The optimization uses MXFP6 quantization, custom GPU kernels, and improved data transfers while retaining 98.3% of BF16 reference image quality, measured by SSIM.
Performance Comparison
| Metric | BF16 Baseline | Optimized |
| DiT inference time | 22.9 s | 3.3 s |
| GEMM kernel latency | 4.34 ms | 1.54 ms |
| GEMM performance | 1× | 2.8× |
| GPU transfer waiting time | 18% | 2% |
| Image quality (relative SSIM) | 100% | 98.3% |
| Precision | BF16M | XFP6/FP8/INT8 |
Total inference time decreased by 85.6%.
Key Technical Optimizations
- MXFP6 GEMM: Reduced matrix multiplication latency by 2.8× compared with BF16.
- Custom attention kernels: Used INT8 queries and keys with MXFP6 values to reduce computation and communication overhead.
- GPU transfer optimization: Reduced exposed waiting time from 18% to 2% through asynchronous transfers and overlapping computation.
- Kernel fusion: Reduced memory traffic and execution overhead.
AMD Instinct MI355X Specifications
| Specification | Value |
| Architecture | AMD CDNA 4 |
| Memory | 288 GB HBM3E |
| Memory bandwidth | 8 TB/s |
| Compute units | 256 |
| Peak BF16 performance | 2.5 PFLOPS |
| Peak MXFP6 performance | 10.1 PFLOPS |
| Maximum board power | 1,400 W |
Krea 3 vs. Krea 2
| Feature | Krea 2 | Krea 3 |
| Primary function | Image generation | Image generation and editing |
| Architecture | Diffusion transformer | Diffusion transformer |
| Native 2K editing | Not a core feature | Supported |
| Instruction-based editing | Limited | Supported |
| MI355X DiT inference benchmark | Not published | 3.3 s |
Note: The 6.9× improvement compares optimized and baseline Krea 3 inference, not Krea 3 against Krea 2.
Image Quality
In the reported precision comparison:
| Format | SSIM | Relative quality |
| BF16 | 0.962 | 100% |
| FP8 | 0.954 | 99.2% |
| MXFP6 | 0.946 | 98.3% |
| MXFP4 | 0.858 | 89.2% |
MXFP6 delivered the preferred balance between inference speed and image similarity.
Conclusion
The optimized implementation achieves 3.3-second DiT inference, 2.8× faster GEMM execution, and a reduction in GPU transfer waiting time from 18% to 2%.
The benchmark measures DiT execution rather than total end-to-end image generation latency.
