Year in Review: Consolidation Rocks Networking Chips
Acquisitions Hit Both Market Leaders and Niche Vendors
By Bob Wheeler
The most surprising aspect of the 2015 consolidation wave is that it didn’t start sooner. In the communications-IC market, too many vendors have been stagnating at less than $200 million in annual revenue. The largest deals were surprising, however, and will have a major market impact.
Intel stands out among acquirers, suddenly grabbing second place in the broadband and FPGA segments. Whether the ultimate outcome is positive or negative, the company will surely change the direction of Altera as it has already done with Lantiq. Avago was the most aggressive dealmaker, as Figure 1 shows, swallowing bigger fish Broadcom. The question in this case is which Broadcom businesses Avago will retain and invest in.
Figure 1. Networking-silicon acquisitions in 2015. The circle sizes represent annual revenue before the combinations (not drawn to scale). Avago was the most aggressive, whereas Qualcomm made only an opportunistic purchase.
The carrier space saw the most deals, many of which were long overdue. In broadband, Intel scooped up Lantiq, Realtek bought Cortina Access, and Qualcomm saved Ikanos. Mellanox pulled in network-processor leader EZchip. Microsemi swallowed Vitesse, followed by PMC-Sierra. Valued at about $2.5 billion, the latter deal is the largest of these transactions, but storage products now represent nearly three-quarters of PMC’s revenue. EZchip sold for about $800 million, whereas Intel paid a modest $383 million for Lantiq. In the bargain basement, Cortina Access and Ikanos each netted about $50 million.
The enterprise networking arena saw fewer deals but plenty of new products. The major technology driver was Wave 2 of 802.11ac, which enables multigigabit Wi-Fi networks. All of the major Wi-Fi chip vendors announced products that handle 4x4 MIMO configurations. The promise of faster access points in turn drove demand for emerging 2.5Gbps Ethernet. Switch chips that handle 2.5GbE ports are now available from leading vendors, but the 2.5GBase-T PHY standard (802.3bz) remains in draft form.
For data centers, most activity revolved around 100G Ethernet and the related 25G and 50G Ethernet rates that remain under development. With its Tomahawk 100GbE switch nearing production, Broadcom filled out the server end with new 25GbE/50GbE adapters and controllers. Intel unveiled its unique FM10000 (Red Rock Canyon) hybrid switch/controller for 25GbE and 100GbE; this device has strong ties to server processors.
Care for Some Gates With Your Server?
Intel turned the FPGA market upside down by acquiring Altera, a deal that should close in 1Q16. The acquirer claims up to one-third of cloud-service-provider (CSP) nodes will use FPGAs by 2020, and it had plans for Xeon processors with embedded FPGAs before the deal. Intel already supplies custom Xeon variants to CSPs, so it should have good visibility into those customers’ plans.
More speculative is Intel’s assertion that the applications it classifies as IoT will represent an incremental $11 billion market for FPGAs by 2020. This projection is based on the assumption that FPGAs become cost competitive with ASICs and ASSPs. We’ve heard this story before, however, and FPGAs have thus far failed to displace such chips in high-volume designs. In any case, Intel’s goals are highly speculative and could defocus Altera from the largest existing FPGA market—communications.
While Altera prepares to join Intel, Xilinx is pressing forward with new products and creating new partnerships. The FPGA leader is shipping high-end 20nm FPGAs and is now sampling SoC FPGAs using 16nm FinFET technology. It has not, however, announced tapeout of a high-end (Virtex) FPGA in 16nm. Meanwhile, Altera has yet to sample Stratix 10 devices built in Intel’s 14nm FinFET process, leaving 28nm chips as its current high-end offering.
With Altera and Intel becoming one, competitors to both companies are lining up new partnerships: IBM dumped Altera in favor of Xilinx for Power8 systems, Qualcomm is working with Xilinx for ARMv8 server acceleration, and Mellanox is using Xilinx for FPGA-accelerated Ethernet NICs. Still, the Intel acquisition places Xilinx at a major disadvantage in data-center workload acceleration.
