Jul 07, 2026 22:00:00

A review of AMD's AI development kit, 'Ryzen AI Halo,' has already appeared, describing it as a powerful mini workstation with a massive 128GB of unified memory, a Zen 5 CPU, and a Radeon 8060S GPU, all for around 640,000 yen.

At AMD's keynote speech at CES 2026, held in January 2026, the ' AMD Ryzen AI Halo ,' a reference mini-PC for AI developers, was

announced , with shipments beginning sequentially from July 10th (Friday) of the same year. The Ryzen AI Halo is equipped with a powerful APU that integrates a 16-core, 32-thread Zen 5 CPU, a Radeon 8060S GPU with 40 CUs (Computer Units), and an XDNA 2 NPU on a single chip, along with a massive 128GB of LPDDR5X unified memory. It is designed as an appliance to compete with rivals such as NVIDIA's mini-workstation ' DGX Spark ' and Apple's Mac Studio . LTT Labs, an overseas tech media outlet, has published a review of the Ryzen AI Halo.

AI Dev Kit, Batteries Included - AMD Ryzen AI Halo | LTT Labs
https://www.lttlabs.com/articles/2026/07/06/amd-ryzen-ai-halo

◆ Hardware
The table below summarizes the specifications of the Ryzen AI Halo, DGX, and Mac Studio.

The Ryzen AI Halo is a mini PC built around the Zen 5 generation Ryzen AI Max+ 395. It features a 16-core, 32-thread CPU, an integrated Radeon 8060S with 40 CUs (RDNA 3.5 generation), and an XDNA 2 NPU. It has 128GB of LPDDR5X-8000 memory with a bandwidth of 256GB/s. A 2TB M.2 SSD is included as standard storage, and the SSD is replaceable.

The main unit has a footprint of approximately 15cm square, a height of less than 5cm, and weighs 1.2kg. The back features four USB 3.2 Type-C ports, an HDMI 2.1 port, and 10GbE Ethernet, and supports Wi-Fi 7 and Bluetooth 5.4 wireless communication. One of the USB Type-C ports, closest to the power button, is dedicated to power supply via USB-C Power Delivery.

The cooling system consists of two blower fans and omnidirectional air intakes. The Ryzen AI Max+ 395 has a TDP of 120W and can boost up to 140W, so the fan speed increases under heavy load. However, LTT Labs states that the sound quality was relatively mild, mainly consisting of wind noise rather than high-pitched noise.

The bottom of the unit being tested reportedly rose to approximately 50°C.

The selling price is $3999.99 (approximately 644,000 yen), and it is available in Windows 11 Pro and Linux versions. The Linux version that LTT Labs tested was reportedly using AMD's own distribution based on Debian 13.4.

◆Performance
LTT Labs used 'llama-bench,' which is included with the local LLM inference engine 'llama.cpp,' to validate Qwen 3.6 35B A3B, Gemma 4 31B IT, and GLM 4.7 Flash. Model sizes ranged from approximately 17GB to 32GB, and the basic test used a pp512/tg128 configuration that processes 512 tokens as input and generates 128 tokens.

For comparison, we used a Framework Desktop (orange and green) equipped with the same Ryzen AI Max+ 395 and 128GB of memory as the Ryzen AI Halo (red and blue), as well as a Mac Studio with an M2 Ultra processor (purple) and a Mac Studio with an M3 Ultra processor (red and purple). The following bar graph compares the token generation speed measured with llama-bench. The Mac Studio has a maximum memory bandwidth of approximately 800GB/s, surpassing the Ryzen AI Max+ 395 machine with 256GB/s. In particular, token generation is highly dependent on memory bandwidth, and in Gemma 4, the Mac Studio showed a token generation speed approximately 2 to 3 times faster than the Ryzen AI Max+ 395 machine.

On the other hand, the prompting process for reading the input text is highly dependent on computational performance, and the difference was relatively small in the Dense model, which uses all parameters in the calculation for each input.

Furthermore, in tests simulating agent usage with context lengths increased to 0, 4096, 8192, 16384, 32768, and 65536 tokens, prompt processing and generation speeds significantly decreased across all models. In a case where Gemma 4 was run on Vulkan, processing with 65536 tokens did not complete within 30 minutes. For agent processing using long conversation histories or tool calls, it is necessary to consider not only the model's inherent generation performance but also the performance degradation due to context length.

For power and cooling testing, we ran the llama-bench prefill process 20 times over approximately 20 minutes. The Ryzen AI Halo maintained a boost state of 140W for about 5 minutes from the start, and then settled down to around 120W TDP. The Framework Desktop, which was used for comparison, maintained around 120W, but intermittently rose to around 130W.

