Mini PC that can outperform gaming machines and run large AI models in South Africa

The GMKtec NUCBox Evo-X2 isn’t just an impressive mini PC — it’s a showcase for what AMD’s new Strix Halo architecture, specifically its Ryzen AI Max+ 395 processor, is capable of.

AMD unveiled its new range of Ryzen AI Max 300-series processors ahead of CES 2025 in January, stating that the first devices would become available from the second quarter of the year.

So far, only a handful of products featuring the new architecture have been announced so far, and even fewer have reached the market.

These include the 2025 edition of the Asus ROG Flow Z13 gaming tablet, the HP ZBook Ultra G1a laptop, and the GMKtec Evo-X2.

The Asus tablet showcases Strix Halo’s potential as a portable gaming powerhouse, while HP’s laptop highlights its strengths for mobile workstations and desktop replacements.

Framework — the repairability-focused company known for its modular laptops — has also announced a 4.5ℓ small form factor desktop powered by Strix Halo chips.

However, the product is yet to launch, and Framework does not sell its products in South Africa. In contrast, GMKtec’s Evo-X2 is already available through local retailer Wootware.

The Evo-X2 also showcases the full capabilities of AMD’s Ryzen AI Max+ 395 processor, including its versatile range of external connectivity options such as USB4 and support for up to 128GB of shared LPDDR5x memory.

Because it’s a larger device, the Evo-X2 can also support better cooling than laptops and promises to sustain the maximum 120-watt thermal design power (TDP) of the chip with zero throttling under load.

A standout feature of the Ryzen AI Max+ is its integrated graphics processor — the AMD Radeon 8060S iGPU — which can access up to 96GB of the system’s shared memory as video RAM.

Integrated graphics (iGPUs) are built into the same chip as the central processing unit (CPU), forming what AMD refers to as an Accelerated Processing Unit (APU).

iGPUs have traditionally had a poor reputation, as they have significantly underperformed compared to discrete GPUs. One reason is that shared memory typically offers lower bandwidth than dedicated video memory.

However, Apple’s M-series chips — particularly the Pro, Max, and Ultra variants — have shown that integrated graphics can deliver excellent performance and energy efficiency, challenging long-held assumptions.

Of note is that Apple uses LPDDR5x memory in its computers, the same technology AMD supports on its Ryzen AI and newer mobile processors.

AMD’s vision for APUs dates back to 2006, when it began development on the Fusion project and acquired graphics accelerator company ATI as part of the strategy.

The company emphasised its commitment by rebranding with the tagline “the Future is Fusion” in 2008. Its first APUs were launched in 2011.

While early APUs aimed at desktop and laptop computers generally underwhelmed gamers, AMD went on to develop custom processors with integrated graphics for the PlayStation 4 and 5, Xbox One, and Xbox S|X.

AMD also developed a custom APU for Valve’s handheld gaming console, the Steam Deck. Valve is best known for its Steam platform — the dominant online marketplace for PC games.

These APUs showed that iGPUs capable of handling graphically intensive games were viable, just not widely adopted in the PC market, where discrete and upgradeable graphics cards were the norm.

However, the rise of large language models (LLMs) and Microsoft’s investment in artificial intelligence (AI) may have created a new opportunity for APUs with powerful built-in graphics.

Last year, Microsoft launched its Copilot+ initiative — a new category of Windows PCs designed for AI tasks.

To qualify for Copilot+ branding, devices must meet minimum system requirements, including a neural processing unit (NPU), as well as high-performance CPUs and GPUs.

With a new era of APUs upon us, we secured a GMKtec Evo-X2 unit in Wootware’s first shipment, which arrived in South Africa on 20 June.

The batch was sold out as it landed, and Wootware’s website currently lists 11 July as the expected date for the next shipment.

Our initial tests and overall experience with the Evo-X2 have been promising, and it showcased the versatility of the Ryzen AI Max+ 395.

The APU had no problem running large language models with billions of parameters or running demanding games like Cyberpunk 2077 in 1440p resolution at ultra-high graphical settings.

AI performance

To evaluate the Evo-X2’s ability to run large language models, we experimented with several open-source tools.

Given how new this APU is and how fast the AI space is moving, our results will likely be outdated before long. However, they offer a useful snapshot of its capabilities at launch.

For our first tests, we tried Ollama with Open-WebUI and the LM Studio technology preview specifically built for AMD Ryzen AI. However, neither currently supports the chip’s neural processing unit (NPU).

We quickly learned that using the NPU requires a special driver and models optimised for Microsoft’s ONNX Runtime. This enables hybrid inference across the NPU and Radeon 8060S iGPU, delivering improved performance.

Since Ollama does not support ONNX Runtime models, we looked for a local LLM server that did — and found AMD’s own Lemonade Server.

Unfortunately, unlike Ollama, Lemonade doesn’t display performance metrics alongside model responses. To work around this, we used a third-party function for Open-WebUI called Time Token Tracker.

