Baseline Benchmarks for stable/easy diffusion in AM4 APUs (vulkan backend)
This brief baseline benchmark article aims to demonstrate the expected performance of various AMD APUs on the AM4 platform when using image generators that utilize the Vulkan backend.
We have previously discussed the Vulkan backend for ggml/stable-diffusion.cpp, which enables the use of both older and current GPUs and iGPUs—without the drawbacks of the closed CUDA ecosystem.
https://civitai.red/articles/30191/draw-ai-on-cheap-cpu-igpu
In this article, we will now look at how the use of this backend performs on various APUs.
For the benchmark, we used a Ryzen 3 3200G with only its Vega 8 iGPU in a system with 16GB of RAM (2x8GB), allocating 8GB to the iGPU. The Stable/Easy Diffusion configuration used on the 3200G involved LCM (Latent Consistency Models) with 7 steps and a CFG scale of 1.5, running Illustrious SDXL at a base resolution of 768x1024px for the initial image.
Calculation Methodology.
To estimate performance and image generation times across various AM4 platform APUs, a mixed approach combining mathematical hardware modeling and empirical data from the community was used:
Base Benchmark: Actual performance data measured on a Ryzen 3 3200G system (8 Vega graphics cores at 1250 MHz) served as the absolute reference, operating under identical software conditions (Easy Diffusion, Illustrious SDXL with LCM at 7 steps and 768x1024px resolution).
Theoretical Compute Scaling: The raw mathematical power (theoretical TFLOPS) of each processor was calculated by multiplying the number of graphics cores (Compute Units) by the factory clock frequency, then applying the resulting percentage difference to the base time.
Architecture and Lithography Adjustment: Calculations were corrected to reflect the leap in transistor efficiency (from older 12nm technology to the 7nm process used in the 4000/5000 series), which allows for lower internal latency when processing artificial intelligence workloads.
Constraint Factor (Bandwidth): Final projections were adjusted downward by applying a saturation loss rate, accounting for the fact that all iGPUs on the platform share the same physical bottleneck: the system's DDR4 RAM bandwidth.
Comparison Table of Performance and Estimated Times:
Final Conclusion
The performance analysis demonstrates that local Artificial Intelligence generation using SDXL models (such as Illustrious) combined with latent accelerators like LCM is perfectly viable on low-cost hardware, even without a discrete graphics card. Optimizing the system by making a massive VRAM allocation in the BIOS (8 GB) proves to be a critical and successful strategy to bypass Out of Memory (OOM) errors in shared memory architectures.
While the Ryzen 3 3200G serves as a decent entry-level option in this ecosystem (establishing generation times of 3 to 5 minutes per image), the AM4 platform offers a clear upgrade path without requiring a motherboard or RAM replacement:
The Silicon Boundary: Stable Diffusion performance does not rely solely on raw graphics core count (CUs), but on a critical combination of high clock speeds and manufacturing node efficiency (nanometers). Because of this, newer 7nm processors easily outperform legacy models that have more cores but lower clock speeds.
The Physical Bottleneck (DDR4): Regardless of the chosen APU, the ultimate limiting factor on the AM4 platform remains the system memory bandwidth. Lacking dedicated high-speed video memory (such as GDDR6 VRAM), data transfer is physically constrained by native DDR4 RAM speeds.
In conclusion, for users looking to optimize their rendering times within AMD's integrated AM4 ecosystem, upgrading to processors like the Ryzen 7 5700G represents the absolute performance peak, cutting current generation times practically in half (~2.1 minutes). However, to scale production to a professional or commercial level (achieving sub-30-second rendering times), the limits of integrated architecture dictate that the next logical step must be adding a discrete graphics card featuring native CUDA (NVIDIA) or ROCm (AMD) support.-