Dual-Threaded DSP Targets AI and Wireless Apps
The Ceva-XC23 DSP core from Ceva is built to handle artificial-intelligence (AI) applications as well as RF chores for 45, 5G, and 5G-Advanced NR applicions. It delivers 5.14 CoreMark/MHz with easy integration via AMBA4 buses.
The dual processing elements feature L1 cache in addition to processing and data tightly coupled memory (PTCM/DTCM). The individual scalar execution units are augmented with a shared, 512-bit, SIMD vector computation unit (VCU). This includes 128 16-bit MACs. Nine-issue VLIW architecture is supported by an optimizing LLVM C compiler.
The processing system is designed to take on AI applications with support for INT8/16/32 data types plus half-, single-, and double-precision floating point. There is dedicated 5G/5G-Advanced instructions to improve 5G NR acceleration.