Microsoft just dropped a targeted update that gives AMD-powered Copilot+ PCs a serious bump in on-device AI performance. KB5064650, which delivers version 1.2507.793.0 of the Phi Silica AI component, began rolling out automatically to Windows 11 24H2 systems in early April 2025. The update focuses squarely on systems with AMD's neural processing units, promising faster, more stable AI operations without a whisper of data leaving the hardware.

For the growing cohort of users who opted for AMD's Ryzen AI–equipped Copilot+ laptops over Qualcomm's Snapdragon X machines, this update marks a key milestone. It's the first Phi Silica component update explicitly tuned for AMD silicon, a sign that Microsoft and its silicon partners are ironing out the wrinkles in the Copilot+ ecosystem.

What Exactly Is Phi Silica?

Phi Silica is Microsoft's homegrown, NPU-optimized language model. It's a Transformer-based architecture—the same family that powers models like GPT—but shrunk and tuned to run efficiently on the dedicated neural processing units baked into Copilot+ PCs. Unlike a cloud-first model such as Copilot in Bing, Phi Silica executes entirely on-device. That means zero round-trip latency to a data center and zero chance of your prompts or responses being processed by remote servers.

The model is designed to handle many of the tasks users would normally throw at a large language model: text generation, summarization, rewriting, and contextual reasoning. But because it runs locally, Phi Silica can respond in milliseconds, powering experiences like Windows Copilot's in-line suggestions, Live Captions translation, and the controversial Recall feature—all with the privacy guarantees local processing provides.

Microsoft first introduced Phi Silica alongside the Copilot+ PC launch in mid-2024. At that time, the model was heavily optimized for Qualcomm's Snapdragon X Elite and Plus chips, which had a head start in the Windows on Arm push. AMD's Ryzen AI 300 series, with its XDNA 2 NPU capable of up to 50 trillion operations per second (TOPS), was always part of the roadmap, but software optimization lagged. KB5064650 is the first salvo in closing that gap.

What's Inside KB5064650?

The update itself is lean, but its impact is felt across any application that taps into Phi Silica through the Windows Copilot Runtime. According to the official Microsoft support document, the update brings three categories of improvement:

  • Performance enhancements: The model's inference engine has been recompiled to take advantage of AMD's specific NPU instruction set, reducing token-generation latency by up to 30% on measured benchmarks. Users should experience snappier text completions, faster on-the-fly translations, and smoother multitasking when AI features are active.
  • Stability improvements: Early adopters of AMD Copilot+ PCs reported occasional hangs or silent failures of AI features, particularly after resuming from sleep. This update includes memory management fixes that prevent model corruption in low-power states.
  • Compatibility updates: As Windows 11 24H2 continues to evolve—with new APIs for AI and security layers like Pluton tapping the NPU—Phi Silica now aligns with the latest system interfaces, reducing the chance of feature breakage after future Patch Tuesday drops.

The version number 1.2507.793.0 follows a calendar-based scheme: 2025, July, build 793. That suggests this code was finalized in July 2025, meaning it went through a rigorous internal validation cycle before public rollout.

How to Get It

There's no need to chase this update manually. Microsoft is delivering KB5064650 through Windows Update as an automatic download for all eligible devices. To confirm you've received it, head to Settings > Windows Update > Update history and look for "Phi Silica AI component update (version 1.2507.793.0) for AMD-powered systems."

One critical prerequisite: your machine must already have the latest cumulative update for Windows 11 24H2 installed. That's because Phi Silica hooks into kernel-level threads that change with servicing stack updates. If you've been deferring monthly patches, Windows Update will install the necessary cumulative fix before KB5064650 appears.

This is a server-side phased rollout, so not all AMD Copilot+ PCs will see it on day one. Microsoft typically uses a gradual deployment ring model, starting with a small percentage of devices and expanding over the course of a week or two. If you're impatient, you can try checking for updates manually, but the "seeker" experience isn't guaranteed for driver-class updates like this.

The AMD Copilot+ Experience: Before and After

When the first AMD Copilot+ laptops from ASUS, HP, and Lenovo shipped in late 2024, reviews were mixed. While the raw CPU and GPU performance often outpaced Qualcomm alternatives—especially for x86-native creative apps—the AI features felt half-baked. Live Captions might take an extra second to kick in; Cocreator in Paint could stutter; and the Phi Silica model sometimes spat out replies that felt a beat too slow compared to the instantaneous responses on ARM-based machines.

KB5064650 addresses the software maturity gap. In practice, this means:

  • Windows Studio Effects, which use the NPU for background blur, gaze correction, and auto-framing, now engage with less CPU overhead, preserving battery life during video calls.
  • Live Captions translation – an accessibility feature that transcribes and translates any audio playing through the system – sees reduced latency, making it usable for real-time conversations.
  • Recall (when enabled) performs faster semantic search across your timeline, as the Phi Silica model more efficiently indexes and retrieves context.
  • Developer scenarios using the ONNX runtime with the NPU execution provider (like local chatbots or custom AI pipelines in Visual Studio) benefit from the updated Phi Silica as a ready-made, production-quality language model.

Beyond the raw speed, the stability improvements are just as welcome. Early forum chatter on windowsnews.ai indicated that some users had resorted to disabling Phi Silica features entirely after encountering blue screens tied to the NPU driver. Those reports have quieted following the update, suggesting the underlying conflicts have been addressed.

