Up first is an updated second-generation Xeon Scalable processor, now at a top speed of 3.9GHz and bolstered by additional AI capabilities to aid with inference applications. The new chip promises up to 36 percent more performance than the first-generation version, with up to 42 percent more performance per dollar, though early second-generation chips were introduced in April 2019.
Intel says the Xeon Scalable is the “only CPU with AI built-in” while this is not completely true, Intel said that it means “the only CPU on the market that features integrated deep learning acceleration”. Xeon Scalable’s Deep Learning Boost feature set promises up to six times more AI performance than AMD’s Rome processors. Intel is not telling the world the number of TOPS available for AI processing but says Xeon Scalable will support the cloud AI needs of Alibaba, AWS, Baidu, Microsoft, and Tencent, as well as other major companies.
Network-optimized “N-SKUs” of the new Xeon Scalable will also be available, offering up to 58 percent more performance for network function virtualisation workloads compared with the first chip. Customers such as China Mobile, SK Telecom, Sprint, and T-Mobile Poland are all using Xeon Scalable in their 5G networks. The boosted Xeon Scalable chips are officially available starting today.
Meanwhile, Chipzilla is introducing the Atom P5900 for base stations and designed for radio access network (RAN) needs. It’s a 10-nanometer chip with hardware-based network acceleration features, including integrated packet processing, ultra-low latency, and a switch for inline cryptographic acceleration. It promises up to 1.8 times the integer throughput versus an Atom C3000, plus 5.6 times the packet security throughput and 3.7 times the packet balancing throughput compared with software-based alternatives.
The Atom P5900 is designed for use in 5G base stations, and Intel already has customers for the chip, including Ericsson, Nokia, and ZTE to use it in their RANs.Intel is aiming to be the market leader in 5G base station silicon by 2021, ahead of its 2022 goal, with a 40 per cent share of the growing business.
Intel also announced the eASIC, codenamed Diamond Mesa, which it’s calling its first 5G structured ASIC — an option for customers seeking custom chip solutions containing low-power, high-performance Intel processors plus customer-specific IP in silicon. Using a footprint compatible with prior Intel FPGA solutions, Diamond Mesa promises twice the performance or half the power consumption versus prior-generation Intel ASICs. The company expects eASIC solutions to be used in 5G wireless data centres.
Chipzilla also announced the Ethernet 700 series, 5G-optimised network adapter cards with hardware-enhanced precision time protocol (PTP) for ultra-low latency and similarly precise timing requirements — it’s targeting 100-nanosecond phase accuracy for 5G network service synchronization. The cards will be used in computers demanding split-second responsiveness for controlling industrial equipment, financial trading, emergency response, RAN connectivity, or video streaming. The Ethernet 700 series is sampling to customers now and will go into production in the second quarter of 2020.