The Xpliant deal sprang from an effort to rethink the way chip companies are funded. It’s an idea I’d reported on nearly three years ago, and while the new magic recipe still hasn’t been found, it’s exciting to see some progress on the concept.
It’s also exciting to see that Cavium is entering the Ethernet switch market, which has been dominated by Broadcom. Xpliant isn’t talking about its technology just yet. But the way the company got funded and nurtured is a story worth telling.
Chips Take Time, Chips Take Money
When it comes to startups, the “silicon” in Silicon Valley is drying up. “The number of series A startups that have gotten funded — the traditional number used to be 50 or 60 a year,” says Dennis Segers, CEO of chip startup Tabula. “In the last five years, it’s been more like two. Nobody gets funded. It’s a real shame.”
The problem is that a chip startup is expensive, even in these days of fabless operations (wherein chip companies outsource all their manufacturing — Broadcom being a classic example). Chips are so complex that getting the first copies produced can take tens of millions of dollars and a couple of years of development.
The “couple of years” part is the real killer. Here’s one example: The next version of the ASICs at the heart of Alcatel-Lucent’s routers will spend at least one year in simulation, says Basil Alwan, president of IP routing and transport. That’s after the chip is designed and before the complex process of manufacturing kicks into gear — one year of checking to make sure the chip works properly.
“We spend a lot of time iterating simulations — fix, simulation, fix — before we cut the chip, because we know: If we cut the chip and miss, and we incur another fab cycle time and another simulation time, it’s at least a one-year slip. And a year can be fatal,” Alwan says.
Even to work on a product like Xpliant’s — an Ethernet switch chip that’s configurable to work with nonstandard header fields and new protocols — you’re going to need plentiful cash and time. And your reward, at the end, is to battle the immovable object that is Broadcom.
That’s why venture capitalists, during the 2000s, turned away from semiconductors. Why bother, when you can get better returns from two guys writing iPhone apps in their kitchen?
By 2010, the problem was acute. The startups that had helped fuel the communications boom circa 2000 were mostly acquired or starved to death, and there was no new crop arriving.
But there was a sense around the industry, even among venture capitalists, that it didn’t have to be this way.
“What are the probabilities of building a large successful business if you get a team that has a track record in that space, has a product, and has a business model that extends into IP [intellectual property] and beyond? The odds can be pretty good if you can afford the cost,” Segers says.
The Company as a Platform
A few different models tried to attack the problem. The Global Semiconductor Alliance tried to find ways for startups to avoid building the parts of a chip that were unremarkable — pieces such as I/O and memory subsystems. At least one venture firm considered starting a fund with money from systems companies — customers of chip startups.
Amer Haider, who then led the mergers and acquisitions efforts at Cavium, was actively involved in trying to get these models off the ground. Most of them didn’t pan out or are still works in progress. But Cavium tried something on its own that turned out to work.
It wasn’t about providing cash or shortening the design cycle. The idea was to attack the factors that had starved out so many chip startups.
“Where most companies get screwed is that they don’t have enough steam in them to secure design wins, because people are afraid [the startup is] going to die. For smaller chip companies, you need a platform that will enable them to build a product, provide access to customers, and get them the feedback they need,” Haider says.
That’s “platform” not in a technological sense, but a business sense. Xpliant used Cavium as a springboard for the business side of being a chip startup. Cavium did provide funding, but more importantly, it provided a sign to the outside world that Xpliant had friends that could help it remain stable and growing.
Maybe most important of all, Cavium provided access to potential customers who could guide Xpliant toward the things they wanted.
Xpliant’s founders didn’t come from Cavium, but from Marvell, so the model was not the same as the spin-in concept that bred Cisco’s Insieme or Alcatel-Lucent’s Nuage Networks. “I would say ‘foster’ is a good word — fostering a company to provide needed solutions,” says Haider, who has since left Cavium on his own to pursue personal interests in medical technology.
A Chip for All Seasons
Of course, before you need funding, you need to be doing something worth funding. Cavium has disclosed little about Xpliant’s technology, but the general idea has been circulating around the rumor mill for some time.
An Ethernet switch chip operates by examining values in the Ethernet header. What it finds there determines where to send the packet or what to do with it. The definition of that header has always been rigid, though. Any change in protocols (or, heaven forbid, any attempt at something besides purely standard Ethernet networking) would require a respin of the chip.
What we’ve heard is that Xpliant’s chip can be configured to interpret packet headers or frames differently — or to adjust to protocol changes, something that became a sore spot in software-defined networking (SDN) as the early versions of OpenFlow arrived in rapid-fire succession. It’s not fully programmable like the network processors from EZchip or Netronome, but it could create a new level of flexibility.
Barefoot Networks is aiming for a similar goal, based on research from a Stanford team that includes Nick McKeown, the professor who helped OpenFlow come into being. But while Xpliant is aiming to have chips sampling in the fourth quarter of this year, Barefoot is still in an early stage.
Configurability does add heft to the chip. But an Ethernet switch chip is no longer just a simple Layer 2 device. Broadcom’s Trident II includes Layer 3 processing, access control lists, tunneling capability in the form of VXLAN support, and stacks of other features.
It’s believable, then, that a configurable chip like Xpliant’s — assuming it does what we think it does — could be comparable to Trident II in performance, size, and power consumption. “The problem with an Ethernet switch today is that they’re carrying a lot of baggage because of the amount of stuff that’s been added over the years,” says Bob Wheeler, an analyst with the Linley Group.
Does that mean Xpliant is onto something? That’s now Cavium’s job to find out. It won’t be an easy fight, bringing a brand new chip to market — but you could say Xpliant already got the hardest part out of the way.