Seven years ago, Microsoft and Intel forged an alliance that would quietly reshape where and how artificial intelligence runs on Windows. The partnership paired Windows Machine Learning—Microsoft’s nascent on-device inferencing runtime—with Intel’s Movidius Myriad X, a dedicated vision processing unit built for low-power AI workloads. It wasn’t just another hardware-software integration. It was an architectural bet that on-device AI would one day become as essential as the CPU itself.

Today, Windows ML sits at the heart of Microsoft’s local AI strategy, powering everything from real-time video effects to Copilot+ PC experiences. But understanding the Myriad X chapter is key to grasping why Windows now treats neural processing as a first-class system resource. This is the story of that shift, what it meant for developers and users, and how it set the stage for the AI PCs shipping today.

What Windows ML Actually Is

Windows ML is the unified, high-performance inferencing framework that ships with Windows. Officially, it’s a Windows-supported and maintained copy of ONNX Runtime, available as a system-wide component or bundled with apps. It lets developers run trained models locally on Windows devices—no cloud required—using ONNX as the standard model format.

Microsoft designed the runtime to abstract away hardware complexity. When an app calls Windows ML, the system automatically picks the best available execution provider (EP) for the device’s silicon: CPU, GPU, or NPU. These EPs are supplied by silicon vendors and updated via Windows Update, so developers don’t need to bundle accelerator-specific SDKs or worry about device fragmentation.

The official documentation emphasizes on-device privacy, offline capability, and “to-the-metal performance” that rivals dedicated toolkits like TensorRT or Qualcomm’s AI Engine Direct. It supports x64 and ARM64 architectures, and requires Windows 11 version 24H2 or later for hardware-optimized NPU execution providers. Older Windows versions can still use CPU and GPU inferencing through DirectML.

The Myriad X: A Vision Processing Unit Ahead of Its Time

Intel’s Movidius Myriad X, launched in 2017, was a specialized piece of silicon with an almost singular focus: running computer vision and deep neural network inference at the edge, inside power envelopes measured in single watts. Its architecture was a stark contrast to general-purpose CPUs and GPUs, packing a dedicated Neural Compute Engine alongside 16 programmable SHAVE vector cores and over 20 hardware-accelerated imaging blocks for tasks like optical flow, stereo depth, and dewarping.

Intel claimed over 1 TOPS of DNN inferencing from the Neural Compute Engine, with aggregate device-level throughput in the multiple-TOPS range. The 2.5 MB of on-chip memory and rich MIPI camera interfaces made it a natural fit for drones, security cameras, AR headsets, and other sensor-laden devices where latency and battery life are paramount.

It was not a general-purpose AI engine. The Myriad X was tuned for real-time perception workloads—exactly the kind of always-on, low-latency inferencing that would become table stakes for modern Windows features.

How Windows ML and Myriad X Came Together

In early 2018, Microsoft announced that Windows ML would support Intel’s Myriad X VPU as a first-class execution provider. The pitch was simple: Windows ML would treat the VPU like any other accelerator, routing compatible ONNX models to it automatically. For an app developer, that meant shipping one package and leaning on the OS to exploit the hardware.

Behind the scenes, Intel supplied a runtime path—an execution provider—that plugged into Windows ML. When a device had a Myriad X chip, Windows could run vision models directly on the VPU, avoiding the CPU and GPU entirely. This offload promised not only faster inference but also dramatic power savings, particularly for sustained workloads like camera-based object detection or facial recognition.

The partnership was announced with concrete use cases: personal digital assistants that understand scenes in real time, on-device image search and tagging, biometric authentication that keeps data local, and smart camera features like auto-framing and background segmentation. These were not far-off visions; they were scenarios that the Myriad X’s imaging accelerators and low power draw made immediately feasible.

Why a Dedicated AI Chip Changes the Game

The move to a dedicated VPU wasn’t about raw TOPS alone. It was about the interplay of latency, power, and system load. In a traditional CPU- or GPU-only laptop, continuous vision inferencing would either drain the battery rapidly or cause the fans to spin up. A VPU like Myriad X could run the same task at a fraction of the power, often below one watt, while leaving the main processors free for user-facing work.

This matters for any feature that runs in the background: an always-on camera that detects when you walk away from your PC and locks it; a video call app that keeps you centered in the frame without maxing out the GPU; a factory kiosk that inspects parts in real time without overheating. In all these cases, the VPU delivers inference-per-watt that CPUs and GPUs can’t match at low throughput levels.

The Windows ML design capitalizes on this by letting the system dynamically choose the best executor. A casual photo search might run just fine on the CPU; a live translation overlay on your webcam feed may require the VPU. The OS makes that call transparently.

Developer Reality: Not Quite Plug-and-Play

While Windows ML and Myriad X abstracted hardware, they didn’t eliminate the need for developer care. ONNX provided a standard model format, but mapping a model efficiently to the VPU’s Neural Compute Engine still required attention to operator support, quantization, and topology. Intel’s SDKs offered model converters and performance tuning tools, but developers who ignored them often saw disappointing results.

