Intel Pushes On-Device AI with New NPUs
Intel is promoting its new neural processing units (NPUs) for AI processing at the edge, emphasizing that on-device AI is about privacy and latency, not just performance. The push validates the industry-wide trend toward localized AI workloads, a core tenet of Apple's product strategy.
Intel's new NPUs are integrated into its "Core Ultra" processors, codenamed Meteor Lake, which represent a major architectural shift. The chips utilize a disaggregated "tile" design with separate CPU, GPU, SoC, and I/O chiplets connected using Foveros 3D packaging technology. This modular approach allows Intel to mix and match process nodes for different components. The NPU itself resides on the SoC tile, a low-power island manufactured on TSMC's N6 process, while the main compute tile uses the new Intel 4 process. The NPU is designed specifically for sustained, low-power AI inference, offloading tasks like real-time background blur and noise suppression in video calls from the CPU and GPU to improve overall system power efficiency. The CPU tile is Intel's first to be manufactured on the "Intel 4" process node, the company's first time using extreme ultraviolet (EUV) lithography in high-volume production. Compared to the prior "Intel 7" process, Intel 4 can deliver over 20% higher clock speeds at the same power level or reduce power consumption by 40% at the same frequency. This move follows a long-established trend at Apple, which has included a dedicated "Neural Engine" in its silicon since the A11 Bionic chip in 2017. Apple's unified memory architecture provides a key advantage by allowing its CPU, GPU, and Neural Engine to share a single pool of high-bandwidth memory, which can accelerate AI data processing. In terms of raw performance, the Neural Engine in Apple's M4 chip is capable of 38 trillion operations per second (TOPS). For comparison, the NPU in Intel's upcoming "Arrow Lake" desktop processors is rated for 13 TOPS, highlighting the intense competition in AI acceleration hardware. The broader semiconductor industry is similarly focused on on-device AI acceleration, with both AMD and Qualcomm integrating NPUs into their latest processors. This industry-wide pivot is creating a new category of "AI PCs," which market research firm Gartner expects to account for more than half of all global PC shipments by 2026.
