AMD Ryzen AI 9 HX 370

The AMD Ryzen AI 9 HX 370 is the flagship silicon of the "Strix Point" architecture, representing a significant pivot in AMD’s mobile strategy toward dedicated AI acceleration. Built on TSMC’s 4nm process, this SoC (System on a Chip) integrates the Zen 5 CPU architecture with the new XDNA 2 NPU, specifically designed to handle persistent AI workloads without offloading to a discrete GPU. For developers and engineers, this chip marks the transition of the "AI PC" from a marketing term to a functional reality for local model execution.

Positioned as a premium mobile processor, the Ryzen AI 9 HX 370 competes directly with the Intel Core Ultra 200V (Lunar Lake) and the Apple M3/M4 series. While its predecessors focused primarily on raw CPU throughput, the HX 370 is engineered for the balanced demands of local AI development and agentic workflows. By exceeding the 40 TOPS requirement for Microsoft’s Copilot+ PC program with its 50 TOPS NPU, it provides a stable, high-performance target for Windows-based AI applications and Linux-based inference stacks.

The AMD Ryzen AI 9 HX 370 AI inference performance is driven by a three-pillar compute architecture: the CPU, the iGPU, and the NPU. While the total platform offers 80 TOPS, the distribution of these resources is what dictates how practitioners should deploy models.

The standout feature is the XDNA 2 NPU, delivering 50 TOPS of INT8 performance. Unlike previous generations, XDNA 2 supports "Block FP16" data types, which allows for near-INT8 speeds with the accuracy of FP16. This is critical for running local LLMs where quantization loss can degrade reasoning capabilities. The NPU is optimized for low-power, "always-on" tasks such as background noise suppression, image generation, and lightweight agentic loops, operating within a lean 28W TDP.

For AI workloads, memory is the primary bottleneck. The Ryzen AI 9 HX 370 supports LPDDR5X-8000, providing a maximum memory bandwidth of 128 GB/s. While this is lower than dedicated desktop GPUs, it is highly competitive for a mobile SoC.

• VRAM Capacity: Since this is an APU, the VRAM is shared with system RAM. In configurations with 64GB of LPDDR5X, users can typically allocate up to 16GB or even 32GB (depending on BIOS settings) to the Radeon 890M iGPU.
• Why it matters: This shared pool allows the HX 370 to load models that would be impossible on mobile GPUs with only 8GB of dedicated VRAM.

The integrated Radeon 890M, featuring 16 Compute Units on the RDNA 3.5 architecture, acts as the heavy lifter for parallel processing. When the NPU isn't supported by a specific framework (like certain custom PyTorch kernels), the iGPU provides a robust fallback for GPGPU acceleration, benefiting from the high-speed 128 GB/s memory bus.

When evaluating the AMD Ryzen AI 9 HX 370 for local LLMs, performance is dictated by quantization. This hardware is the "sweet spot" for models in the 3B to 14B parameter range.

Using frameworks like LM Studio (via llama.cpp) or AMD’s own Ryzen AI Software (leveraging ONNX Runtime), you can expect the following:

• Llama 3.1 8B (Q4_K_M quantization): This is the primary target for this chip. You can expect a highly usable 15-20 tokens per second (t/s). At Q8 quantization, speed may drop toward 8-12 t/s, but the 128 GB/s bandwidth keeps the experience fluid.
• Mistral 7B / Nemo 12B: These models fit comfortably within the shared memory pool. The 12B models run effectively at 4-bit or 5-bit quantization, making them excellent choices for local RAG (Retrieval-Augmented Generation) tasks.
• Phi-3.5 Mini / Qwen 2.5 3B: These smaller models run at exceptional speeds, often exceeding 40+ t/s, making them ideal for real-time agentic workflows where low latency is required.
• DeepSeek-R1 (Distilled 7B/14B): The 7B version runs natively with high performance. The 14B version is viable at 4-bit quantization but will begin to test the limits of the 128 GB/s bandwidth.

The 50 TOPS NPU is particularly well-suited for Stable Diffusion XL Turbo or Whisper v3 (speech-to-text). Generating a 512x512 image via SDXL Turbo on the iGPU/NPU hybrid path typically takes under 2 seconds, making it a viable tool for on-the-fly asset generation in developer workflows.

The AMD Ryzen AI 9 HX 370 is one of the best AI chips for local deployment in a mobile form factor. It is specifically designed for:

• AI Application Developers: If you are building software that leverages the Windows Copilot+ ecosystem or utilizing the Ryzen AI library to optimize models for XDNA, this is your primary development target.
• Local Agentic Workflows: For practitioners running local agents (e.g., AutoGPT or CrewAI) that require a persistent LLM in the background, the HX 370’s energy efficiency allows these agents to run without significant battery drain or thermal throttling.
• Privacy-Conscious Researchers: The ability to run a 14B parameter model entirely offline with 32GB+ of accessible memory makes this an ideal platform for processing sensitive datasets that cannot be sent to the cloud.
• Edge AI Prototyping: Engineers designing systems for edge deployment can use the HX 370 to prototype models that will eventually run on smaller AMD Versal or embedded Ryzen chips, as they share the XDNA architecture.

The Ryzen AI 9 HX 370 generally wins on raw multi-core compute with its 12-core/24-thread configuration (4 Zen 5 + 8 Zen 5c), compared to Intel’s 8-core design. While Intel’s Lunar Lake NPU is highly competitive at 48 TOPS, AMD’s XDNA 2 architecture currently holds a slight edge in INT8 throughput (50 TOPS). For developers who also need high CPU performance for compiling code alongside AI inference, the Ryzen is the stronger choice.

Apple’s M-series chips benefit from higher unified memory bandwidth (e.g., 100 GB/s on base M3, significantly higher on Pro/Max). However, the Ryzen AI 9 HX 370 offers a more flexible environment for developers who require Windows or Linux-specific toolchains. While the M4's Neural Engine is formidable, the XDNA 2 NPU is more accessible for developers using the ONNX and OpenVINO (via cross-compatibility) ecosystems in a PC environment.

When selecting best hardware for local AI agents 2025, the HX 370 stands out because it doesn't require a power-hungry discrete GPU to be effective. For the first time in the Windows ecosystem, a mobile SoC can handle medium-sized LLMs with the efficiency previously reserved for specialized silicon. If your workflow involves 7B to 14B parameter models and you require a mobile workstation, the Ryzen AI 9 HX 370 is currently the performance leader in the integrated AI PC category.