Google Open-Sources NPU IP (www.eetimes.com)

Google Research has open-sourced the Coral NPU IP (formerly “Kelvin”) and is handing it to the industry to reduce fragmentation, improve security, and accelerate edge AI adoption. Synaptics is the first production implementer, embedding the Coral NPU in its Astra SL2610 family of IoT SoCs (dual Cortex‑A55, 12 nm). Google’s strategy is to offer an IP- and standards-based stack rather than a proprietary ASIC, targeting wearable- and IoT-class SoCs before platform fragmentation becomes entrenched. The move leans on RISC‑V for openness, extensibility, and no licensing fees, and Google also open-sourced a compiler/runtime pipeline (MLIR/IREE) to lower models from multiple ML frameworks to the NPU front end. Technically, the Coral NPU today is a 4‑way superscalar 32‑bit RISC‑V CPU with a vector engine and matrix engine planned (matrix to follow once a RISC‑V profile is ratified); CHERI support is on the roadmap. Synaptics’ Torq subsystem pairs the Coral NPU (for scalar ops tightly coupled to PPA needs) with its T1 fixed‑function 1 TOPS (INT8) accelerator for transformers/CNNs, keeping common math extensions to limit fragmentation. Verisilicon is productizing and offering verification/support for the IP. For the AI/ML community this creates an open, end‑to‑end path from frameworks to silicon for low‑power edge inference and LLM-capable accelerators, lowering vendor lock‑in and encouraging portable, secure edge deployments.