Rockchip RK3668 Targets Edge AI with 44 TOPS NPU and PCIe 5.0 - LinuxGizmos.com

Rockchip's 44 TOPS NPU in the RK3668 is a deliberate positioning play, slotting between NVIDIA's Jetson Orin Nano and commodity IoT silicon to capture the underserved mid-tier inference segment. The PCIe 5.0 interface, however, is the real tell: 64 GT/s bandwidth signals multi-chip and GPU-expansion architectures, repositioning the product from dev-board territory toward edge AI servers and shifting the competitive axis entirely. Supply-chain exposure is the quiet, persistent cost. A 44 TOPS NPU almost certainly requires sub-7nm fabrication, and if that foundry capacity sits in Taiwan, China, the chip remains perpetually exposed to US export-control gray zones. The 2023 HBM restrictions demonstrated that interconnect bandwidth thresholds—not just raw compute—can trigger entity-list expansion. PCIe 5.0 sits uncomfortably close to that line, and compliance overhead will be structural, not episodic. On competition, Rockchip's edge is price: 30-40% below equivalent-performance rivals. But if NVIDIA cascades Thor-class compute down to consumer edge products, the 44 TOPS window may close within 18 months. Qualcomm and MediaTek are accelerating their own edge-AI roadmaps in parallel, eroding the price moat. The 12-24 month thesis: the competitive axis is shifting from peak TOPS to effective inference throughput per watt and full-stack model efficiency. Without a TensorRT-class software ecosystem, Rockchip's hardware specs will be commoditized rapidly, collapsing the product into pure cost competition.