Social Network giant Facebook has open-sourced a next-gen hardware platform for Artificial Intelligence model training, named Zion, along with custom Application-Specific Integrated Circuits (ASICs) optimized for AI inference, Kings Canyon, and video transcoding, Mount Shasta. According to the company, the platforms, which donating to the Open Compute Project, will considerably stimulate AI training and inference. The Open Compute Project is an organization which shares designs of data center products among its members.
In a statement, Facebook engineers Kevin Lee, Vijay Rao, and William Christie Arnold stated that AI is used across a range of services to help people in their daily interactions and provide them with unique, personalized experiences. In addition, AI workloads are used throughout Facebook’s infrastructure to make their services more relevant and improve the experience of people using the company’s services. Zion, which is designed to handle a gamut of neural networks architectures, such as CNNs, LSTMs, and SparseNNs, comprises three parts- a server with eight NUMA CPU sockets, an eight-accelerator chipset, and Facebook’s vendor-agnostic OCP accelerator module (OAM). It boasts high memory capacity and bandwidth, owing to two high-speed fabrics, a coherent fabric which links all CPUs, and a fabric which connects all accelerators, and a flexible architecture that can measure to various servers within a single rack using a top-of-rack (TOR) network switch.
On the other side, Kings Canyon, designed for inferencing tasks, and classified into four components- Kings Canyon inference M.2 modules, a Twin Lakes single-socket server, a Glacier Point v2 carrier card, and Facebook’s Yosemite v2 chassis. Moreover, Kings Canyon’s every server combines M.2 Kings Canyon accelerators and a Glacier Point v2 carrier card that link to a Twin Lakes server; two of these are installed into a Yosemite v2 sled, which has more PCIe lanes than the first-gen Yosemite, connected to a TOR switch through a NIC. Facebook pointed out that it’s collaborating with Esperanto, Habana, Intel, Marvell, and Qualcomm to develop ASIC chips to support both INT8 and high-precision FP16 workloads.