Intel is to buy Barefoot Networks, in a bid to fill a gaping hole in its portfolio, for strong Ethernet switch-chips.
The company was clear, in conversations at Mobile World Congress this year, that it needed to assemble an end-to-end platform if it was to dominate cloud infrastructure and, in particular, the increasingly cloudified 5G network. It has been putting together many elements, through internal development or acquisition, for this 5G infrastructure play, including the purchase of Altera for its FPGAs (field programmable gate arrays) and of eASIC. Both of these provide technology to surround the core x86 processors and accelerate specialized and highly demanding cloud and 5G workloads.
But Intel identified one of its main weaknesses, in this end-to-end objective, as being the switch-chips which Broadcom dominates. Although the larger firm has acquired switch silicon makers in the past – notably Fulcrum in 2011 – it has fallen behind Broadcom in market share and technology, and not kept pace with the rise of open systems and white box architectures in the switching space.
Barefoot, a Silicon Valley start-up which has been central to AT&T’s market-leading white box switching initiative, will fill many of the gaps, bringing an open, modern approach with no legacy. It makes Ethernet switch silicon and software for data centers and telecom networks, including the Tofino range of programmable ASICs, which are liked by cloud providers for their flexibility and use of the emerging open source standard language, P4.
The company was founded in 2013 by chief scientist and Stanford University professor Nick McKeown, CTO Pat Bosshart and chief development officer Dan Lenoski, and now has about 200 staff. As of last October, it had raised $150m from investors including Sequoia Capital, Google Cloud, Tencent, Alibaba, Dell and HPE.
The purchase price was not disclosed but was likely to have been lower than the $5.5bn Intel reportedly bid recently for another switch-chip maker, Mellanox. It was outbid for that firm by Nvidia, which paid $6.9bn. Intel expects the transaction to close in the third quarter of 2019, when Barefoot CEO Craig Barratt and the rest of his team will join Intel.
Initially, this deal seems like a consolation prize for losing Mellanox to Nvidia, Intel’s greatest threat in the hyperscale world. But longer term, it may have got a better deal with Barefoot anyway, with fewer ageing products to consider; some strategic partnerships including AT&T; and a leading role in the shift of telco networks towards software-defined networking (SDN), high performance standardized hardware and easy programmability – all trends which Intel has identified as favoring the spread of the x86 architecture into previously proprietary areas such as telecoms backbones.
Of course, Intel will have to execute well on an acquisition which raises some red flags, based on its patchy track record in successful integration of companies. Barefoot is culturally very different and a lot further down the open source road than Intel. In the past, Intel has made many acquisitions which looked strategically wise and important, but then failed to integrate them well enough into its working structures or its technology platforms. Full integration of the hugely significant Altera products has been slow, according to insiders; other interesting acquisitions were wasted completely, including the StrongARM processors bought from Digital Equipment (eventually sold on to Marvell), and probably the Infineon modem business, which now looks set to be wound down or sold (see separate item).
The StrongARM and Infineon examples relate to the multi-episode saga of Intel’s failure in mobile devices. In mobile infrastructure, however, it cannot afford to fail, since 5G networks, and the switches and routers that underpin them, will increasingly run on the cloud platforms that are essential to Intel’s future growth. So this is not just about data centers, though those are the most logical first target – it is closely related to the bid for an end-to-end 5G platform.
Navin Shenoy, general manager of Intel’s Data Center Group, wrote in a blog post that Barefoot “will support our focus on end-to-end cloud networking and infrastructure leadership … Barefoot Networks will add deep expertise in cloud network architectures, P4-programmable high speed data paths, switch silicon development, P4 compilers, driver software, network telemetry and computational networking.”
Zeus Kerravala, principal analyst at ZK Research, told SDxCentral: “For all of Intel’s success, they have never been that strong in hyperscale data centers, particularly in the area of networking. Their arch-nemesis, Nvidia, bought Mellanox to drive greater integration between compute and networking, and Intel can do the same. Barefoot’s Tofino chip is well regarded as best in class for data center interconnect. As the world becomes more distributed and edge driven, having the capability to connect the cloud to the edge gives Intel a strong competitive position. This also aligns well with the edge push that Intel has had over the past year. It could address compute and now it can bring networking into the fold.”
