Why Arista Doesn't Like ASICs

SAN JOSE, Calif. -- The best way for networking to catch up with Moore's Law is to abandon Application-Specific Integrated Circuit (ASIC) design, according to Andy Bechtolsheim, a founder of Arista Networks Inc.

That opinion isn't surprising, considering Bechtolsheim is betting Arista's success on merchant semiconductors -- i.e., non-ASICs.

Still, Bechtolsheim is a Silicon Valley celebrity, having also helped found Sun Microsystems and a key, early Cisco Systems Inc. (Nasdaq: CSCO) acquisition, Granite Systems, so he was bound to hold his audience captive. In a keynote Thursday at the Linley Tech Processor Conference, put on by The Linley Group , he touched on a wide range of topics related to today's data-center networks, loosely centering the whole talk on Moore's Law.

His main point was that networking hasn't kept up with Moore's Law. That's not necessarily surprising, considering Moore's Law is about transistor density and doesn't promise anything about system performance.

Still, Bechtolsheim said it's annoying how networking has been so far behind Moore's Law -- with speeds taking 12 years to go from 1Gbit/s to 10Gbit/s -- and the way to catch up, he thinks, is to avoid using ASICs in networking.

"I'll just flat-out say ASIC designs will never be on Moore's Law," he said.

The difference is in the way chips are designed. ASICs have an established procedure behind them, a sequence of software tools that build the chip. Specifically, the chip's functions are dreamed up independent of chip layout, Bechtolsheim said.

That means that certain considerations, such as the chip's clock rate, become secondary. The way to overcome that, if clock rate happens to be important to you, is to not use the ASIC design process.

That's why Arista can find more suitable chips for its needs by buying off the shelf from the likes of Broadcom Corp. (Nasdaq: BRCM) and Intel Corp. (Nasdaq: INTC), he said. Bechtolsheim didn't mention this explicitly, but the message was clearly that these design differences are what, in his view, make Arista's approach superior to the ASICs designed by Cisco, Juniper Networks Inc. (NYSE: JNPR), Alcatel-Lucent (NYSE: ALU) and others.

"We are actually getting back on Moore's Law here, by a combination of packaging technologies and, let's say, a more competitive industry where people are getting these chips out faster," he said.

Competition among switch-chip vendors does seem to be on the rise, with Intel (through the recent acquisition of Fulcrum Microsystems) and Freescale Semiconductor Inc. apparently taking on Broadcom.

That Intel acquisition is important, because Bechtolsheim also noted that the creation of more complex, "fancy" chips can now cost $20 million, whereas it used to cost just $1 million. "If you can't sell 100,000 chips, it's hardly worth doing," he said. "What this all means is that some of these smaller vendors can't design the chips, because they can't get the volume."

Where 100G went wrong
Moore's Law was the general topic of Bechtolsheim's talk, but really, he wandered through several topics relevant to high-end networking, speaking quickly in his soft German accent.

Regarding high-speed networking, Bechtolsheim said he doesn't see the data center being the driver for speeds of 400Gbit/s and faster. The throughput in a fabric woven of 100Gbit/s connections is massive enough that a jump to something faster doesn't seem necessary yet; rather, it's the long-haul connections that will drive speeds up to 1Tbit/s, he said.

Bechtolsheim added an aside about silicon photonics. His big hope for the technology is that it improves the state of 100Gbit/s optical interfaces; he complained that there are too many types already (citing CXP, ZHD, QSFP2 and the three types of CFP) and that they're too costly, with longer-reach interfaces running at $50,000 per port.

Afterwards, Bechtolsheim graciously hung around to answer questions in the hallway. He seemed to relish the interaction, even when some of the questions struck topics he found trite or unworthy.

OpenFlow was one such topic, as Bechtolsheim considers it overhyped. "We actually support OpenFlow. You can run OpenFlow on our product. It's just that nobody uses it. We did it for one customer," he said.

