George Gilder, editor of the Gilder Technology Report, regularly responds to subscriber questions at a subscriber-online chat. Here are some excerpts from Gilder’s recent question and answer session on the topic of network processors and companies that are leading the move to win “share of network.”
GTR Subscriber Question: Why would IBM start selling its interest in EZchip after all these years, when EZchip is ready to ramp up with the NP2 or are they?
George Gilder: IBM was a partner with EZchip on the first-generation, NP1c, which was based on IBM’s on-chip dynamic random access memory (DRAM) cell technology and an application-specific integrated circuit (ASIC) design and pricing process. For the NP2, EZ abandoned this IBM relationship because of cost concerns. Although the on-chip DRAM gave the NP1c world-beating memory bandwidth, the NP2 has adequate memory channels and is easier to produce in volume. EZ is no longer an IBM partner or a fab contractor for future chips. I would assume that IBM’s sales of LNOP shares reflect these changes.
The general prospects of the company have improved as a result, though during my trip to Silicon Valley last week I discovered widespread skepticism about the overall network processor market, which IBM long-ago abandoned, and which is being downplayed by Intel and others as well. This is good news and bad news. The venture capitalists and in-house financiers will be less willing to fund rivals to EZchip, but to achieve its goals, EZ must gain a large share of the market. Over 60 design wins would seem to give them this share. But it is not a sure thing.
The networking industry was shell-shocked by the crash and remains incredibly cautious about its prospects. It will not believe in a resurgent market for network processors until it happens. This will give EZchip a lot of leverage for an upside surprise as broadband is rolled out and the “access” market, wireline and wireless, becomes a 10-gig arena.
GTR Subscriber: All, the big three have moved toward Network Processing Units. So, why the “not believe in a resurgent market for net processors?”
George Gilder: In an informal presentation last week in Silicon Valley, I ventured the idea that network processors would hollow out Cisco Systems routers (and deplete their margins) in the same way that Intel hollowed out the PC industry and depleted its hardware margins. I suggested that some NPU companies such as EZchip, Broadcom or Raza Micro could emerge as the Intels, AMDs and Motorolas of the new era. This in essence, was the proposition behind my possibly imprudent choice of EZchip as my new Qualcomm .
EZchip still has good prospects to become a ten-bagger over the next few years. But the path will be more treacherous than implied by my original analogy with Intel (NPUs are the new CPUs). My Silicon Valley critics, who as venture capitalists and router strategists had turned down various chances to invest in network processors, claimed that the router market is simply orders of magnitude smaller than the PC market. There are hundreds of millions of PCs and billions of embedded microprocessors, while there are perhaps a million routers and switches of all kinds produced per year. Cisco knows the market best and can produce most of its own NPUs.
A man from Procket Networks, who believed he had designed the world’s leading terabit router in the late 1990s, figured out that just two of them could manage all the data traffic in Asia. He sold out to Cisco (cheaply) and they dissolved his team. (Now at Mobilygen, he is making low-power H.264 high-definition video encoders for mobile applications.) The Linley Group report on NPUs, the canonical source on the subject, estimates a total market of under around $500 million for merchant NPUs.
Moreover, real-time wirespeed network slots need a different kind of central processor than PCs do. Wirespeed devices require specialized hardware. With every turn of the technological rachet, the network processor market changes. Now it’s XML processors; last month it was security processors, knowledge processors, DSLAM backend processors, 3G wireless processors, residential gateway processors—each with new wirespeed requirements.
I have assumed that EZchip with its capacious headroom and seven layers of programmability could win many of these slots, but the folks from NetLogic , Cavium Networks, et al contend that the demands are so exacting and specialized that a general purpose NPU comparable to the Pentium is impossible. A market fragmented into special-purpose devices offers less opportunity for EZ’s general-purpose seven-layer design.
Thus from Intel to Motorola C-Port, to Applied Micro Circuits to Vitesse , and on and on, strategists have contemplated the general purpose NPU market and decided against it. Instead, Silicon Valley is producing an array of low-end access devices (Wintegra, AMCC) and high-end co-processors (Cavium, Raza, NetLogic) and specialized hardwired Ethernet switches (Broadcom, Marvell). Only Broadcom, with their Sandburst acquisition, shows a serious intent to compete with EZchip in high-end general-purpose NPUs.
That’s the consensus, but I think Silicon Valley is wrong. What Silicon Valley misses is the premise of my EZchip choice from the beginning, namely widespread deployment of broadband and high-definition everywhere. Silicon Valley is building for a world of television and data, but the world is moving toward life after television, where video is just another form of data on the network. Every node in the enterprise and metro is going to have to switch at ten gigabits per second to handle the sea change in the network from narrowband voice and slow-data applications to various forms of full-motion video. It’s this change that will create large markets for Broadcom, Sigma Designs , Texas Instruments , Corning and much of the rest of our list. It will also create volume markets for ten-gig NPUs.
Emerging over the next five years in the context of the existing menagerie of networks, the rachet in broadband traffic will require millions of network processors that can handle more than just Ethernet connections. (For Asia at least IPV6 is necessary, as EZchip foresaw.) NPUs will have to offer rudimentary seven layer functions, including some security and packet scanning, but they will have to be on a volume track toward pricing for the residence and the enterprise. For high-end slots, they will have to interoperate with the custom devices from Cavium et al. EZchip is the best situated company to take advantage of these opportunities as they arise. Perhaps the supposedly runaway success of Huawei-3Com is a portent, making 3Com an inviting turnaround investment story and enhancing EZchip as a supplier. My view is that the huge bet that Silicon Valley in making on high definition IP products is going to require a similar response in the network.
And because Silicon Valley was traumatized by the crash, it will not respond until too late. Broadcom and EZchip will have the field mostly to themselves. And EZ is the one with most of the design wins today.
Excerpted from posts on the Gildertech.com GTR subscriber-only message board.