Qualcomm’s Snapdragon X2 Elite SoC brings significant advancements in both AI processing and device security, with the Hexagon NPU and Guardian technology standing out as key innovations. This deep dive explores the evolution and capabilities of the Hexagon NPU and the controversial aspects of Guardian. The Hexagon NPU, now officially designated as NPU6, marks a major leap from its predecessors—the Hexagon DSP and earlier Hexagon Processor. While the name change reflects a shift in focus toward AI workloads, the underlying architecture has evolved to handle increasingly complex tasks. The new NPU is built around three core components: the scalar unit, vector unit, and matrix unit, each optimized for specific types of computation. The scalar unit has doubled in thread count, now supporting 12 threads compared to the previous 6. Each thread uses a 4-wide VLIW architecture and includes advanced features like multi-level branch prediction, user-mode DMA for reduced latency, and hardware synchronization. The unit also features two master ports and supports 64-bit DMA, even though the internal core remains 32-bit—highlighting a design choice that balances performance and efficiency. The vector unit has also been significantly upgraded, now featuring eight threads, up from four. Each engine can process four 128-bit SIMD vectors per cycle and supports FP8 and BF16 precision, essential for modern AI models. This results in a claimed 143% increase in vector throughput—matching the scalar unit’s improvement, which is no coincidence. The increased bandwidth and larger L2 caches further enhance performance, with a 127% boost in bus bandwidth to the NPU. The matrix unit is the most powerful addition. It supports 2-bit weights and includes dedicated weight and activation caches, along with direct access to vector tightly coupled memory. It operates on a separate power rail for efficiency and is controlled by one or two scalar threads, leaving the rest free for other tasks. The new architecture shifts from a 6:4:1 scalar:vector:matrix thread ratio in the X1 to 12:8:1 in the X2, with all components larger, faster, and more capable. AI model utilization data shows the matrix unit is the most active, but the scalar and vector units are also heavily used. The design allows for dynamic resource reallocation, minimizing idle time and maximizing throughput. The NPU delivers a claimed 80 TOPS of AI performance, up from 45 TOPS in the X1—though real-world performance will depend on software optimization. On the security side, Qualcomm’s Guardian technology aims to address long-standing enterprise needs for device management, remote tracking, locking, and wiping. It’s similar to Intel vPro but includes a built-in cellular modem (X75 for 5G or a 4G IoT modem for basic functions), enabling always-on connectivity. However, this comes with major drawbacks. First, Guardian is Windows-only, limiting its appeal in multi-platform environments. Second, it requires a cellular connection for full functionality, including location tracking. While it can work over Wi-Fi, location data is not available without a cellular link. This adds cost and complexity, as the modem is mandatory and cannot be disabled—creating a persistent attack surface. The service is not free. Qualcomm charges OEMs about $20 per year, which is passed on to consumers. For a three-year period, this adds $60 to the system’s cost, significantly increasing the total price. This forced inclusion of a security feature with no opt-out option raises concerns about transparency and user control. Even more troubling is the integration of Microsoft’s Pluton security block. Pluton is a third-party, remotely updatable CPU that can access any part of the system silently. According to sources familiar with its deployment, this creates a critical vulnerability. Unlike Intel’s approach, which avoids such centralized access, Pluton enables remote manipulation with no user oversight. Combined with the always-on cellular modem, Guardian creates a system that is both expensive and inherently insecure. While the intent behind Guardian is sound—enterprise-grade device management—the execution undermines trust and adds unnecessary risk. In summary, the Hexagon NPU delivers impressive AI performance and architectural flexibility. But Guardian, despite its potential benefits, introduces serious security and cost concerns. For users and enterprises alike, the trade-offs may outweigh the advantages—especially when better alternatives exist.