As the Broadband World Turns
Acquisitions helped consolidate narrowly focused chip vendors in the broadband market. In mature end markets, subscriber and technology churn are driving customer-premises-equipment (CPE) sales while operators pursue cost-effective infrastructure upgrades.
Thanks to DOCSIS 3.1, cable operators enjoy a straightforward path to gigabit-class service. Although subscriber numbers are relatively flat, Broadcom and Intel are benefiting from rising prices for CPE silicon. Broadcom was first to announce a DOCSIS 3.1 CPE SoC when, in January, it introduced its BCM3390. In September, Intel unveiled its Puma 7 SoC, which combines a dual-core Atom processor and DOCSIS 3.1 modem in a 14nm device.
Intel has been quietly gaining share in cable modems— undoubtedly a factor in its decision to acquire Lantiq. The two companies previously worked on Lantiq’s DSL+LTE gateway reference design, which used Intel’s LTE modem chip. The acquisition made Intel the second-place supplier of DSL chips, giving Broadcom a deep-pocketed competitor. Last month, Intel announced an investment in Sckipio, which offers G.fast silicon and had collaborated with Lantiq (see NWR 7/13/15, “Sckipio Unites G.fast and PON in SFP”).
In growing markets, particularly China, operators favor PON for new infrastructure deployments. At Broadband World Forum, Broadcom announced a new “universal” controller chip for PON OLTs. The chip not only handles EPON and GPON but also 10G-EPON and XGPON; the company promises NGPON-2 support in 1H16. The PON-silicon leader, however, faces increasing competition from Asian vendors. As a Huawei subsidiary, HiSilicon poses the greatest challenge. Although it developed its initial products as ASICs for Huawei, HiSilicon now offers PON chips to other OEMs including Alcatel-Lucent. Purely merchant suppliers MediaTek and Realtek are also emerging from Taiwan.
Wildcards include Marvell and Qualcomm, which have limited broadband share but are investing. By acquiring Ikanos, Qualcomm added DSL technologies to its home-gateway offering, which already included IPQ processors, Wi-Fi radios, and HomePlug AV2 chips. Marvell has a growing PON CPE business, but it’s also positioning G.hn as complementary for multiple-dwelling-unit (MDU) access. Neither company, however, has laid out a clear strategy to capture significant market share.
Carrier Networks Aren’t Sexy
Microsemi is consolidating two old-guard communications-IC companies that for years bore the burden of declining legacy products. Vitesse had narrowed its focus to Ethernet products but lacked the resources to develop a competitive portfolio. In networking, PMC continued to invest in OTN chips while merely sustaining its existing network processors and EPON products. This narrower focus yielded share gains for Vitesse and PMC in Gigabit Ethernet and OTN chips, respectively, but the absolute revenue from these products remained modest. Microsemi should improve operating margins through its greater scale, but it’s unlikely to alter product roadmaps dramatically.
Facing customer, product, and technology transitions, EZchip’s future was loaded with risk. On the bright side, the company finally taped out its next-generation NPS-400 in 3Q15. As a part of larger Mellanox, EZchip is better positioned to bridge these transitions as it brings NPS-400 designs to production in 2017. Mellanox should also help EZchip reach hyperscale data-center operators, whereas the latter company’s strength in service-provider networks could open new opportunities for the former.
In wireless infrastructure, 2015 saw few new-chip announcements, as most vendors awaited demand for their existing small-cell SoCs. The major exception was Cavium, which moved up-market to address macro base stations. The new Octeon Fusion-M CPU+DSP chips place Cavium in direct competition with the DSP incumbents, Freescale and Texas Instruments. Although we expect macro-base-station shipments to be relatively flat in the next few years, these systems will still represent more merchant-silicon revenue than small cells.
In place of small cells, the new wave of hyperbole surrounds wireless networks for IoT. Because IoT applications are so varied, they leave room for wireless technologies with different power, cost, and range attributes. For wide-area networks, LoRa has carved out a position for low-power sensor networks (see NWR 10/19/15, “LoRa Alliance Drives Low-Power WAN”). The primary supplier of LoRa chips, Semtech, is forecasting revenue of $100 million within three years. Although that’s still a small market, it’s a high-growth opportunity for Semtech.
