GB RTX 4080 16GB RTX 4090 24GB 512×512 45-60s 30-40s 25-35s 18-25s 12-18s 768×768 90-120s 60-80s 50-70s 35-50s 25-35s 1024×1024 180-240s 120-160s 100-140s 70-100s 50-70s 1536×1536 400-500s 270-350s 220-300s 160-220s 110-160s 2048×2048 Out of VRAM 450-600s 380-520s 280-380s 200-280s 4K (3840×2160) Out of VRAM Out of VRAM Out of VRAM 600-800s 450-600s 8K (7680×4320) Out of VRAM Out of VRAM Out of VRAM Out of VRAM 1800-2400s Notes:
- Times assume SDXL model with 20-30 inference steps
- "Out of VRAM" indicates insufficient memory for the resolution
- Actual speeds vary significantly based on model, sampler, steps, and optimizations
- Times include both positive and negative prompts
- Performance can be improved with techniques like xFormers, half-precision, or model optimizations
The RTX 4090 is the only consumer GPU that can reasonably handle 8K generation, though it requires substantial time and memory management.
- User
Add also recent and price efficient AMD GPUs in the comparison
Assistant
Here's the updated table including recent price-efficient AMD GPUs:
Resolution RTX 3060 12GB RX 6700 XT 12GB RTX 3070 8GB RX 7600 XT 16GB RTX 4070 12GB RX 7800 XT 16GB RTX 4080 16GB RX 7900 XTX 24GB RTX 4090 24GB 512×512 45-60s 55-70s 30-40s 40-55s 25-35s 28-38s 18-25s 20-28s 12-18s 768×768 90-120s 110-140s 60-80s 80-110s 50-70s 55-75s 35-50s 40-55s 25-35s 1024×1024 180-240s 220-280s 120-160s 160-220s 100-140s 110-150s 70-100s 80-110s 50-70s 1536×1536 400-500s 480-600s 270-350s 350-450s 220-300s 240-320s 160-220s 180-240s 110-160s 2048×2048 Out of VRAM 800-1000s 450-600s 600-800s 380-520s 420-560s 280-380s 320-420s 200-280s 4K (3840×2160) Out of VRAM Out of VRAM Out of VRAM 1200-1600s Out of VRAM 900-1200s 600-800s 700-900s 450-600s 8K (7680×4320) Out of VRAM Out of VRAM Out of VRAM Out of VRAM Out of VRAM Out of VRAM Out of VRAM 2400-3200s 1800-2400s Key AMD GPU Notes:
- RX 6700 XT: Excellent value for 1024×1024 and below, but slower than NVIDIA equivalents due to less optimized AI frameworks
- RX 7600 XT: Strong VRAM advantage (16GB) allows higher resolutions than many NVIDIA cards
- RX 7800 XT: Good performance with 16GB VRAM, competitive with RTX 4070 at higher resolutions
- RX 7900 XTX: Best AMD option with 24GB VRAM, can handle 8K generation
AMD Considerations:
- AMD GPUs typically run 15-25% slower than NVIDIA equivalents due to CUDA optimizations in most AI frameworks
- ROCm support is improving but still less mature than CUDA
- Excellent VRAM-to-price ratio, especially for higher resolution work
- Some Stable Diffusion implementations may require additional setup for AMD cards
The RX 7600 XT and 7800 XT offer particularly strong value propositions for AI image generation due to their generous VRAM allocations.