— Craig Matsumoto, Managing Editor, Light Reading

fozzy 10/20/2013 | 4:22:47 AM
Moore's Law The comparison is flawed. Moore's Law was actually an observation about transistor count doubling every two years. This translated to faster processors and more RAM. Why should network speeds follow this trend? Distances in the real world can't be shrunk, which is why satellite Internet will always be slow.
netkopf 12/5/2012 | 5:19:18 PM
re: Why Arista Doesn't Like ASICs

Andy is speaking conventional wisdom.  What's interesting about that? I guess cuz he's Andy. And it's probably not what the Linley Group wants to hear, since alot of their employers will be likely crushed by the coming commoditization.

What i want to know is: what will stop the datacenter/sp/enterprise switch vendor industry from turning into a Broadcom distribution outlet? Who can compete with them in making switch chips? I dont see Intel/Fulcrum doing it, fulcrum had some serious methodology issues that seem to have set them way back. Marvel is way behind. I dont know of anything from Freescale. Was Freescale a typo, btw?


gtchavan 12/5/2012 | 5:19:18 PM
re: Why Arista Doesn't Like ASICs

Does this guy getting a better price on merchant ASIC than the rest of the network hardware makers that use merchant chips that have never made a GAAP profit with today's accounting standards?

MMQoS 12/5/2012 | 5:19:14 PM
re: Why Arista Doesn't Like ASICs

I have the highest regard for Andy and his knowledge of both networking and silicon so I'm dubious of the comment attributed to him about Ethernet ".....speeds taking 12 years to go from 1Gbit/s to 10Gbit/s ".  The 1GE standard, IEEE802.3z was finished in 1998.  The 802.3 Task Force completed the 10GE standard, 802.3ae four years later in 2002 and at Nortel we already had a 10GE switch port based on ASICs running in the lab as well as a 10GE adapter card.  The same was true for switch ports at Force10 too.  The problem was not time, but BOM costs, especially optics.  Copper 10GE for LANs helped somewhat but the market volumes for 10GE nodes were not yet in place and so the prices remained high for silicon.  Some of the same vendors mentioned in this story were early developers for merchant silicon but again, awaited the necessary volume to bring thier costs down.

But if that is what Andy said then he is comparing apples to oranges.  Demand drives the pricing for semiconductor components so IMHO, regarding Moore's Law, the market demand for faster PC processing in the early part of the 21st  century was much greater than the demand for faster (>1G) Enet nodes.  The volume in the Ethernet chip market, like the market for processors, is in the PC and at that time, PC's were just moving from 10Mb/s to 100Mb/s on motherboards.  Even 1GE adapter nodes were a small volume market.

I'd say then that Ethernet speeds are close to following Moore's Law, moving from 1GE to 10GE in about 10 years.  The explosion of the blade server and cloud computing centers has played a large part in driving the 10GE volumes. 




Pete Baldwin 12/5/2012 | 5:19:13 PM
re: Why Arista Doesn't Like ASICs

Desmoden -- yep.  The "jack of all trades" argument is the usual counter to the merchant silicon case. So far, it's proven out well for Cisco et.al.

Pete Baldwin 12/5/2012 | 5:19:13 PM
re: Why Arista Doesn't Like ASICs

netkopf -- Oop, freescale was indeed a mistake.  I was cut-and-pasting company names out of a standard list we use (so that the first reference will be the 'full' name) and grabbed for the wrong name.  Thanks, I'll go correct that.

Desmoden 12/5/2012 | 5:19:13 PM
re: Why Arista Doesn't Like ASICs However, I think when Cisco did this http://low-latency.com/article... a few weeks ago they showed that once again it's not always good to be a "jack of all trades" as merchant silicon is. For specialized goals, latency, heat, power etc, custom ASICs can and will still dominate.
co123 12/5/2012 | 5:18:50 PM
re: Why Arista Doesn't Like ASICs

It looks like Andy still trying to push the merchant silicon agenda after Cisco came out with their own ASIC to beat them in the HFT market (7150 vs 3548).  Keep in mind if your are using merchant silicon, that means EVERYONE is going to be the same hardware capabilities, then you need to differentiate yourself in the software stack.  There still a lot of value in some cases to develop ASIC...

Sign In