Cellular networks provide another WAN option for IoT, but most deployed applications use 2G (EDGE) technology. Mobile-network operators (MNOs) wishing to shut down their 2G networks require new alternatives, but the 3GPP appears confused as to the best approach. Simple LTE derivatives, such as Category 0, are overkill for many IoT tasks (see NWR 8/31/15, “Low-Rate LTE Delivers IoT WANs”). But 3GPP Release 13 could include two lower-bandwidth alternatives, Category M and Narrowband IoT (NB-IoT). Meanwhile, some MNOs plan to deploy LoRa networks in areas where they have already deployed LTE.
Vendors are also talking up IoT applications for wired networks, but much of this positioning really represents the transition of industrial networks to Ethernet. Recall that the IEEE 1588 packet-based timing protocol originally targeted industrial applications before extending to carrier networks. So, IoT is largely the new name vendors are using for various embedded applications, which do present growth opportunities.
Although embedded unit volumes are generally low, specialized chips can carry much higher prices than commoditized Ethernet components. Both Microsemi and Microchip—which acquired Vitesse and Micrel, respectively—are targeting these high-margin designs. Microchip is a leader in 10/100 Ethernet products for industrial designs; Microsemi is entering the market with GbE chips.
Whereas IoT drove acquisitions, data centers drove most of the revenue growth for Ethernet chips. In data-center switching, Broadcom has become so dominant that most vendors are avoiding direct competition with it. The major exception is newcomer Cavium, which sampled its first Xpliant switch chip in 1Q15. The CNX88091 is a 32x100GbE switch that competes directly with Broadcom’s Tomahawk. Platforms based on Tomahawk are already available from OEMs, however, so Broadcom is first to market.
Mellanox introduced its Spectrum 32x100GbE switch chip, but it continues to focus on system-level switch products. Meanwhile, Centec introduced its GoldenGate 10GbE switch, which competes with Broadcom’s Trident II+.
Intel’s Red Rock Canyon is an intriguing chip that spans multiple applications. It lacks the port density, however, to compete with Tomahawk for top-of-rack switch designs. On the other hand, the company should sell the FM10000 with relative ease into appliance designs, where it can pair the controller-and-switch chip with Xeon E5 processors. For enterprise servers, Intel updated its 10GBase-T controller design with the new 28nm Sageville X550, which adds support for the pre-standard NBase-T version of 2.5Gbps and 5Gbps Ethernet.
Ethernet Enabling Wi-Fi to Kill Ethernet
Cisco and HP (now HPE) both launched switches that support NBase-T during 2015, and these designs should be forward compatible with the 802.3bz 2.5/5GBase-T standard. The standards effort is moving quickly, with a sponsor ballot expected in May 2016. Vitesse (now Microsemi) was first to market with a 24x2.5GbE switch chip in 2014, whereas Broadcom and Marvell introduced switch chips optimized for this speed during 2015. Aquantia remains the only vendor shipping an NBase-T PHY, while Marvell has demonstrated an unannounced chip. Broadcom plans to support 802.3bz, but its initial PHY was designed for MGBase-T, which uses different modulation than the draft standard.
The impetus for 2.5Gbps Ethernet came from 802.11ac Wave 2 designs, which deliver greater than 1Gbps throughputs thanks to 4x4 MIMO and multiuser MIMO (MU-MIMO). Using beam forming with multiple antennas, MU-MIMO enables an access point to simultaneously send independent streams to multiple clients, as Figure 2 shows. (For a refresher on MU-MIMO, see NWR 4/28/14, “Vendors Tussle Over Multiuser Wi-Fi.”)
Figure 2. Complex beam forming for MU-MIMO. The access point strengthens the signal for each client in its own direction while creating a local null to reduce crosstalk from other clients.
Although Wave 2 features are most important for enterprise Wi-Fi, Broadcom, Marvell, and Qualcomm Atheros all introduced new chipsets at the Consumer Electronics Show (CES). Smaller Quantenna had been shipping Wave 2 chips for more than a year and was waiting for the market to catch up. The Wi-Fi Alliance still has yet to introduce a Wave 2 certification program, however, more than two years after beginning Wave 1 testing (see NWR 8/12/13, “Wi-Fi Alliance Blesses 802.11ac”).
