Article Info

Battery Man

I sat down with Alex Benik of Battery Ventures Tuesday morning to swap perspectives on the Communication components business, semiconductors in particular.

We lamented the growing barriers to entry for fabless startups, and talked about what needed to be done differently to make Telecom chip startups a reality. While everyone is avoiding Telecom chips like the plague, investors have not lost their moon-shot mentality of funding 6 businesses all chasing the latest IEEE standard (in this case, Copper 10GE Ethernet transceivers). Two land successfully and folks forget about the four that didn’t.

Other than Cortina, and the FTTH PON guys, there have been virtually no Telecom semi startups in the past 5 years. Zero. Not only that, but the incumbents have been doing little except make duals, quad, and octal port versions of the same designs from 2002.

Pundits proclaim the rebirth of optical yet no money is flowing. Which either means the pundits are wrong or a very big opportunity is lurking out there.

I’m convinced that a massive innovation opportunity exists in this sector that either an incumbent or startup can capitalize on. As is usual for me, I think it exists in the dirtiest and ugliest sector of all – Optical Transport. If you have an idea, I suggest you speak with Alex, because private investments just are not our bag.


Comments are disallowed for this post.

  1. This post will be a question. By next post will be a comment.

    What exactly are you talking about in terms of the “growing barriers to entry”? Are you talking about the difficulty in obtaining financing, which can change with the wind; or, are you talking about something more structural, such as the increasing complexity and intellectual property needs of these chips?

    Posted by Herb Chen | February 23, 2007, 4:18 PM
  2. The big one is the growing cost of developing as process geometries shrink. 90nm proto lots cost near $1MM. Therefore, you need to cram as much functionality into a device in order to amortize the fab costs. This makes the development of the device more expensive.

    I’m sure a more detailed analysis exists, but the cost per gate to develop a chip has not dropped at the same rate as fab capabilities. Companies with scale, size, and access to capital have an advantage.

    The difficulty in obtaining financing is directly related to this problem. If someone had a great idea manufacturable in 0.18u funding would not be as difficult. It is also possible that someone can break the development cost barrier as well.

    Posted by Andrew Schmitt | February 25, 2007, 7:24 PM
  3. Andrew-
    I’m not sure I agree with your funding comment. I have heard of a lot of startups (optical and otherwise) who are “pushed” into 90 nm CMOS by their VC’s.

    If you’re looking for great ideas in 0.18um technology, it requires internal R&D funding. And opportunities are there.

    Maybe the fact that the VC’s are unwilling to invest in passe technology has led to the extinction of optical chip startups.

    But, I do agree that there hasn’t been much new for these guys to spend money on in the optical market.

    Posted by DC | February 26, 2007, 11:58 AM
  4. I think it is not so much about shrinking process geometries as about simple economics. In my mind, why on earth would anyone want to fund a startup when public market valuations are as depressed as they are?

    The end game for a venture capitalist is taking a company public at an exalted valuation. The easy money is made when markets become frothy and indiscriminate, as they were in the 2000 period. VCs could bring companies public sooner and at considerably higher valuations than ever before. Even better, they could unload dogs that would never have seen the light of day in a more normal investment environment.

    What about today? Why fund a company when you get buy R&D for 50 cents on the dollar in the public market with a free look at the technology? What’s more, you often get a legacy business, an established customer base and a bunch of NOLs thrown in for free. If funding fabless comm IC startups is to be a good business again, then the existing stock of companies will have to be valued much more highly in the market. Why not start one’s investment search there? If I were a VC investor, I would be trolling through the public markets. Just as sometimes it is cheaper to buy oil on Wall Street than to drill, so may it be cheaper to buy promising technologies in the public market.

    The dream of a comm. IC venture capitalist is to fund a talented group of engineers who develop cutting edge products and achieve marquis design wins with blue chip customers. Let’s say you invest $50mm into a start up chip company. I am not a VC so I am just guessing at the numbers, but for starters you probably aren’t going to own more than 1/3 of the firm. The other 2/3 is going to the founders and employees who are going to work night and day to make everyone’s dream a reality. So overnight, you’ve written a check for $50mm, and you own a 1/3 interest in a firm capitalized at $150mm, and you’ve still got nothing but cash in the bank, some smart guys, an idea, and a lot of hard work in front of you. Now the real work begins. You’ve got to design a chip, get it to work, hope it’s timely to market, hope there is a market, hope you get the design wins that matter with the customers that matter, and hope the time to money won’t be ridiculously long. Over the years your company will take that $50mm and probably $35mm will be spent on R&D and the balance will be spent on rent and administrative and marketing costs as you try to build a real company.

    If you are good and smart and lucky, you will make a 20 % return on your money. For this to happen, you’ve got to be able to sell your company for about $500 million in six or seven years. In other words, an initial investment of $50mm must somehow morph into a company worth $500mm. To get this kind of return, you’ve got to buy low and sell high, with ultimate selling valuations largely determined by the public market.

    Currently, the public market for comm. IC stocks is horribly depressed, and it would be difficult to bring a company public at an attractive valuation. It has been so long since these companies enjoyed a normal business environment that no one can imagine better times, and no one is prepared to give these companies credit for anything good that may happen in the future, but as any long time market participant knows, everything goes in cycles and better times almost surely come.

    Let’s look at TranSwitch (TXCC), a large holding of mine. There are no doubt other companies to which this analysis would apply, but I will discuss the one with which I am most familiar. The shares have been languishing for the past few years, with a current enterprise value of just $150mm. From a public markets perspective, there is just no interest. Revenues have been stagnant and unless one looks carefully, it appears as if nothing good is happening. No one wants to buy it because no one else is buying, so the shares languish. What I mean is that everyone says why should I buy this stock when revenues are flat? Why don’t I just wait until revenues actually start rising, and then I’ll be all over it. That’s a public markets perspective at times, although it isn’t always that way. Remember back in the bubble when the merest hint of something positive, no matter how trifling or ephemeral, was enough to send shares skyward? When stocks are moving, people will find a reason to buy; when they’re languishing, they’ll find a reason not to.