Practical optimization steps included exporting models to ONNX, validating correctness on CPU, profiling on actual VPU hardware, and quantizing to 8-bit integer precision—a critical step because VPUs typically deliver peak efficiency with low-bit quantization. Blind reliance on vendor TOPS numbers, community members warned early on, could lead to “wildly inaccurate latency and battery estimates.”

The community also flagged real-world friction: early driver and firmware bugs, occasional thermal management quirks, and the fact that not all ONNX operators were accelerated equally on the VPU. Some models fell back to the CPU for certain layers, erasing the power advantage. Microsoft’s evolving EP model and automatic provider updates later smoothed many of these rough edges, but the lesson was clear: heterogeneous hardware demands rigorous profiling.

Enterprise Stakes: Policy, Privacy, and Procurement

For IT departments, the VPU integration raised immediate questions about fleet management. Windows Update became the primary channel for VPU drivers and EPs, which meant that a flawed driver could impact many devices at once. Community guidance from the time stressed staged rollouts, firmware validation, and close monitoring of Windows reliability logs after updates.

Privacy also emerged as a double-edged sword. On-device inferencing kept sensitive data local, a clear improvement over cloud-based AI services. Yet features that indexed user activity for “recall” or made local biometric data searchable introduced new governance challenges. Administrators needed to audit what was stored on the device, how it was encrypted, and whether any cloud fallbacks existed. The consensus: on-device AI improves privacy posture, but not without deliberate policy controls.

Device procurement teams faced a harder calculation. A general office worker might never benefit from a VPU, while a frontline retail device running continuous vision would see immense value. The advice: pilot hardware in representative environments and avoid blanket deployment of VPU-equipped PCs across the board.

The Limitations Intel and Microsoft Didn’t Lead With

Vendor marketing often touted the Myriad X’s TOPS figures as a proxy for all AI performance. Community analysts pushed back, noting that TOPS capture only peak theoretical multiply-accumulate throughput, not real-world bottlenecks like memory bandwidth, operator coverage, or thermal limits. Two accelerators with identical TOPS could deliver vastly different inference times on the same model.

Another caution was ecosystem fragmentation. Each silicon vendor’s EP came with its own quirks—operator support, precision preferences, even different acceptable ONNX opsets. While Windows ML’s automatic selection helped, developers targeting multiple device classes frequently needed to maintain alternative model variants or graceful fallbacks. This remains true today, even as NPUs proliferate.

Finally, the driver maturity curve was steep. Early VPU-enabled laptops occasionally exhibited stability issues, and some enterprise users reported that critical features worked only after specific OEM firmware updates. Microsoft’s evergreen model has since matured, but the lesson stood: treat new AI hardware deployments like any other platform shift, not a simple feature toggle.

From Myriad X to the NPUs of 2025

The Myriad X collaboration was an early beacon of where Windows was heading. Today, Windows ML is no longer tied to a single partner or chip; it supports NPUs from Qualcomm, Intel, AMD, and others. The execution provider model has expanded to cover every major silicon class, and the runtime now delivers performance that Microsoft claims matches dedicated SDKs like TensorRT.

Windows 11’s Copilot+ PCs, with their 40+ TOPS NPUs, are the direct descendants of the architecture that Myriad X helped prove. The bets placed in 2018—on ONNX portability, heterogeneous offload, and OS-managed acceleration—are now mainstream. Developers who adopted Windows ML early found themselves with a foundation that scaled from a tiny vision chip to a powerful AI engine without rewriting their apps.

But the core truths haven’t changed. Dedicated AI hardware still wins on energy efficiency and sustained throughput. Privacy still demands careful implementation. And TOPS numbers still make for better marketing than engineering. The Myriad X era taught the industry those lessons, and they remain as relevant as ever.

What to Do Now: Practical Advice for the Modern AI PC Era

For developers building AI features on Windows today, the path is clearer than in the Myriad X days, but not effortless. Start with ONNX, profile on every target hardware class, and quantize aggressively. Don’t assume the OS will magically optimize everything—use Windows ML’s diagnostics to measure time-to-first-token and battery impact. Fallbacks are still necessary; not every device has an NPU, and even when present, not every operator is accelerated.

IT teams should adopt a lifecycle mindset. Test EP and driver updates in pilot rings, set group policies for AI-powered features that touch user data, and demand clear driver support SLAs from OEMs. The hardware is powerful, but it introduces a new dependency chain into the Windows servicing stack.

End users, especially in privacy-conscious organizations, should lean on on-device-only modes whenever possible and verify that any cloud fallback is both transparent and controllable. Windows Hello and device encryption are essential companions to local AI, not optional extras.

The Myriad X chapter closed long ago, but its legacy shapes every Windows PC shipping today. When Microsoft and Intel decided to make a dedicated vision processor a first-class citizen of the operating system, they didn’t just accelerate a few camera features. They defined a new tier of device intelligence—one that runs locally, sips power, and respects user privacy. The VPU era may be over, but the architecture it launched has never been more alive.