Ceva has added DSPs to the Ceva-XC20 line. The XC23 updates the infrastructure-focused XC22, and three XC21 models target wireless IoT and other low-speed and midrange devices. They target chips for the newest 5G and forthcoming 6G standards (including satellite connectivity) and also improve INT8 support for AI functions. The DSPs will be available to license by the end of next month.
Ceva-XC23 DSP Enhances High-Performance XC22 Architecture
The original XC22 was a conjoined-twin design in which two DSPs shared a vector unit (VCU), dynamically allocating processing resources between the DSPs’ threads. Because the VCU is physically large but typically utilized during a fraction of computing cycles, sharing it conserves die area with little performance degradation.
The XC23 enhances the XC22 while retaining the same high-level architecture. After analyzing XC22 customers’ code, Ceva added instructions to speed up time-critical operations. The new DSP also adds accumulator bits to mitigate overflow in long operation sequences.
Updates for INT8 processing include byte-aligned addressing and native 8-bit operations for improved throughput and memory utilization. The INT8 support helps new designs employ AI-based heuristics for applicable wireless functions. For the past few years, wireless-chip companies, such as MediaTek, have touted the use of AI to improve smartphones’ and base stations’ radio performance.
Ceva-XC21 DSP Targets RedCap Devices and NTN Satellite Terminals
The XC23 can be used in high-end user equipment such as smartphones, but other devices connect to the cellular network. These include IoT devices, vehicles, and satellite communicators. Most require lower-cost, lower-power modems and don’t need the fastest data rates. Thus, the 5G RedCap and the new eRedCap standards for low-speed connectivity are sufficient. For these applications, Ceva offers the XC21 series.
The XC21 comes in three versions to address different performance levels. The XC212 is similar to the XC23 but halves the VCU’s width to 256 bits. Like the XC23 and the XC22, its VCU has two vector arithmetic (VA) blocks and is a dual-thread, conjoined-twin core. The SC211 is a unitary, single-thread DSP. Performance on vector-heavy code should be good because it doesn’t share the vector unit between threads, but this decreases area efficiency. Designs requiring less vector throughput can employ the SC210, which integrates only a single VA to reduce cost (die area).
Ceva states the XC211 delivers the same performance as the older Ceva-XC4500 with a 37% area reduction. Compared with that DSP, the XC212 delivers 80% more throughput but is 12% smaller, while the XC210 isn’t as powerful but occupies half the area.
Bottom Line
Whether developing user-equipment or infrastructure modems, companies either employ a proprietary DSP or license one from Ceva. For customers requiring more than a DSP, the company also offers its PentaG platform, which adds accelerators and software to complete a modem design. We expect future PentaG offerings to integrate the new XC23 and SC21 cores.
Cellular standards continually evolve. Recent 5G developments include satellite communications (NTN) and RedCap/eRedCap. On the horizon lies 6G. It may be five years before the first 6G networks go live, but backward scheduling to meet this target indicates DSPs must be available now.
To address evolving standards and to better handle existing 5G protocols, Ceva has developed new DSPs. Source compatible with the XC22 and delivering more performance, the XC23 is a natural upgrade for the company’s infrastructure and high-end user-equipment customers. For less costly devices, the XC21 cores enable connectivity using the newest low-speed, low-power 5G standards and NTN. Other licensing options exist in theory, but only Ceva focuses on cellular technology and is, therefore, the only viable choice for companies without proprietary DSPs.