Lacking independent testing, vendors can make dubious claims about throughput, sometimes even using proprietary PHY rates (e.g., QAM-1,024). On paper, however, Broadcom, Qualcomm, and Quantenna all provide fully featured Wave 2 chipsets that handle 4x4 MIMO with four streams (4x4:4), MU-MIMO, and standard QAM-256, yielding a 1.7Gbps maximum PHY rate in an 80MHz channel. Marvell’s design supports four antennas but only three spatial streams (4x4:3), reducing its maximum rate to 1.3Gbps (using an 80MHz channel).
To stay ahead of the pack, Quantenna demonstrated at CES a “10Gbps” Wi-Fi platform followed by a formal announcement of its “Wave 3” chipset in September (see NWR 9/28/15, “Quantenna Debuts 8x8 MU-MIMO”). The maximum 9.6Gbps aggregate rate, however, requires two radios and proprietary QAM-1,024 modulation. Still, the company is first to market with an 8x8:8 MIMO design, which supports four simultaneous MU-MIMO clients. For enterprise networks, such a design provides a more spectrum-efficient alternative to multiple radios or 160MHz channels. Quantenna somehow remained independent through the 2015 merger-and-acquisition (M&A) frenzy, likely hoping for a higher valuation in an IPO.
Catching Our Collective Breath
Following the M&A activity of 2015, many companies will focus on integration and execution in 2016. This operational focus may cause a lull in product announcements across multiple markets. Avago must digest Broadcom, reviewing product lines and roadmaps. The new Broadcom will likely introduce some products, however, that were already sampling or taped out in 2015. They should include the company’s first 16nm networking chips, such as network search engines (KBPs).
Intel will first deliver new NICs based on Red Rock Canyon as well as Altera’s previously announced Stratix 10 line. It also plans to deliver FPGA-integrated Xeon processors to lead customers in 1Q16, marking the initial step in its longer-term FPGA plan. Having refined its home-gateway strategy, the company may find it needs to fill holes by acquiring small vendors like Quantenna for Wi-Fi or Sckipio for G.fast.
Although Marvell faced operational challenges in 2015, it also quietly hired an Ericsson veteran as the new head of its networking business. Before joining Marvell in August, Maya Strelar-Migotti spent the prior six years running Ericsson’s development center in Silicon Valley. This recruitment is unusual, in that president and cofounder Weili Dai has traditionally directed the networking business. In 2016, Strelar-Migotti could put her stamp on the organization by focusing investment on a smaller number of market segments.
Fewer acquisition opportunities remain, but 2016 could see some deals among more-analog-oriented names. Potential targets include Inphi, Maxim, and Semtech. Another name that needs greater scale is AppliedMicro, but that company appears uninterested in separating its OTN line from its dissimilar ARM processor line. Having networking processors plus an emerging Ethernet switch business, Cavium is an attractive target. A price tag in the $5 billion range, however, limits the list of potential buyers.
Consolidation can be healthy when it culls the herd of the old and diseased. In communications, however, some markets require patience and long-term investment. Our hope is that the 2015 acquirers will take the time to understand their new businesses before revising roadmaps or divesting products. Excessive focus on the bottom line could undermine the development of new products and technologies, stifling future growth.
The Year in M&A
Ethernet specialists grew with Microchip’s deal for Micrel (see NWR 6/8/15).
Mellanox picked up NPU leader EZchip (see NWR 10/19/15).
Networking-Chip Events of 2015
It was the year of 802.11ac Wave 2, beginning with announcements at CES from Broadcom, Marvell, and Qualcomm Atheros (see NWR 2/2/15).
MWC saw Cavium attack macro base stations with Octeon Fusion-M (see NWR 3/9/15).
Broadcom introduced 28nm generations in its StrataDNX (see NWR 4/6/15) and Trident switch families (see NWR 5/11/15). It also launched all-new 25G/50G Ethernet adapters and controllers targeting cloud servers (see NWR 8/3/15).
Mellanox introduced a 100G Ethernet switch chip as well as NICs for 25G/50G Ethernet (see NWR 7/6/15).