Community Pulse: Cautious Optimism

Within the enthusiast forums, the reception has been largely positive, though seasoned users are taking a wait-and-see approach. "Finally, my ASUS Zenbook S 16 (Ryzen AI 9 HX 370) feels like a true Copilot+ machine," one anonymous commenter wrote on a related thread. Others expressed relief that Microsoft hasn't abandoned AMD optimization efforts in favor of the Qualcomm exclusivity window that marked the platform's debut.

Not all feedback is glowing. A handful of users have noted that the update's installation can fail silently if the prerequisite cumulative update hasn't been applied, leading to confusion. Microsoft's decision not to bundle KB5064650 into the main Windows Update package means Joe Average might not even realize he's missing out until he digs into update history.

There's also the elephant in the room: app support. Phi Silica is just one piece of the AI puzzle. Even with a sped-up language model, the value proposition of a Copilot+ PC hinges on software developers embracing the Windows Copilot Runtime. For now, the first-party Microsoft apps (Paint, Photos, Camera) account for the bulk of AI features, and no amount of NPU tuning can compensate if third-party ISVs drag their feet.

On-Device AI: The Strategic Play

KB5064650 isn't just a bug fix; it's a strategic signal. Microsoft is betting that on-device AI, rather than purely cloud-based solutions, will define the next decade of personal computing. The rationale is threefold: latency, privacy, and cost.

A local language model eliminates the network round-trip, enabling truly real-time interactions. It also keeps sensitive data under the user's control—critical for enterprise adoption of AI features like Recall, which records screen activity. Finally, offloading AI inference to the NPU reduces pressure on Microsoft's Azure infrastructure, a non-trivial expense when serving hundreds of millions of Copilot queries.

AMD's Ryzen AI hardware plays a crucial role in this vision. With the Ryzen AI 300 series (Strix Point), AMD became the first x86 vendor to ship an NPU exceeding the 40 TOPS Copilot+ threshold, matching and sometimes exceeding Qualcomm's Snapdragon X NPU performance. Intel's Lunar Lake has since joined the fray, but AMD's head start and aggressive pricing in the notebook market give it a significant installed base. Optimizing Phi Silica for AMD is therefore not a niche concern—it's essential for the Copilot+ brand to avoid being seen as a Qualcomm-only ghetto.

This update also hints at what's coming with Windows 11 25H2 (expected late 2025). Leaked builds suggest deeper integration of local models into the shell, including 'AI widgets' that would demand consistent NPU performance across architectures. KB5064650 lays the groundwork by ensuring the Phi Silica pipeline is battle-tested on AMD silicon before those features hit mainstream.

How It Compares: AMD vs. Qualcomm vs. Intel

While Microsoft doesn't publish benchmark comparisons between NPUs, independent testing has started to paint a picture. The original Phi Silica implementation on Snapdragon X delivered sub-10ms token latency for small prompts. Pre-update AMD systems often hovered in the 20–30ms range. With KB5064650, anecdotal reports suggest a drop to the 12–18ms band, closing the gap significantly.

Intel's Core Ultra 200V (Lunar Lake) with its 48 TOPS NPU remains untested in the wild for Phi Silica, as the current component targets only AMD and Qualcomm. That suggests separate optimization paths for each silicon vendor—a fragmented reality that Microsoft's unified Windows Copilot Runtime aims to mask.

Users shouldn't expect the kind of raw generative throughput you'd get from an NVIDIA RTX GPU running a full-sized LLM. Phi Silica is designed for efficiency, not brute force. Its token-per-watt efficiency, however, far outstrips any discrete GPU, making it the right tool for always-on, background AI tasks.

Potential Caveats and What to Watch Out For

No update is without risk. A few considerations for those installing KB5064650:

  • Storage footprint: Phi Silica model files reside on the system drive and can consume several gigabytes. The update may prompt a model redownload if a prior version is corrupted, potentially eating into free space on 256GB base-model laptops.
  • Thermal behavior: Sustained NPU workloads can cause additional heat, especially in thin-and-light designs. While modern AMD chips throttle gracefully, users pushing AI features alongside heavy gaming or rendering may notice a performance dip.
  • Driver dependency: The Phi Silica component interacts closely with the AMD NPU driver (part of the Adrenalin or chipset package). An outdated GPU driver can undermine the gains. Make sure you're on the recommended OEM driver rather than a generic Windows-inbox version.

Microsoft's documentation notes that the update is available only for "AMD-powered systems"—meaning systems with an integrated AMD NPU that meets the Copilot+ spec. Older AMD laptops with Ryzen 6000 or 7000 series (which lack a 40+ TOPS NPU) won't see the update. That line in the sand protects the Copilot+ brand from dilution but may frustrate owners of very recent non-Copilot+ AMD machines.

The Road Ahead

KB5064650 is a stepping stone, not a destination. Microsoft's AI roadmap through 2026 includes larger on-device models, multimodal capabilities (combining vision and text), and a potential Copilot+ certification for desktop Ryzen 9000 processors with discrete NPUs. The Phi Silica update cadence will likely accelerate, with a new version expected alongside each Windows feature drop.

For now, AMD Copilot+ PC owners have a practical reason to hit that check for updates button. The boost may not revolutionize their workflow overnight, but it delivers on the promise of a balanced, privacy-focused AI PC that doesn't need to phone home for every query. In a computing landscape increasingly defined by AI, that local edge matters more than ever.