AMD Ryzen AI Max+395: A Leap Forward in Generative AI Performance with Consumer PC

Key Takeaways

• AMD Ryzen™ AI MAX+395 with 128GB unified memory (up to 112 GB allocatable by the GPU) space provides unique capability of running generative AI workloads on a laptop form factor
• Powerful RDNA™ 3.5 based GPU technology with up to 40 compute units
• Up to 3.9x performance advantage over MacBook Pro 48GB with M4 Pro silicon running Stable Diffusion models^[1]
• On device 70 billion parameter LLM models on AMD Ryzen AI MAX+395 with unified memory architecture

Brief Overview of AMD Ryzen AI Max+395 codenamed Strix Halo

AMD Ryzen AI Max+395 is a groundbreaking SoC (system on a chip) that sets new standards for an APU in memory capacity and generative AI capabilities.  Ryzen AI Max+395 code named Strix Halo, is designed to excel in demanding GenAI workloads requiring large memory space and AI computational needs, making it an ideal choice for high-end client PCs demanding the best-in-class Generative AI experiences.

Figure 1: AMD Ryzen AI Max PRO Series Processors

Figure 2: AMD Ryzen AI Max PRO Series Processors with Unified Memory Architecture

Ryzen AI Max+395 boasts several key features and capabilities that distinguish it in the market. With its leadership memory capacity in its class, it ensures seamless multitasking and efficient handling of large datasets. Its advanced GPU capabilities with up to 40 RDNA™ 3.5 based compute units and optimized software stack provides significant performance boost and capabilities for generative AI and creator workloads. With unified memory capabilities, the data can be shared between different accelerators as shown in Fig2 without needing expensive memory copies and saving memory utilization for demanding generative AI workloads.

In today's rapidly evolving tech landscape, the importance of GenAI workloads in client PCs cannot be overstated. These workloads drive innovation across various PC client applications and experiences. AMD Ryzen AI Max+395 is optimized to meet these demands, providing users with the performance and reliability they need to stay ahead.

Comparative Analysis with Apple MacBook Pro

GenAI Performance Comparison

• Generative AI Performance Advantage: Ryzen AI Max+395 demonstrates a 3.9x performance advantage over the Apple MacBook Pro 16” equipped with M4 Pro silicon with 48GB unified memory running Stable Diffusion 3.5 image generation, attributed to software optimization, efficient memory management and leadership memory capacity.

• Concurrent Workloads: It achieves up to 2.6x faster token generation^[2] and 3.3x faster image generation^[3] when running concurrent GenAI workloads, thanks to its large memory pool and enhanced GPU capabilities. 

Image Generation with Stable Diffusion 3.5

For image generation test, we used Stability AI’s state of the art text to image generation model Stable Diffusion3.5 and ran it on both AMD Ryzen Max+ 395 (Strix Halo) and Apple MacBook M4 Pro machines. Running this model on Strix Halo shows a 3.9x performance advantage over the MacBook M4 Pro 48GB system.

You can refer to this technical blog on how AMD worked with Stability AI to enhance generative AI workloads including Stable Diffusion 3X on all AMD GPU platforms.

Figure 3: Speedup vs Apple M4 Pro MacBook

On AMD Ryzen AI Max+395, we used Amuse, while the Apple M4 Pro (48GB) utilized Comfy UI running Stability AI Stable Diffusion 3.5 Large. The key software optimizations are based on fused attention nodes, unpacked attention nodes, and weight pruning. For more details on the software optimizations that contributed to the 3.9x performance benefit on AMD devices, please refer to the above Stability AI Blog and AMD Technical Blog.

Concurrent Stable Diffusion and Large Language Models (LLM)

Strix Halo achieves up to 2.6x faster token generation and 3.3x faster image generation when running concurrent GenAI workloads.

 Figure 4: Speedup vs. Apple M4 Pro MacBook

During testing, the M4 Pro 48GB was observed to rely on swap memory, which significantly slowed its performance. When running Stable Diffusion 3.5 Large and the Phi4 14 billion parameter models concurrently, the AMD Ryzen™ AI Max+395 demonstrated a clear advantage due to its large memory pool, superior GPU capabilities, and optimized software stack.

Real-World Applications

Generative AI workloads are transforming various industries by enabling new applications and driving innovation. For instance, in the healthcare sector, AI models can assist in diagnosing diseases and personalizing treatment plans. In the entertainment industry, AI can generate realistic graphics and animations, enhancing user experience. The AMD Ryzen™ AI Max+395, with its superior performance in concurrent AI workloads, is well-suited for these and many other applications.

Configurations Tested

To evaluate the performance of AMD Ryzen AI Max+ and Apple MacBook Pro platforms, we conducted a series of tests using various configurations and environments. Below are the details of the setups used:

Workloads Tested

Stable Diffusion:

• Models: Stability AI Stable Diffusion 3.5 Large, 3.5 Medium, 3.0 Medium
• Image Size: 1024x1024
• Configuration: Classifier-Free Guidance (CFG) scale=4.5, steps=20
• For more details on the models, you can visit:
  https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main
  https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/tree/main

Large Language Models (LLM):          

• Phi4 14B
• Windows: https://huggingface.co/microsoft/phi-4-onnx
• macOS: https://ollama.com/library/phi4
• DeepSeek-R1-Distill-Llama-70B 
  • Windows: https://huggingface.co/amd/DeepSeek-R1-Distill-Llama-70B-dml-int4-awq-block-128

Concurrent Models:

• Stable Diffusion 3.5 Large and Phi4 14B parameter model (referenced above)

Conclusion

In summary, the AMD Ryzen AI Max+395 APU sets new standards in memory capacity and generative AI performance. Its ability to handle demanding GenAI workloads with ease makes it a powerful platform for developers and researchers. Whether you need superior performance for image generation or large language models, Ryzen AI Max+395 is optimized to meet your needs.

References

• https://gpuopen.com/learn/accelerating_generative_ai_on_amd_radeon_gpus/
• https://stability.ai/news/stable-diffusion-now-optimized-for-amd-radeon-gpus
• https://stability.ai/news/stable-diffusion-now-optimized-for-amd-radeon-gpus
• https://www.amd.com/en/products/graphics/radeon-ai.html
• https://www.amd.com/en/blogs/2024/llm-on-amd-gpu-memory-footprint-and-performance-i.html