The HNPU is the primary AI compute engine for on-device LLM inference, Copilot+ compliance, real-time upscaling, and heavy background AI workloads. Manufactured on Intel 04A—the same cutting-edge node as the HPiGPU—the HNPU delivers unprecedented neural network density and power efficiency for next-generation client compute.
The HNPU employs a distributed tile-array architecture optimized for sustained AI inference and model execution. Each tile operates as an independent compute island, enabling fine-grained parallelism across neural network operations without introducing latency penalties or synchronization bottlenecks.
INT2 matrix multiplication is the cornerstone of the HNPU's breakthrough throughput. By reducing datapath width to 2-bit integer operations, the HNPU achieves 4× the effective throughput of INT4 at the same silicon area—critical for quantized LLM inference where model weights occupy the vast majority of memory bandwidth. This precision mode unlocks aggressive model compression without sacrificing inference latency.
The 32 MB on-die scratchpad SRAM eliminates external memory round-trips for model weight staging and temporary activations, drastically reducing power consumption during inference. The scratchpad acts as a unified working memory for tile-to-tile data movement, enabling high-bandwidth, low-latency model execution.
The Hardware Thread Director coordinates task routing between HNPU and LPNPU, escalating background AI workloads upward only when power budget and thermal headroom permit sustained execution.
The HNPU seamlessly integrates with the Elementalist GPU's XeSS 5.0 upscaling engine, decoupling AI-driven temporal frame reconstruction from traditional rasterization latency. While the Elementalist handles pixel rasterization, the HNPU executes low-resolution neural networks for temporal motion estimation, feature synthesis, and perceptual enhancement.
Model inference runs entirely on the HNPU's tile array, freeing GPU XMX5 matrix units for parallel compute offloading. This co-execution model enables real-time upscaling at 4K+ resolutions without forcing games to choose between AI enhancements and maximum frame rate.
XeSS 5.0 is production-ready across AAA titles, with dedicated support in Unreal Engine 5.5 and a growing ecosystem of validated game publishers. The HNPU's INT2 quantization pipeline ensures model size stays under 64 MB, fitting entirely in tile SRAM for zero-copy execution.