Baidu and Samsung to Collaborate on Advanced AI Chip Production in Early Next Year

2019-12-20

Samsung and Baidu announced this week that the two companies will begin mass production of AI accelerator chips in early 2020. Baidu's Kunlun chip will be manufactured using Samsung's mature 14-nanometer process technology and will utilize Samsung's Interposer-Cube 2.5D packaging structure. The XPU neural processor architecture serves as the foundation for Baidu's Kunlun AI accelerator, which utilizes thousands of small cores to support various applications at the cloud and network edge. This chip delivers 2.6 trillion operations per second (TOPS) at 150 watts of power and provides a memory bandwidth of 512 GB/s through two HBM2 memory packages. According to Baidu, the Kunlun chip is three times faster than traditional GPUs or FPGAs in ERNIE (Enhanced Representation through Knowledge Integration) inference applications. Additionally, it can be used for autonomous driving, speech recognition, image processing, and deep learning. Kunlun is one of the earliest AI accelerators to use I-Cube packaging, manufactured by Samsung Foundry. The 2.5D packaging, which utilizes interposers, is expected to enable Samsung to manufacture other accelerator chips that require high memory bandwidth, leveraging its HBM2 memory products. Furthermore, the company is developing other advanced packaging solutions, including redistribution layers (RDL) and denser HBM packaging. Through their first foundry collaboration, Baidu will provide an advanced AI platform to maximize AI performance, while Samsung will expand its foundry business to high-performance computing (HPC) chips dedicated to cloud and edge computing. Ouyang Jian, architect at Baidu, said, "We are delighted to lead the high-performance computing industry together with Samsung. Baidu Kunlun is a highly challenging project that requires not only high levels of reliability and performance but also the most advanced technologies in the semiconductor industry. Thanks to Samsung's exquisite process technology, we look forward to achieving and surpassing our goal of delivering a superior AI user experience." Ryan Lee, Vice President of Sales and Marketing at Samsung Electronics Foundry, stated, "We are pleased to provide Baidu with our 14-nanometer process technology for their new foundry service. Baidu Kunlun is an important milestone for Samsung's foundry as we expand our business domain from mobile to data center applications through the development and mass production of AI chips. Samsung will provide a comprehensive foundry solution from design support to cutting-edge manufacturing technologies, such as 5LPE, 4LPE, and 2.5D packaging." As applications such as AI and HPC demand higher performance, chip integration technology becomes increasingly important. Samsung's I-Cube TM technology connects logic chips and high-bandwidth memory (HBM)2 through interposers, providing higher density/bandwidth in the smallest form factor by leveraging Samsung's differentiated solutions. These solutions improve power/signal integrity by over 50% compared to previous technologies, thereby maximizing product performance. The I-Cube TM technology is expected to mark a new era in the heterogeneous computing market.

RECENT AI NEWS

RECENT AI TOOLS