What Is the Ryzen AI 9 HX 370 NPU?
AMD’s Ryzen AI platform is a family of neural‑processing units (NPUs) that sit alongside the company’s well‑known CPU and GPU blocks. The “Ryzen AI 9 HX 370” is the latest iteration aimed at premium laptop designs. It is not a standalone chip you can buy off the shelf; instead, it is integrated into AMD’s HX‑series mobile processors (for example, the Ryzen 9 7945HX) and works as a dedicated accelerator for artificial‑intelligence workloads.
In practice, the HX 370 NPU provides a hardware path for tasks such as on‑device speech recognition, image enhancement, and real‑time translation. By offloading these operations from the CPU and GPU, the system can achieve lower latency, reduced power draw, and smoother multitasking – all without relying on an internet connection.
How the HX 370 Is Built
AMD has not disclosed every transistor‑level detail, but the public technical brief outlines a few key architectural points:
- Matrix‑multiply engine: A core array of multiply‑accumulate units optimized for the 8‑bit and 16‑bit integer math that dominates deep‑learning inference.
- On‑chip SRAM cache: A modest amount of high‑speed memory sits directly next to the compute array, reducing the need to fetch data from DRAM during inference.
- Unified memory interface: The NPU shares the same memory bus used by the CPU and GPU, allowing it to access the same data structures without costly copies.
- Power‑gating controls: The silicon can be turned off completely when no AI workload is present, helping laptops stay within thin‑and‑light thermal envelopes.
These building blocks are similar in spirit to the Tensor cores found in Intel’s Xe‑HPG GPUs or the NPU blocks in Qualcomm’s Snapdragon platform, but AMD’s implementation is tightly coupled to its Zen 4 CPU architecture, which gives software developers a single programming model across CPU, GPU, and NPU.
Software Stack: From Model to Metal
The biggest hurdle for any accelerator is getting developers to actually use it. AMD addresses this with a layered software stack that starts with open‑source frameworks and ends with low‑level drivers.
At the top, developers can use familiar tools such as PyTorch, TensorFlow, or ONNX Runtime. AMD provides a set of conversion utilities that translate a trained model into a format the Ryzen AI runtime understands. The conversion process inserts “hints” that tell the runtime which layers can be executed on the NPU and which should fall back to the GPU or CPU.
Under the hood, the Ryzen AI runtime leverages the open‑source ROCm (Radeon Open Compute) ecosystem. This means that developers who have already written GPU‑accelerated code can often reuse parts of their pipeline, while the runtime automatically schedules NPU‑friendly kernels. The driver layer communicates with the silicon via a lightweight ABI, exposing performance counters that can be read with standard Linux tools.
For Windows users, AMD ships a companion SDK that integrates with Visual Studio and the Windows ML (WinML) API, allowing developers to target the HX 370 with a few extra configuration steps. The result is a workflow that feels like “just another device” rather than a specialized hardware add‑on.
Real‑World Use Cases in Laptops
While the HX 370 is a piece of silicon, its value is measured by the experiences it enables. Here are a few scenarios where the NPU makes a tangible difference:
- Voice assistants: Local speech‑to‑text conversion can run in under a hundred milliseconds, eliminating the need for a round‑trip to the cloud and preserving user privacy.
- Camera enhancements: Real‑time denoising, HDR tone‑mapping, and background‑blur for video calls happen on‑device, keeping the video feed smooth even on bandwidth‑constrained networks.
- Gaming overlays: AI‑driven upscaling (similar to AMD’s FidelityFX Super Resolution) can be applied to a secondary UI layer without taxing the main GPU.
- Productivity tools: Features such as live translation in document editors or AI‑generated summaries in note‑taking apps run locally, delivering instant feedback.
Because the NPU is integrated into the same package as the CPU and GPU, laptop manufacturers can keep the overall footprint small, which is critical for ultra‑thin designs that still want AI capabilities.
Power and Performance: What to Expect
AMD does not publish a single “TOPS” (trillion operations per second) figure for the HX 370, but independent testing from a handful of early adopters gives us a sense of its behavior. In synthetic inference benchmarks, the NPU typically delivers a 2‑3× speedup over a comparable CPU‑only run for models that fit within its on‑chip cache. When the same workload is shifted to the integrated GPU, the NPU still wins on power efficiency, consuming roughly a third of the energy for a similar throughput.
The key takeaway is that the HX 370 shines when the model is modest in size (a few megabytes) and the latency budget is tight. Larger models that exceed the NPU’s memory capacity will automatically fall back to the GPU, which still provides a performance boost compared to the CPU alone. This dynamic fallback is handled by the Ryzen AI runtime without developer intervention.
How It Stacks Up Against the Competition
The mobile AI accelerator market is crowded. Intel’s Meteor Lake CPUs ship with a built‑in X‑evo NPU, while Apple’s M‑series chips include a 16‑core Neural Engine. Qualcomm’s Snapdragon platforms have long featured an AI Engine, and Google’s Tensor SoCs are another major player.
What differentiates the HX 370 is its integration strategy. Rather than being a separate silicon block on a different package, it shares the same silicon die and memory interface as AMD’s Zen 4 cores and RDNA 3 graphics. This shared‑die approach reduces latency between the three compute engines and simplifies board design for OEMs.
In practice, the performance gap is not a binary “faster or slower” comparison but rather a trade‑off between ecosystem support and power envelope. Developers who already work within the ROCm ecosystem may find the Ryzen AI stack more approachable, while those invested in Apple’s Core ML or Google’s TensorFlow Lite may prefer the native tools of those platforms.
Looking Ahead: The Future of On‑Device AI
AMD’s roadmap suggests that the HX 370 is just the first step in a broader AI strategy for the PC market. Upcoming generations are expected to increase the number of matrix‑multiply lanes, expand on‑chip SRAM, and support higher‑precision data types such as bfloat16. Those upgrades will open the door for more demanding workloads, including on‑device natural‑language generation and advanced computer‑vision pipelines.
From a consumer perspective, the real impact will be felt when software vendors embed AI features directly into the operating system or popular applications. Imagine a Windows desktop that can intelligently rearrange windows based on your workflow, or a video editor that suggests cuts and transitions in real time—all powered by the laptop’s own NPU.
For developers, the growing maturity of the Ryzen AI SDK means a lower barrier to entry. As more open‑source projects add support for the HX 370, we’ll likely see a virtuous cycle: more apps using the NPU → more performance data → better tooling → even more apps.
In short, the Ryzen AI 9 HX 370 NPU is a modest but strategically important piece of AMD’s vision for AI‑first computing. By marrying a dedicated inference engine with the existing strengths of Zen 4 and RDNA 3, AMD offers laptop makers a way to add genuine on‑device intelligence without sacrificing form factor or battery life. As the ecosystem matures, the HX 370 could become the hidden workhorse that makes everyday AI feel seamless rather than spectacular.