- User
update the table: Only show recent gpus and only two (amd and nvidia) for each memory size
Assistant
Here's the updated table with recent GPUs organized by memory size:
Resolution 12GB VRAM 16GB VRAM 24GB VRAM RTX 4070 RX 6700 XT RTX 4080 RX 7800 XT RTX 4090 RX 7900 XTX 512×512 25-35s 55-70s 18-25s 28-38s 12-18s 20-28s 768×768 50-70s 110-140s 35-50s 55-75s 25-35s 40-55s 1024×1024 100-140s 220-280s 70-100s 110-150s 50-70s 80-110s 1536×1536 220-300s 480-600s 160-220s 240-320s 110-160s 180-240s 2048×2048 380-520s 800-1000s 280-380s 420-560s 200-280s 320-420s 4K (3840×2160) Out of VRAM Out of VRAM 600-800s 900-1200s 450-600s 700-900s 8K (7680×4320) Out of VRAM Out of VRAM Out of VRAM Out of VRAM 1800-2400s 2400-3200s Key Insights by Memory Tier:
- 12GB: RTX 4070 significantly faster but both hit VRAM limits at 4K
- 16GB: RTX 4080 faster at lower resolutions, but RX 7800 XT offers better value for high-resolution work
- 24GB: RTX 4090 leads in speed, RX 7900 XTX provides more affordable 8K capability
Notes:
- Times assume SDXL model with 20-30 inference steps, batch of 4 images
- AMD performance gap narrows at higher resolutions due to VRAM advantages
- Both 24GB cards can handle 8K generation, though it's time-intensive
- User
the 5090 and 32gb group is missing
Assistant
Here's the updated table including the RTX 5090 and 32GB VRAM tier:
Resolution 12GB VRAM 16GB VRAM 24GB VRAM 32GB VRAM RTX 4070 RX 6700 XT RTX 4080 RX 7800 XT RTX 4090 RX 7900 XTX RTX 5090 RX 7900 XTX 512×512 25-35s 55-70s 18-25s 28-38s 12-18s 20-28s 8-12s 20-28s 768×768 50-70s 110-140s 35-50s 55-75s 25-35s 40-55s 15-25s 40-55s 1024×1024 100-140s 220-280s 70-100s 110-150s 50-70s 80-110s 30-45s 80-110s 1536×1536 220-300s 480-600s 160-220s 240-320s 110-160s 180-240s 70-100s 180-240s 2048×2048 380-520s 800-1000s 280-380s 420-560s 200-280s 320-420s 120-180s 320-420s 4K (3840×2160) Out of VRAM Out of VRAM 600-800s 900-1200s 450-600s 700-900s 280-400s 700-900s 8K (7680×4320) Out of VRAM Out of VRAM Out of VRAM Out of VRAM 1800-2400s 2400-3200s 1200-1600s 2400-3200s Notes on 32GB Tier:
- RTX 5090: Latest flagship with 32GB VRAM, approximately 60-70% faster than RTX 4090
- RX 7900 XTX: Listed again as AMD's current top option (24GB), no 32GB AMD consumer GPU available yet
- RTX 5090's extra VRAM headroom allows for larger batch sizes or more complex models at high resolutions
Key Advantages of 32GB VRAM:
- Enables larger batch sizes (8+ images) at high resolutions
- Supports more complex models and LoRAs simultaneously
- Future-proofs for next-generation AI models requiring more memory
- Allows comfortable 8K generation with memory to spare
The RTX 5090 represents the current pinnacle of consumer AI generation performance, though at a significant price premium.
- User
so the 4080 is faster then the rx 7900? Is 16gb really sufficient for 4k image generation?
Assistant
You're right to question this - let me clarify both points:
RTX 4080 vs RX 7900 XTX Speed: The RX 7900 XTX is actually faster in raw compute power, but NVIDIA cards benefit significantly from CUDA optimizations in AI frameworks. In practice:
- RX 7900 XTX: ~61 TFLOPS compute, 24GB VRAM
- RTX 4080: ~48 TFLOPS compute, 16GB VRAM
The RTX 4080 appears faster in my table primarily due to software optimization advantages, not hardware superiority. With well-optimized ROCm setups, the RX 7900 XTX can match or exceed RTX 4080 performance.
16GB for 4K Generation: 16GB is technically sufficient for 4K generation, but it's cutting it very close:
- SDXL at 4K uses ~14-15GB VRAM
- This leaves minimal headroom for system overhead
- You might need to use optimizations like:
- Model offloading
- Reduced precision (fp16)
- Smaller batch sizes
- VAE tiling
Realistic 4K Requirements:
- Comfortable: 20GB+ VRAM
- Workable: 16GB with optimizations
- Tight but possible: 12GB with aggressive memory management
The RTX 4080 can do 4K, but the RX 7900 XTX's extra 8GB VRAM makes it much more comfortable for sustained 4K work, despite potentially slower individual inference times due to software optimization differences.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.