Chiplet Set For Artificial Intelligence

The design reuse strategy has significantly shortened the time required to create complex System on Chips (SoCs). However, when introducing new intellectual properties (IPs), the monolithic SoC methodology requires a re-run of system-level validation steps, incurring significant costs. Partitioning the design into chiplets over an interposer would mitigate these issues by consigning the IP updates to the individual chiplet. This paper presents a chipletized design used for Artificial Intelligence (AI). This design details a scalable AI chiplet set, along with Central Processing Units (CPUs). The AI chiplet set includes an Long Short Term Memory (LSTM) Application Specific Instruction Set Processor (ASIP) for accelerating inference and training and an Sparse Convolution Neural Network (SCNN) ASIP for accelerating inference through a zero-skipping technique. The CPUs control AI accelerators and handle general tasks. The accelerators and CPUs have an AXI crossbar Network on Chip (NoC) for memory and one for controlling the accelerators. This project has two phases: phase one, IP validation with an emulated interposer (No interposer, connect chiplets through back end of line (BEOL) metal layers), and phase two, connecting validated IP through an interposer. This paper focuses on phase one, which uses the United Semiconductor Japan Co. (USJC) 55 nm LP process to fabricate the design. The chiplets' clock frequencies range from 200 - 400 MHz.