◆Software
LTT Labs highlighted that the most significant feature of Ryzen AI Halo is not the Ryzen AI Max+ 395 itself, but the software environment provided by AMD. While several mini PCs already exist that use the same processor, Ryzen AI Halo allows users to manage software installation, updates, documentation, and system restore all in one place, starting from the 'AMD Ryzen AI Developer Center.'

At its core is BKC (Best Known Configurations). These are configurations that AMD has verified for compatibility among driver, package, and various software combinations, aiming to reduce dependency issues that often occur when building AI environments. Developers can start with these verified configurations and then freely modify them as needed to create a standard Linux or Windows environment.

With a standard Ryzen AI Max+ 395 machine, you need to adjust the amount of memory allocated to AI processing via command line. Ryzen AI Halo allows you to change this setting through the AMD Ryzen AI Developer Center using a slider on Linux or a dropdown menu on Windows, which LTT Labs considers a helpful feature for initial setup.

'AMD AI Playbooks,' which guide users through running local LLMs and building AI coding environments, are also an important element. They provide step-by-step instructions for remote connection via AMD Sync, running local LLMs with LM Studio or Lemonade, coding support using VS Code and Cline, and running and fine-tuning LLMs with PyTorch. LTT Labs appreciates that the process from setup to operation verification is compressed into just a few steps, but points out that it is not a resource for deeply learning the mechanisms of scripts or the meaning of each setting.

On the other hand, AMD AI Playbooks still has some unresolved bugs on GitHub, and LTT Labs itself has encountered some issues. LTT Labs believes that the value of Ryzen AI Halo lies in its software environment, which includes BKC and AI Playbooks, and points out that it will be crucial for AMD to continuously update the content and quickly fix configurations that experience problems.

◆NPU
Ryzen AI Halo features an XDNA 2 NPU. LTT Labs ran the gpt-oss-20b-FLM model on the NPU using AMD's Lemonade and FastFlowLM. Although telemetry showing NPU utilization alone was not available, they reported that with CPU and GPU utilization at near zero, the total power consumption of the Ryzen AI Max+ 395 package was a maximum of 35W, and the generation speed was 20 tokens per second.

While NPUs don't offer the same level of computing power as GPUs, they are characterized by their superior power efficiency. They are well-suited for applications that require low power consumption and continuous AI processing, such as sensor processing in cameras, allowing the CPU and GPU to be freed up for their primary uses. LTT Labs believes that adopting the Ryzen AI Max+ 395, which includes an NPU, rather than relying on a massive external GPU, could lead to the development of more power-efficient local LLMs.

◆Power supply via USB-C PD
The Ryzen AI Halo operates solely via USB-C Power Delivery (PD). The included power adapter supports USB-C PD Extended Power Range (EPR) and can supply up to 48V x 5A = 240W. However, according to LTT Labs' measurements, the Ryzen AI Halo's power consumption did not exceed 200W from the power supply unit.

With USB-C PD, the power adapter first informs the connected device of the voltage and current it can handle. The default Standard Power Range (SPR) is 20V, 5A, and 100W, but the Ryzen AI Halo requests EPR mode to obtain extended power supply capabilities from the power source. Ultimately, the device requests a fixed output of 48V and 5A, operating under power supply conditions of up to 240W.

LTT Labs also observed PD communication using Infineon's CY4500-EPR. They noted as interesting behavior that, despite operating at a fixed voltage, the Ryzen AI Halo continuously sent extended control messages to the power adapter. The ability to power a small development machine with a 120W-class processor using only USB-C power supply simplifies wiring when installing multiple units in racks or clusters.

◆Summary
LTT Labs concluded that Ryzen AI Halo 'doesn't offer any particularly novel performance features compared to other Ryzen AI Max+ 395-based machines.' However, they stated that AMD's BKC (Backup Knockout) configurations, along with AI Playbooks that outline procedures for LLM execution, coding assistance, and fine-tuning, are valuable for lowering the barrier to entry for AI development on AMD hardware.

Furthermore, LTT Labs acknowledges that 'the Ryzen AI Halo is expensive at around 644,000 yen, and in some cases, a small PC with a comparable processor can be obtained for less.' However, they also praise it as 'a meaningful product as a development kit that comes with everything you need' for developers who want to start local AI development on the AMD platform immediately. They also pointed out that this value will largely depend on whether AMD can continuously update BKC and Playbooks and quickly resolve software issues.