All Tier One networking and communications OEMs use NPUs in a range of systems, including edge/services routers, Carrier Ethernet switch/ routers (CESRs), optical-transport platforms, and wireless and broadband infrastructure. Many of these OEMs continue to invest in internal NPUs, which limit the adoption of merchant products. But most of those with internally developed NPUs still use merchant NPUs in some designs.
Because semiconductors are subject to occasional inventory corrections, the merchant-NPU market displayed a somewhat volatile growth rate in the last decade. This trend continued through the recession of 2009, followed by an upswing in 2010. What followed, however, was an unprecedented four-year decline in 2011–2014. This decline was due to several factors including an initial inventory correction followed by weak macroeconomic conditions and falling revenue from the access market. Looking forward, we expect the NPU market will return to significant growth in 2017 as the access segment becomes a smaller contributor and high-end NPUs continue to grow.
Despite being smaller than most competitors, EZchip is the leading supplier of high-end merchant NPUs. The company is shipping multiple product generations including the 200Gbps NP-5, which combines a customer-programmable packet processor, traffic manager, and Ethernet MACs. It has worked with Marvell to supply special versions of its NPUs to Cisco, which primarily uses them in the ASR 9000 router. EZchip is developing a new line of NPUs, called NPS, that handle Layer 4–7 features and promise easier programming than traditional NPUs.
In the access market, PMC-Sierra is the leading vendor. The company entered the NPU market in late 2010 by acquiring Wintegra, which had become a leading vendor of access NPUs because of its flexible architecture and complete data-plane software. Since the acquisition, PMC has increasingly focused on wireless-backhaul designs, where support for legacy protocols continues to differentiate the WinPath NPUs from Ethernet-only alternatives. In 2014, PMC reached production with WinPath4, which scales performance to 40Gbps.
With its BCM88038, Broadcom shipped the industry’s first full-duplex 100Gbps NPU. The company combines this customer-programmable NPU with its 200Gbps configurable traffic manager, switch fabric, and search coprocessors to offer a packet-processing solution that is more complete than any competitor’s. Customers can combine these components in 400Gbps line cards or use them independently in different designs. Broadcom also provides ASIC services to OEMs wishing to develop in-house NPUs, such as Alcatel-Lucent with its FP3. For access and aggregation applications, it instead offers configurable Carrier Ethernet switch chips in its popular StrataXGS line.
In addition to its manufacturing role with EZchip, Marvell supplies high-end NPUs of its own design, which come from the company’s acquisition of NPU startup Xelerated. In late 2013, it sampled the HX41xx, which offers throughputs up to 400Gbps. The derivative AX line delivers programmable packet processing for access- and aggregation-switch designs. Marvell has also added traffic-management technology from its NPUs to Prestera configurable Carrier Ethernet switch chips, which address lower-cost access designs.
Netronome is now the only privately held company developing NPU silicon. Rather than competing for carrier designs, it principally targets security appliances and other data-center applications. The company started to generate NPU revenue in 2011 with initial shipments of its 40Gbps NFP-3240. In 2014, Netronome sampled the NFP-6xxx, its next-generation NPU that scales throughput to 200Gbps. Designed for stateful flow processing, the NFP devices are unique in their ability to perform advanced services such as IPSec, SSL, firewall/NAT, load balancing, and deep packet inspection (DPI).
Also referred to as TCAMs or network search engines (NSEs), search coprocessors offload NPUs by performing lookup functions. Cisco and other leading OEMs combine search coprocessors with both merchant and internal NPUs, resulting in a merchant search-coprocessor market about two-thirds the size of the merchant-NPU market. As with NPUs, some OEMs also possess proprietary search designs.
Through its NetLogic acquisition, Broadcom became the dominant supplier of search coprocessors, which it calls knowledge-based processors (KBPs). Its traditional competitor is Renesas, which primarily supplies TCAMs to Cisco. Two new entrants, Cavium and Marvell, hope to compete with Broadcom using algorithmic implementations that promise to reduce power and cost. Depending on their application, OEMs now have at least two vendors to choose from.