Chiplets Interface Protocol (ChIP) for Ultra-Large-Scale Applications
Arvin Delavari, Boris Vaisband · IEEE Journal on Emerging and Selected Topics in Circuits and Systems · 2025
As computational workloads continue to grow, heterogeneous integration of chiplet-based systems is becoming critically important for data-intensive applications such as high-performance computing, large language models, and artificial intelligence. Scaling to ultra-large-scale (ULS) systems introduces, however, significant communication challenges due to the limitations of network architectures and packaging technologies. Efficient data transfer across a network of thousands of chiplets remains a critical bottleneck. A robust, low-latency, area- and energy-efficient communication architecture for ULS system named Chiplet Interface Protocol (ChIP) is proposed in this work. ChIP supports burst transfers and out-of-order transactions while leveraging the simple universal parallel interface for chips (SuperCHIPS)—a simple area- and energy-efficient streaming channel at the physical layer. Evaluated on a wafer-scale platform, ChIP was compared to state-of-the-art (SOTA) chiplet-based interfaces, including LIPINCON, BoW, UCIe, and AIB, in performance, hardware efficiency, and unified signaling figures of merit. From the comparison results, ChIP significantly outperforms the SOTA alternatives (5.53× better) in bandwidth per shoreline, reaching 2.2 Tbps/mm in pipelined mode and up to 7.3 Tbps/mm in burst transactions. In addition, the transceiver area per link in ChIP is 485 μm2—46.1% smaller than the best SOTA alternative—while achieving 0.38–0.53 pJ/bit energy and 1 ns latency in 45 nm CMOS over a 0.5 mm link, with efficiency sustained across longer channels and varied packaging due to minimal handshaking and optimized point-to-point specifications. The performance of ChIP is evaluated across multiple network configurations on a fine-pitch integration platform, and also for a customized hybrid topology, referred to as the network on interconnect fabric (NoIF), that is introduced and analyzed in this work. The architecture of the NoIF forms the foundation for ULS computing platforms, delivering exceptional results as compared to SOTA solutions. The superior hardware efficiency and advanced inter-chiplet communication features of ChIP position this proposed protocol as an ideal candidate for chiplet communication in ULS architectures.