AMD has fundamentally reimagined the mobile processor by combining desktop-class computing power with integrated AI capabilities on a single chip. The new Ryzen AI Max+ 395, codenamed Strix Halo, breaks with two decades of conventional design by fusing 16 full Zen 5 CPU cores, a 40-compute-unit graphics engine, and a 50+ TOPS (tera operations per second) neural processing unit (NPU) into a unified architecture with up to 128GB of shared memory. This represents AMD's direct challenge to Apple's M-series chips and discrete mobile GPUs like NVIDIA's RTX 4070 Laptop.

Why Does Unified Memory Matter for AI Chips?

Traditional laptop and desktop designs separate CPU, GPU, and memory into distinct components, forcing data to shuttle between them through slower interconnects. This creates what engineers call a "memory wall" that severely hampers artificial intelligence workloads, which require constant data movement between processing units. The Strix Halo eliminates this bottleneck by giving all components direct access to the same 128GB memory pool through a 256-bit wide LPDDR5X-8533 memory bus capable of delivering 273 gigabytes per second of bandwidth. This unified approach enables the chip to run large language models (LLMs) with 70 billion parameters entirely on the device, without relying on cloud servers.

The architecture achieves this through a sophisticated three-die multi-chip module design. Two separate CPU dies, each containing 8 Zen 5 cores, connect to a massive graphics and input/output controller die via ultra-high-speed Infinity Fabric links capable of sustaining over 400 gigabytes per second of bidirectional bandwidth with latency under 45 nanoseconds. All three dies are manufactured on TSMC's advanced 4-nanometer process node, allowing AMD to reuse proven CPU designs from its desktop Ryzen 9000 series while avoiding the manufacturing defects that plague massive monolithic chips.

How Does Strix Halo Compare to Previous Mobile Processors?

The Ryzen AI Max+ 395 represents a generational leap in mobile computing capability. Unlike standard mobile processors that use a mix of high-performance and efficiency-focused cores, every single CPU core in Strix Halo is a full-sized Zen 5 performance core with dedicated 1-megabyte L2 cache and complete AVX-512 vector execution units. This means the chip can sustain desktop-class performance across all 16 cores simultaneously without the thread-scheduling overhead that hybrid designs introduce.

In real-world benchmarks, the Ryzen AI Max+ 395 running at 115 watts achieved 2,180 points in Cinebench 2024 multi-threaded tests, outperforming Intel's Core Ultra 9 285H by 65 percent and matching performance of desktop-class Ryzen 9 9900X processors. The GPU subsystem, branded as Radeon 8060S, delivers 2,560 stream processors running at 2.9 gigahertz, enabling desktop-class gaming and professional rendering workloads in a mobile form factor.

What Makes the NPU Critical for Local AI?

Beyond CPU and GPU, the Strix Halo integrates a dedicated 50+ TOPS XDNA 2 neural processing unit optimized specifically for AI inference tasks. This specialized hardware accelerates matrix operations that power language models, image recognition, and other machine learning workloads far more efficiently than general-purpose CPU cores. The combination of high-bandwidth unified memory, powerful GPU compute, and dedicated AI acceleration means developers can deploy sophisticated AI features directly on devices without sending data to remote servers, addressing growing privacy and latency concerns.

How to Evaluate Strix Halo for Your Use Case

• Performance Requirements: Assess whether your application needs desktop-class CPU performance (16 cores at 5.1 gigahertz boost), high-end graphics (40 compute units), or AI inference capability (50+ TOPS). Strix Halo excels when all three are needed simultaneously.
• Memory Bandwidth Constraints: If your workload involves moving large datasets between processing units, the 273 gigabyte-per-second unified memory bandwidth eliminates traditional bottlenecks that plague hybrid designs with separate GPU memory pools.
• Thermal and Power Envelope: The configurable TDP range of 45 to 130 watts allows tuning for different form factors, from ultra-portable devices at lower power to high-performance workstations at maximum thermal headroom.
• AI Model Deployment: If you need to run 70-billion-parameter language models locally without cloud connectivity, the 128GB unified memory capacity and 50+ TOPS NPU provide the necessary hardware foundation for real-time inference.

The Strix Halo lineup includes four tiers beyond the flagship 395 model. The Ryzen AI Max 390 reduces CPU cores to 12 and maintains 40 GPU compute units, while the 385 and 380 variants scale down further to 8 and 6 cores respectively with reduced GPU configurations. All models share the same 32-megabyte system-level cache, 256-bit memory bus, and 50+ TOPS NPU, ensuring consistent AI acceleration across the product family.

What's Driving the Global Push for Local AI Chips?

AMD's architectural shift reflects a broader industry trend toward processing AI workloads on-device rather than in cloud data centers. South Korean chip startups are aggressively pursuing this same strategy, with companies like FuriosaAI, DEEPX, Mobilint, and Rebellions securing overseas customers and establishing international sales bases. FuriosaAI recently opened a Singapore office to expand adoption of its second-generation RNGD neural processing unit across Asia-Pacific, following mass production that began in January 2026. The company is simultaneously deploying 1,800 RNGD units in a 15-megawatt AI data center in Stockholm, with plans to add more than 7,000 units in subsequent expansion phases beginning in early 2027.

DEEPX has secured 77 commercial purchase orders worth more than $13 million for its DX-M1 on-device AI chip across more than 10 countries and regions, with applications spanning robotics, smart factories, manufacturing automation, healthcare, and video surveillance. Mobilint has begun mass production of its Regulus system-on-chip and is expanding deliveries to the United States, Germany, Taiwan, and Japan. These Korean companies emphasize that chip performance alone is insufficient to compete globally; securing real-world customer references and demonstrating stable operation in production environments is essential to breaking into markets dominated by NVIDIA.

"A reference showing that you are actually running AI services is the most important thing in overseas business," stated Park Sung-hyun, chief executive of Rebellions.

Park Sung-hyun, Chief Executive at Rebellions

This competitive landscape underscores why AMD's Strix Halo matters beyond raw specifications. By delivering unified memory architecture, desktop-class CPU performance, and dedicated AI acceleration in a single package, AMD provides system integrators and device manufacturers with a compelling alternative to discrete GPU designs. The chip enables a new category of mobile workstations and edge AI devices that can handle demanding professional workloads and sophisticated AI inference without cloud dependency, addressing the core value proposition that Korean startups and other NPU makers are pursuing globally.