AMD: Ryzen AI Max+ 395 Runs LLM Models Up to 122 Billion Parameters with Unified Memory

What Does AMD Ryzen AI Max+ Enable for Local LLM Inference?

AMD published a detailed analysis of LLM inference capabilities on the Ryzen AI Max+ 395 processor through the ROCm blog — the most advanced chip in AMD’s Ryzen AI Max lineup for laptops and mini PC devices. The key advantage of this processor is UMA (Unified Memory Architecture), where the CPU and the integrated Radeon 8060S GPU share the same physical memory pool of 128 GB LPDDR5X memory.

Unlike discrete GPUs with separate VRAM, Ryzen AI Max+ processors enable transparent memory sharing between CPU and GPU, which practically eliminates one of the key limitations of local LLM inference.

What Are the Concrete Benchmark Results on Tested Models?

Testing was conducted on Ubuntu 24.04 LTS with ROCm 7.2.1 and Ollama 0.20.x, using Q4_K_M model quantization:

Model	Size	Generation Speed	GPU Offloading
Qwen3.5 9B	6.2 GB	29.84 tok/s	100% GPU
Qwen3.5 35B-A3B (MoE)	20.5 GB	42.04 tok/s	100% GPU
Qwen3.5 122B-A10B (MoE)	76 GB	8.59 tok/s	61% GPU / 39% CPU

Particularly interesting is the result for the Qwen3.5 35B-A3B MoE (Mixture-of-Experts) model, which outperforms the smaller 9B model despite significantly more parameters — this is because the MoE architecture activates only 3 billion parameters per token, while the rest remains inactive.

How Does Ryzen AI Max+ Position for Professional Users?

For interactive applications, AMD recommends Qwen3.5 35B-A3B, which ensures 42 tok/s — sufficient for a fluid chat experience. For maximum capacity, Qwen3.5 122B-A10B is available, exceeding the GPU’s 64 GB capacity and automatically distributing the load between GPU and CPU, achieving 8.59 tok/s.

The Ryzen AI Max+ 395 positions AMD directly in the segment previously dominated by Apple M4 Max chips with high memory bandwidth — but with the advantage of the open ROCm ecosystem and Linux compatibility with Ollama and related tools.