Chip startup Xpliant is indeed working on a programmable competitor to Broadcom‘s Ethernet switch chips, betting on a need for protocol flexibility as software defined networking (SDN) continues to develop.
We relayed rumors about Xpliant’s plans last week, in a story that emphasized greater question of how semiconductor startups can get funded.
Cavium, a vendor of multicore processors and security accelerators, recently announced it’s acquiring Xpliant, a deal expected to close before April. Today, the companies are revealing details of what’s now called the XPliant Packet Architecture. They’re also announcing a new spelling — XPliant, with a capital “P.”
We did have a good idea of what the architecture is like. But the capital “P?” Totally did not see that coming.
Programmable Switch Chip
Former Marvell employees, including Guy Hutchinson and Sachin Gandhi, founded XPliant based on the question: “How can we continue to innovate in networking without spinning up a new chip every year?” says Eric Hayes, vice president of Cavium’s switch platform group.
Their answer was to make the chip programmable. Ethernet chips are rigid when it comes to things like the formats of lookup tables or the options for editing packets. This actually makes sense, because Ethernet is a well defined protocol that doesn’t change quickly.
But as software-defined networking (SDN) has progressed, protocols have been shifting. OpenFlow 1.2 required new hardware, and one proposal for OpenFlow 2.0 would involve a new type of switch, one that can adapt to different protocols. (Barefoot Networks, an XPliant competitor, is working along those lines.)
There’s also the case of the different encapsulation protocols that have emerged — including Geneve, which emerged in February aiming to unify VXLAN, NVGRE, and others. As if to show off the point of its programmability, XPliant’s first chips will support Geneve. (Intel also showed Geneve in hardware at the recent Intel Developer Forum.)
A programmable switch chip could also opens the door for systems vendors to stand out from one another. Most high-end systems now use Broadcom’s XGS series, including the high-end Trident II, and that leaves little room to differentiate. The systems “might have different colors on them, but they all taste the same,” Hayes says.
Finally, programmability could extend the chips’ lifetime in the network. As time passes and 3.2 Tb/s stops being the top of the heap, those chips could be moved, say, out of the network core and closer to the edge — and they could be reconfigured for different tasks as well.
Not a DIY Project
XPliant’s plan isn’t just to hand over the chips and assume systems vendors will figure out how to best configure them. Rather, XPliant will do all the programming, providing chips preconfigured for particular tasks — Layer 3 MPLS routing or plain Layer 2 switching, for instance.
That will let systems vendors get to market quickly, Hayes argues. It’s a departure from the model of network processors in the early 2000s, where it was assumed that the customers would program the chips.
In the long run, it’s possible that systems vendors will program the chip themselves.
“Our customers can program in a language that’s very much like C. When we’ve brought in customers to program in that language, they get it,” Hutchinson says. (Machine language is another option, but of course it would require an even higher level of expertise.)
Programmability wouldn’t be enough unless XPliant could also compete at the high end. The company says the throughput of its chip — 3.2 Tb/s for the highest-end CNX88091, which beats the 1.28 Tb/s of the Broadcom XGS architecture.
The next version of Broadcom’s Trident family, reportedly code-named Tomahawk and expected to launch soon, is rumored to match the 3.2 Tb/s figure. The bigger question, then, will be what it would take for Broadcom to match XPliant’s programmability. It’s possible it could require a redesign, which would take time to produce, says Bob Wheeler, an analyst with the Linley Group.
XPliant is arriving at a time when Broadcom really does dominate the market, so it might be facing a customer base that’s hungry for an alternative.
“There’s been a hole,” Wheeler says. “[Broadcom’s] Trident line, according to our numbers, had about 90 percent market share last year. Intel is trying and is currently investing. It’s less clear what Marvell’s plans are here.”
XPliant’s chips are due to start sampling in the fourth quarter. For now, customers are developing around the chips, using software simulations.
“My estimation is that before the end of 2015, you’ll see [end] products shipping in production,” Hayes says.