AMD Ryzen™ AI MAX+ 395 Processor: Breakthrough AI Performance in Thin and Light

The AMD Ryzen AI MAX+ 395 (codename: "Strix Halo") is the most powerful x86 APU and  delivers a significant performance boost over the competition. Powered by "Zen 5" CPU cores,  50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD  RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade for the premium thin  and light form factor. The Ryzen AI MAX+ 395 is available in options ranging from 32GB all  the way up to 128GB of unified memory – out of which up to 96GB can be converted to VRAM  through AMD Variable Graphics Memory.  

The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered  application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM  Studio allows users to locally run the latest language model without any technical knowledge  required. Deploying new AI text and vision models on Day 1 has never been simpler.  

The "Strix Halo" platform extends AMD performance leadership in LM Studio with the new AMD  Ryzen AI MAX+ series of processors.  

As a primer: the model size is dictated by the number of parameters and the precision used.  Generally speaking, doubling the parameter count (on the same architecture) or doubling the 

precision will also double the size of the model. Most of our competition's current-generation  offerings in this space max out at 32GB on-package memory. This is enough shared graphics  memory to run large language models (roughly) up to 16GB in size.  

Benchmarking text and vision language models in LM Studio 

For this comparison, we will be using the ASUS ROG Flow Z13 with 64GB of unified memory.  We will restrict the LLM size to models that fit inside 16GB to ensure that it runs on the  competition's 32GB laptop. 

From the results, we can see that the ASUS ROG Flow Z13 - powered by the integrated  Radeon 8060S and taking full advantage of the 256 GB/s bandwidth - effortlessly achieves up  to 2.2x the performance of the Intel Arc 140V in token throughput.  

The performance uplift is very consistent across different model types (whether you are running  chain-of-thought DeepSeek R1 Distills or standard models like Microsoft Phi 4) and different  parameter sizes. 

In time to first token benchmarks, the AMD Ryzen AI MAX+ 395 processor is up to 4x faster  than the competition in smaller models like Llama 3.2 3b Instruct. 

Going up to 7 billion and 8 billion models like the DeepSeek R1 Distill Qwen 7b and DeepSeek  R1 Distill Llama 8b, the Ryzen AI Max+ 395 is up to 9.1x faster. When looking at 14 billion  parameter models (which is approaching the largest size that can comfortably fit on a standard  32GB laptop), the ASUS ROG Flow Z13 is up to 12.2x faster than the Intel Core Ultra 258V  powered laptop – more than an order of magnitude faster than the competition!

The larger the LLM, the faster AMD Ryzen AI Max+ 395 processor is in responding to the  user query. So whether you are having a conversation with the model or giving it large  summarization tasks involving thousands of tokens – the AMD machine will be much faster to  respond. This advantage scales with the prompt length – so the heavier the task – the more  pronounced the advantage will be.  

Text-only LLMs are also slowly getting replaced with highly capable multi-modal models that  have vision adapters and visual reasoning capabilities. The IBM Granite Vision is one example  and the recently launched Google Gemma 3 family of models is another – with both providing highly capable vision capabilities to next generation AMD AI PCs. Both of these models run  incredibly performantly on an AMD Ryzen AI MAX+ 395 processor. 

An interesting point to note here: when running vision models, the time to first token metric also  effectively becomes the time it takes for the model to analyze the image you give it. 

The Ryzen AI Max+ 395 processor is up to 7x faster in IBM Granite Vision 3.2 3b, up to 4.6x  faster in Google Gemma 3 4b and up to 6x faster in Google Gemma 3 12b. The ASUS ROG  Flow Z13 came with a 64GB memory option so it can also effortlessly run the Google Gemma 3  27B Vision model – which is currently considered the current SOTA (state of the art) vision  model.  

A very cool demo showcasing this capability is shown here. When given a stock image of a CT  scan, the model was able to identify the organs and provide a diagnosis. This is just one of  many possible use cases for vision models that have complex vision modalities and visual  reasoning capabilities – and these are only going to get better with time. 

Another example is running the DeepSeek R1 Distill Qwen 32b in 6-bit precision (while 4-bit is  industry standard for everyday use cases, coding can require higher precision to maintain  coding accuracy) – which you can use to code a gaming classic in roughly 5 minutes.

Setting up for LLM runs

Now let's talk about how to tune your AMD Ryzen AI Max+ 395 processor for maximum performance and capability for large language models.
 

Image: VGM options on a 32GB laptop. VGM High = 16GB dedicated graphics memory.

Please make sure you are on the latest AMD Software: Adrenalin Edition driver update. AMD  laptops powered by AMD Ryzen AI 300 series processors feature Variable Graphics Memory.  AMD recommends turning on VGM for any LLM workloads to help token throughput and allow  larger model sizes to run. A VGM setting of High is recommended. You can access the VGM  options through the Performance > Tuning tab in AMD Software: Adrenalin Edition. 

You can download and install LM Studio from their website.  

When running LLMs – please check "manually select parameters" and set the GPU Offload  setting to MAX. AMD recommends Q4 K M quantization for everyday use and Q6 or Q8 for  coding.  

Experiencing AI locally on laptops powered by the AMD Ryzen AI MAX+ 395 processor is a  great way for the power user to experience state-of-the-art AI models while having a portable,  thin and light, gaming and productivity powerhouse.