The AMD Ryzen AI Max+ 395 dominates as the “most powerful” APU on the market, but its competition is questionable

AMD has published official benchmarks for the Ryzen AI Max+ 395 APU
It demonstrates a clear lead over the Intel Core Ultra 7 258V
The tests were conducted in LM Studio with various LLMs

Official benchmarks have backed up the “Strix Halo” AMD Ryzen AI Max+ 395’s performance as the “most powerful x86 APU” on the market for AI computing.

The AMD Ryzen AI Max+ 395 is a 16-core (32 threads) processor with a 50+ peak AI TOPS XDNA 2 NPU, and Radeon 8060S integrated graphics (40 RDNA 3.5 Compute Units) for some serious processing power for the form factor. It’s being primarily marketed by AMD for its handling of AI workloads, such as in applications such as LM Studio.

This is evident in AMD’s published benchmarks for the Ryzen AI Max+ 395, which are measured in ‘tokens per second’ and ‘time to first token’ in LM Studio against its competition. Specifically, we see how the new “Strix Halo” processor inside of the Asus ROG Flow Z13 with 64GB RAM compares to a similar spec Asus Zenbook S14 with 32GB RAM.

The latter machine has half the unified memory and is using the Intel Core Ultra 7 258V APU with its baked-in Arc integrated graphics clocked at 140V, so it’s not necessarily a 1:1 comparison. However, AMD has showcased the prowess of its latest chipset in LM Studio 0.3.11 with “various LLMs” with a 16GB model size, demonstrating at least twice the effective tokens per second with DeepSeek R1, Phi 4 Mini Instruct, and Llama 3.2 compared to its rival.

The lead becomes more dramatic when comparing time to first token in text models, with up to 12.2x faster as evidenced by the benchmarks in DeepSeek R1 Distill Qwen 14b, with a similar lead of 11.3x in Phi 4 14b. It’s not a consistent lead across all text models, however, as the Ryzen AI Max+ 395 is anywhere from 4x to 9x faster in Llama 3.2 and other DeepSeek R1 distilled models.

AMD’s Ryzen AI Max+ 395 is also claimed to be up to seven times faster in SOTA vision models in the time to first token, this can be seen in IBM Granite Vision 3.2 2B while the chip is six times faster in Google Gemma 3 12b; it’s roughly halved when comparing against Gema 3 4b, though.

Powerful performance that should not be too shocking

AMD’s leading “Strix Point” APU is head and shoulders above the Intel Core Ultra 7 processor in a way that should not be surprising to those interested in AI computing. That’s because Team Blue’s hardware was made with lower threshold AI computing in mind, and this can be seen in the architectural differences when analyzing the two.

The Intel Core Ultra 7 258V features eight cores and eight threads with a maximum boost clock of up to 4.8 GHz and a maximum TDP of 37W. In contrast, the AMD Ryzen AI Max+ 395 has 16 cores, 32 threads, a boost clock of up to 5.1 GHz, and a default TDP of 55W. However, this TDP is configurable up to 120W, so it’s a night and day hardware difference in the chipsets. Of course, AMD’s hardware was going to come out on top; it’s far more powerful across the board.

Then we have to consider the two tested machines used for the benchmarks, the differences between the Asus ROG Flow Z13 (a leading gaming laptop) and the Asus Zenbook S14 (a midrange ultrabook). We reviewed the latter device late last year giving it a four-star write-up, citing the “solid performance” from the Lunar Lake processor. The chip debuted inside this machine (and similar) back in September 2024, while the AMD Ryzen™ AI Max+ 395 hit the scene this month.

It’s not just AI laptops that are using the flagship Ryzen AI chipset for its performance capabilities as a myriad of mini PCs are using them for productivity and even gaming use. It’s become a race to launch the most powerful AI mini PCs possible as mid-March to mid-May are targeted from companies, such as GMKTec and Aoostar, which are leading the charge.