After several hours spent downloading hundreds of gigabytes of model files, we put the AMD Ryzen AI Max+ 395 to the test.

We used a range of prompts to evaluate programming ability, factual accuracy, logical reasoning, and creative output. These included:

• Write a Python function that takes a list of numbers and returns the second-largest number.
• Explain the difference between a black hole and a neutron star in simple terms.
• If Alice is twice as old as Bob, and Bob is three years older than Charlie, and Charlie is 10, how old is Alice?
• Write a short story about a robot that wants to become a painter.

The table below summarises each model’s response time and token generation rate. To avoid repetition, only results from the Python programming and creative writing prompts are included.

The number of parameters in a model is given at the start of its name, with “B” indicating billions of parameters. All of the models in the table were ones AMD indicated were optimised for hybrid inference.

As the results show, responses only took a few seconds to generate, with longer response times generally correlating with larger models.

For the Deepseek-R1 models, the second number indicates the amount of time spent “thinking”.

Interestingly, the distilled DeepSeek-R1 model based on Qwen could not complete the Python programming question as it got stuck while reasoning through it.

AMD also recently added a non-hybrid version of Meta Platforms’ 17-billion parameter Llama 4 model to Lemonade. The Ryzen AI Max+ 395 was able to run this model with 96GB of VRAM allocated to the iGPU.

However, the response from the Llama 4 model was extremely slow. It was also only able to complete the Python programming question. It seemed to crash for all the other questions.

Gaming and Geekbench performance

When it comes to more traditional performance metrics, our experience with the Ryzen AI Max+ 395 mirrored that of international reviewers who tested devices built around the chip.

It delivers excellent single- and multi-core performance, and its integrated graphics compete with — and occasionally outperform — discrete GPUs like the GeForce RTX 4060 and 4070.

In Geekbench, the APU achieved a single-core score of 2,967 and a multi-core score of 17,474, placing it among the highest-performing processors in the benchmark’s rankings.

Its single-core performance is comparable to the AMD Ryzen 9 7900X and Intel Core i7-14700KF, while its multi-core score is close to that of the Intel Core i7-13700KF and AMD Ryzen 9 7900X3D.

Regarding gaming performance, we tested the Radeon 8060S in two demanding games, Cyberpunk 2077 and Red Dead Redemption 2, using some of their highest graphical settings.

All tests were conducted at 1440p (QHD) resolution and High Dynamic Range (HDR) was enabled where available, as was Windows 11’s “Auto HDR” feature.

While expected, it was still notable that our Cyberpunk 2077 benchmark — using the Ultra preset with no frame generation — produced significantly better results than those reported by reviewers testing the chip in laptops.

For example, the Asus ROG Flow Z13 running at an 80W TDP achieved an average of around 47 frames per second (FPS), whereas the Evo-X2 hit 59 FPS at a 120W TDP.

This highlights the substantial performance gains possible when the APU is given more thermal and power headroom.

With this level of performance, AMD’s new Strix Halo chips are well-positioned to challenge the longstanding stigma around APUs and integrated graphics — and may offer the most credible threat yet to Nvidia’s dominance in laptop GPUs.

Elephant in the room — The price

Whenever the AMD Ryzen AI Max+ 395 comes up, its price is often raised as a concern. Although devices based on the APU aren’t cheap, they’re not overpriced within their performance class.

In South Africa, the GMKtec NUCBox Evo-X2 with 128GB of unified memory sells for R35,999 on Wootware. By comparison, the Asus ROG Flow Z13 with 32GB of RAM was recently discounted to R39,999 — now back up to R45,999.

While it’s natural to compare those prices to a desktop gaming PC you can build for the same price, it’s important to remember that these APUs serve a different market segment.

For those seeking something more power-efficient that generates less heat, therefore requiring less noise to cool, AMD’s new Strix Halo processors offer an excellent trade-off between performance, size, and efficiency.

In the portable PC market, a R36,000–R46,000 budget won’t get you a gaming laptop with more than 8GB of VRAM — and you certainly won’t see Radeon graphics at all.

As for the mini PC segment in which the Evo-X2 competes, Strix Halo has opened several new possibilities — from running local LLMs to high-performance homebrew Steam Box gaming consoles.

Intel introduced its concept for small-form-factor barebone computer kits in 2012, calling them the Next Unit of Computing (NUC) — from which GMKtec’s NUCBox machines derive their name.

In mid-2023, Intel announced that it would stop direct investment in the NUC business and leave future innovation and growth in the space to its ecosystem partners.

While Intel is no longer directly involved in the NUC market, and the definition has expanded to include devices as small as an Apple TV to some as large as shoeboxes, the segment is more vibrant than ever.

With the arrival of powerful APUs like Strix Halo, devices like the Evo-X2 are no longer just a “Next Unit of Computing” — they’re an Absolute Unit of Computing.

GMKtec NUCBOX Evo-X2 photos