Most-powerful DSP for 5G and 5G-Advanced

• Infrastructure
• Automotive 

The Ceva-XC23, based on the advanced Ceva-XC20 architecture, features dual execution threads with independent memory subsystems and caches. It includes a shared Vector Computation Unit (VCU) with four engines, dynamically allocated to maximize utilization and reduce contention. The VCU consist of two arithmetic engines, with one non-linear engine and one move-and-scale engine that supports 128 16×16 MACs. The architecture ensures efficient data handling with wide memory connections and dedicated DMA engines.

The Ceva-XC23 DSP offers exceptional efficiency in terms of power, performance, and area, making it ideal for 4G, 5G, and 5G-Advanced NR implementations. With AI support and programmability, these DSPs not only enhance modem performance but also enable advanced AI and ML workloads.

It delivers 2.4 times the performance of Ceva’s previous-generation Ceva-XC architecture, while maintaining a compact DSP core design. This allows clusters of Ceva-XC23 cores to achieve extraordinary performance levels while saving precious SoC real estate.

The Ceva-XC23 is software compatible with the Ceva-XC4500 and provides scalable performance across various 5G profiles, includinguRLLC, and eMBB. With a high control code performance of 5.14 CoreMark/MHz, it ensures top-notch performance. The core features a unified programming model across all Ceva-XC20 variants and supports low power modes for system power optimization. The design allows for easy integration into SoCs with standard AMBA4 buses. Additionally, the complete software tool chain includes an optimizing LLVM C compiler for high-level programming, and best-in-class debug capabilities for effective system development.

• Industry’s most powerful vector DSP core, with 2.4 times the performance of Ceva’s previous Ceva-XC architecture
• True dual-threaded hardware, including dual processing elements and dual instruction and data memory subsystems for contention-free multithread execution
• SIMD vector size of 512bits
• 128 MACs (16bits) with very high utilization on multithreaded workloads
• AI workloads support with INT8/16/32 data types
• Supports half-precision, single-precision, and double-precision floating point
• Advanced pipeline implementation assures high clock frequencies
• Low power operating modes for system power optimizations
• Nine-issue VLIW instruction set architecture and optimizing LLVM C compiler deliver high throughput without extensive assembly coding
• Up to six AXI4 bus connections in AMBA matrix interconnect architecture assure enormous bandwidth but simple, standards-compliant interfacing.
• Dedicated 5G / 5G-Advanced ISA for 5G NR acceleration
• ETM compliance, JTAG, RTT, Profiler, APB slave