ASRock AI BOX-A395 Runs Ryzen AI Max+ 395 with Up to 128GB LPDDR5x

Key Takeaways

• Up to 128 GB LPDDR5x unified memory.
• Integrated Radeon 8060S GPU shares system memory.
• XDNA 2 NPU provides up to 50 TOPS AI acceleration.
• Dual Ethernet: 10 GbE and 2.5 GbE for high throughput.
• Supports Windows 11 and Linux for local model inference.

Summary

ASRock Industrial unveiled the AI BOX‑A395, a compact bare‑bone system built around AMD’s Ryzen AI Max+ 395 platform. The device combines a 16‑core Zen 5 CPU, integrated Radeon 8060S GPU, and an XDNA 2 NPU delivering up to 50 TOPS, all fed by up to 128 GB of 8000 MT/s LPDDR5x unified memory. It targets on‑premise inference of large language models, offering dual 10 GbE/2.5 GbE, USB4, Wi‑Fi 7, and support for Windows 11 or Linux. The enclosure packs high‑speed storage, 8K display outputs, and a configurable 120‑160 W power envelope.

Pulse Analysis

Edge AI is moving from experimental labs to production floors, driven by the need for real‑time insights and data privacy. Traditional server deployments are costly and introduce latency, prompting vendors to deliver compact platforms that match data‑center performance. ASRock’s AI BOX‑A395 answers this demand by integrating a Zen 5 CPU, Radeon 8060S graphics, and an XDNA 2 NPU into a 200 × 100 × 232 mm chassis, offering a single‑box solution for on‑site inference of quantized LLMs such as Llama and Mistral.

The technical architecture hinges on a unified LPDDR5x memory pool up to 128 GB, allowing the GPU to access up to 96 GB for graphics‑intensive workloads while the NPU processes AI tensors at 50 TOPS. This eliminates the need for separate discrete GPUs or external accelerators, simplifying system design and reducing power draw. High‑speed connectivity—including two USB4 ports, dual Ethernet (10 GbE and 2.5 GbE), and Wi‑Fi 7—ensures rapid data movement, while PCIe Gen4 storage slots provide fast local datasets, positioning the box as a competitive alternative to larger edge servers.

For enterprises, the AI BOX‑A395 translates to faster model deployment, lower latency, and tighter control over proprietary data. Industries such as manufacturing, healthcare, and finance can run locally quantized models for predictive maintenance, diagnostic assistance, or fraud detection without relying on cloud round‑trips. As AI workloads continue to scale, solutions that blend high compute density, unified memory, and versatile I/O will become essential, and ASRock’s offering signals a broader shift toward edge‑first AI architectures.