Central to the appeal of Barefoot is P4, the language which the start-up helped to develop and which is described as “taking SDN to the next level” by enabling the programming of forwarding planes. This is important in telco networks, as highlighted by AT&T’s ground-breaking work in using Tofino (as well as Broadcom chips) to power white boxes and improve the efficiency and flexibility of its data transport network in readiness for 5G.
The work done so far by AT&T will provide the kind of environment in which Intel needs to get embedded in order to make an impact on 5G. In late 2017, AT&T’s CTO, Andre Fuetsch, demonstrated the use of P4 on merchant silicon for a white box switch, saying in a conference keynote: “This is more than just about lowering cost and achieving higher performance. Frankly that’s table stakes. This is really about removing barriers, removing layers, removing all that internal proprietary API stack that we’ve lived with these legacy IT systems, now we can bypass all of that and go straight to ONAP” to achieve fine-grained per-packet visibility.
AT&T worked with Barefoot and SnapRoute on its switch. Barefoot provided its Tofino chip and surrounding software, while SnapRoute (in which AT&T is an investor) developed the network operating system, Flex Switch. That has now evolved into AT&T’s dNOS network OS, which it has placed into open source via the Linux Foundation-hosted DANOS project.
Fuetsch said the firm “took this very novel approach of building a hardware abstraction layer and running open source networking modules like BGP and OSPF on top”. This layer can then operate independently of the silicon, and AT&T said it has other white boxes, based on different chips, in the works, with the same network OS.
All this will, operators hope, add up to a transformed cost base and supply chain for 5G. As Fuetsch said: “Here’s the big message for the OEMs. It is really a call for them to open up their architectures, open up their software and their own hardware so they can participate. They are going to have to make a choice here – do you want to be at the table or on the plate?”
The next step will be to push the open hardware and virtualized network efforts into the wired and wireless access space. “That’s really where the big money and big opportunity is,” Fuetsch said at the start of 2019, adding that AT&T is trialling the ONF’s SDN-Enabled Broadband Access (SEBA) reference design in Georgia and Texas. This is a lightweight version of another open source project, R-CORD (Residential Central Office Re-architected as a Datacenter). SEBA supports access technologies, such as PON, G.Fast and DOCSIS, at the edge of service providers’ networks. Deutsche Telekom and Turk Telecom are also planning SEBA deployments.
The US operator has also tested white box switches and CPE, the former powered by dNOS and the P4, which has replaced Openflow as the Open Networking Foundation’s default protocol for switching. Chip vendors including Broadcom, Barefoot, Cavium and Mellanox now make merchant switch-chips for white box platforms, and AT&T has tested systems with the first two of those suppliers. Quanta and Delta are among the first companies to make white box switches themselves.
Broadcom unleashes its own open language to challenge P4:
Broadcom will be in Intel’s sights, but remains a powerful company to beat in switches. It recently announced that it was sampling its first 7-nanometer network switch-chip, the Trident 4 – but it is pushing its own network programming language rather than, as many had expected, succumbing to industry pressure for a single language by adopting P4.
The Trident 4 family spans switches with 2-128Tbps aggregate bandwidth, and the 21bn-transistor chip can mannage up to 5bn packets per second in a single chassis.
“No one has really made a significant dent so far” in Broadcom’s share, estimated at 80% of the $3bn merchant Ethernet switch-chip market in 2018 by chip analysts the Linley Group (this excludes company-specific ASICs, notably Cisco’s). Linley’s Bob Wheeler told EETimes: “Barefoot has gotten traction with Arista and Cisco designs and there’s a lot of talk about Google using them, but it’s not clear how broadly the systems are being deployed.”
Trident 4 can be programmed with existing APIs (application programming interfaces) or the new open language, called Network Programming Language (NPL), which is a clear bid to squeeze Barefoot. One key difference, according to Wheeler, is that Barefoot has focused on making its chips customer-programmable, while the Trident 4 is more focused on letting OEMs program them.
Broadcom will provide tools to support NPL on Trident 4 and its Jericho 2 chips, which are now sampling. Importantly, other companies can port the code to their architectures for free, as Broadcom aims to establish a rival platform to P4 across the industry. The firm isstill determining which open source licence to use. Existing Trident OEMs include Arista, Dell, Huawei and Juniper, and their migration from existing APIs to NPL will be crucial to Broadcom’s endeavor. China’s web giant, Tencent, has also expressed interest in NPL.