- Analysts are not employed by SDxCentral.
- Views and opinions expressed in analyst content belong solely to the author and do not reflect the views of SDxCentral, LLC.
- SDxCentral does not fact check analyst content. If you believe there is a factual error in analyst content, please notify firstname.lastname@example.org. Should we find factual irregularities, that article will be unpublished from the SDxCentral website.
Effective April 18th, 2019, the SDxCentral analyst blog syndication program has been terminated.
SDxCentral Statement about AvidThink, LLC
- Roy Chua, the founder of AvidThink, was a co-founder of SDxCentral. As of September 30, 2018, Roy is no longer affiliated with SDxCentral.
- The views expressed by AvidThink and Roy Chua are independent of SDxCentral and do not represent the views or journalistic principles of SDxCentral.
- As of April 18th, 2019, SDxCentral is no longer publishing AvidThink analyst blogs on the SDxCentral website.
By the time you read this we’ll likely be at Open Networking Summit 2019, where I’m sure there’ll be plenty of news for the press to cover. The advantage to AvidThink being analysts is that we can cover topics retrospectively without looking like we’re slacking by not giving you the up-to-the-moment coverage. Plus, you’re probably exhausted and SDN-out and … ahem … ready for ONAP or two.
Moving on … “IBM and the Seven Dwarfs” was a popular term in the business press back in the 1960s when IBM dominated computing. Only a few of the “Seven Dwarfs” – Burroughs, Sperry Rand/UNIVAC, Control Data, Honeywell, General Electric, RCA, and NCR – survive to this day and in slightly different lines of business. Given that, it’s remarkable how IBM has been able to re-invent itself to remain relevant.
The Seven Dwarfs of Merchant Silicon
In the data center switching chipset or merchant silicon market today, it feels a little like Broadcom (Avago) is the new Snow White. And I’m not sure I can name seven competitors, but here we go: Barefoot, Centec, Innovium, Marvell, nVidia/Mellanox, and Nephos (from Mediatek). We could have had seven with Cavium/Xpliant, but they are now part of Marvell. And yes, networking vendors like Cisco have their own switching chipset, but they typically don’t provide them to third parties for use.
Back when SDN started its ascendency around 2012, the other merchant silicon giant, Marvell, took a little hiatus. Prior to that, when building data center switches, you would always look at Broadcom and Marvell silicon as a starting point. For the last six years, it feels like Broadcom dominated the merchant silicon market. After a last hurrah with Marvell’s chips in Juniper’s data center switches, it seemed like Marvell – a previous Snow White contender – took a long nap in data center switching and woke up a Dwarf. To be fair to Marvell, it had its hands full with other issues and did continue to innovate in other product lines. But it struggled for a while, including dealing with a $1.5 billion judgement against them for infringing on Carnegie Mellon University’s patents, which ultimately resulted in their paying $750 million in 2016. Eventually their business woes led to the ouster of their founder/CEO, and president – a husband-and-wife team – from active management in 2016.
Around the time Marvell faded from the data center switching scene, Intel showed up with its then-recent Fulcrum acquisition. The Intel FM 6000 with its sub-400 nanosecond latency helped Arista – which built a switch on the chipset – drive hard into the financial markets, which valued low-latency for high-frequency trading. Soon after the FM 6000 was launched, Cisco upped the ante with their own ASIC Monticello-powered switches, the Nexus 3548, which boasted typical latencies of 250 nanoseconds, which were driven down to 190-nanosecond latency in what they called “Warp” mode.
And from then on it appeared that new players on the scene would continue to drive innovation despite the loss of the Broadcom-Marvell duopolistic dynamics. Intel subsequently dropped out of the merchant silicon switch market, but Barefoot, Centec, Innovium, nVidia/Mellanox, and Nephos jumped in. And some of them –Barefoot, Innovium, and Nephos – managed to get to the 12.8 Tb/s switching capacity bar. Impressive, considering the challenges and difficulty in fundraising for startup chip manufacturers and the level of investment required to bring a chip to market.
At the same time, these vendors continue to innovate in terms of the level of programmability, with Barefoot pushing hard to get programmability adopted by data center operators with its bet on the P4 language.
Marvell’s Back in the Game?
At the OCP Summit 2019, it seemed Marvell finally got back into the fray with its Prestera CX 8500 family, showing off a working 12.8 Tb/s switching fabric at the show complete with blinky lights and traffic generators spewing 64-byte packets in a full-mesh configuration. Marvell indicated that they will be providing a whole series of lower-power, modular switching solutions that can be managed and programmed the same way for a range of core to edge data center use cases. And they talked about “just enough” fabric programmability to meet most data center needs, though it’s clear the Xpliant-level of flexibility will likely not show up again soon. They even hinted at some design wins to be announced soon. And in speaking to friends at leading switch manufacturers, these manufacturers are again seriously considering Marvell, along with Broadcom and other players, for upcoming high-end switch platforms.
Now the game for Marvell, Broadcom, and the other players will be a multifaceted one, balancing the requirements of latency, power, telemetry, switching capacity, and programmability in a play that spans not just the core data centers, but one that likely involves the new edge data centers as well.
Part of chipset adoption in the data center switching game will depend on ease of orchestration, configurability, and programmability. To that end, OCP’s two projects, Switch Abstraction Interface (SAI) and Software for Open Networking in the Cloud (SONiC), which were both originally seeded and proposed by Microsoft, saw strong uptake and support at this year’s OCP Summit. It’s quite remarkable that Juniper has jumped on the SONiC bandwagon as well. With Arista, Big Switch, Cisco, Dell, Juniper, nVidia/Mellanox, and many more players all onboard in some form, there’s likely to be innovation on that front. There’s always concern that these players will only pay lip service and not contribute their best and brightest to the effort, but I remain hopeful that SAI will help drive change in the switching market and that SONiC and projects like Free Range Routing (FRRouting) will make progress. I know there are some who are critical of Quagga’s scalability and questions about maturity of the modules in SONiC, but we’ve got to start somewhere and there’s plenty of hard work involved – Cisco, Juniper, Huawei, and Nokia didn’t get to be routing leaders without significant investment over many years.
AvidThink will be watching the dynamics in the space to see how all of this will play out and whether Marvell can climb its way past the new contenders to duke it out again with Broadcom. The combined melee between Broadcom (with no chance of acquiring Qualcomm for now), Marvell/Cavium, and nVidia/Mellanox is made all the more interesting by the presence of SAI and SONiC.
Unfortunately we don’t know who the wicked queen would be in this mix, and please don’t ask us to name all the dwarfs or associate names of dwarfs with merchant silicon vendors. Though you’re certainly welcome to message us with your votes for which one is Grumpy.