    But now let’s look at it from a venture capitalists perspective. Over the past six years, since the bubble burst, this company has spent over $230mm in R&D, virtually all of it directed towards Ethernet-over-SONET and carrier Ethernet, two of the key enabling technologies in next generation networks. They also bought Mysticom which itself had invested over $90mm of R&D. Today, you can buy it all for 50 cents on the dollar, and that’s without giving any credit for the significant sales and marketing infrastructure,
    large intellectual property portfolio, and established customer relationships. Add in about $300mm in tax loss carryforwards, and you’re talking some pretty serious value.

    But most importantly, you are buying a company with a free peek at what all this spending has produced, and that goes a long way towards mitigating risk. In the case of TXCC, all the heavy lifting has already been done. The products have been developed, hundreds of design wins have been won, and the arduously long time to money has been mostly suffered by someone else. They have a blue chip list of customers and key design wins at Fujitsu and Alcatel, two of the most important customers in the space. They will be a big beneficiary from British Telecom’s 21CN build out, and management has stated the expectation for in excess of $100mm in revenues from this project, an amount that could potentially double revenues from the current run rate and bring the company well into profitability. Since then, the company has indicated it will have content in ADVA’s FSP150 box, which will form the foundation for the Ethernet access device used in this deployment, an opportunity likely worth an incremental $20mm.

    If the R&D hasn’t been on target, the company wouldn’t be playing such a prominent role in this and other deployments; if the R&D is on target, there will be a lot more revenues to come from as other carriers build out their next-gen networks.

    From my perspective, it is hard to see how an investment in a de novo chip company can yield anything near the returns that are available in the public market with much less risk, and maybe that’s why venture funding in this area is difficult to come by.

    Posted by Herb Chen | February 28, 2007, 11:50 AM
  5. Herb:

    I agree with the concept – “Why buy private when you can get public for 1/2 the price?”. R&D is a sunk cost. It’s value is diluted if it is duplicated in five other companies. Unique R&D that is sunk (i.e. company has a jump on others) gets a multiple. All R&D is not created equal…

    I think the issue is there is too much money in the existing public companies chasing too few ideas. Someone sniffs a good idea and three companies pounce. Only afterward do they look to see if what they grabbed is worthwhile. In general I don’t believe the VC’s are seeing proposals for unique R&D.

    That being said, if I was a VC, I would consider taking a public company private and using it as a shell to launch new ideas. That is similar to what Cortina is doing, and why I think they are such an interesting company.

    I don’t think the VC guys are foolish enough today to fund duplicative R&D that already exists in public markets. They seem to have no hesitation funding multiple startups pursuing the latest 802.x standard. I think the path to creating shareholder value is consolidating the duplicative R&D and market share.

    Posted by Andrew Schmitt | February 28, 2007, 4:59 PM
  6. “I don’t think the VC guys are foolish enough today to fund duplicative R&D that already exists in public markets.”

    I think the VC cycle is just like any other financial cycle, subject to the same manic highs and lows. When returns are great, as they were in the 1999/2000 period, investors throw money at venture capital funds. VCs, inundated with cash that they have to invest, inevitably lower their standards. It doesn’t hurt either that every dogcrap company they’ve funded in the last five years is going public at an exorbitant valuation. After all, there’s nothing like success to make you think you’re smart. And so begins a period of lower returns.

    I think the issue is really one of where we are in the cycle. In essence, pick a cycle, any cycle. If you invested in energy in 1980, or real estate in 1989, you suffered poor returns. If you invested in energy in 1998 or real estate in 1994, you had great returns. I think you are giving the VC community too much credit when you say they “are not seeing proposals for unique R&D.”

    I think there are a couple of important issues. One is that when the industry is doing well, R&D yields a good return; and when the industry is doing poorly, it doesn’t. Another is that in the wake of the superb returns of 1999/2000, a lot of money was thrown at the industry, yielding an overinvestment in that time period, and consequently low returns. In the last few years, there has probably been underinvestment, and I’ll be returns will be pretty good.

    Posted by Herb Chen | February 28, 2007, 9:33 PM
  7. The reality is that Mr. Chen’s statements about VCs are right. The only reason the VCs still run around stating that they are looking is to do their job today, which is show deal flow to their limiteds. The reality is that Mr. Chen’s statements about VCs are right. The only reason the VCs still run around stating that they are looking and moaning about the lack of “quality” deals is to generate deal flow activity. They are really not funding much and mainly C and later rounds. Meanwhile, to earn their pay, they will need to show “deal flow”.

    Also, as far as tech and communications is concerned, the public market offers better opportunities. The overhang of VC funded companies are the best place to get already funded R&D at closer to 20 cents on the dollar that the 50 cents that Mr. Chen mentioned.

    At the same time I don’t think these public companies that Mr. Chen refers to should be public entities. They need to be taken private, cleaned up, consolidated with a few others and then put back in the public markets. However, no private capital wants to do this in tech for small companies. They are too scared of tech, especially if it is a small company. Freescale and NXP are of a different scale.

    On Mr. Chen’s eos comments I sadly think it is not going to take off. I believe he is wrong. Native Ethernet is the way to go and also the eos that is happening is at at OC-48 and OC-192 speeds.

    Posted by JV | March 1, 2007, 4:52 PM
  8. Agree with above, with the exception of native Ethernet as a transport technology. If you’d like, we can step outside and settle that argument…

    Posted by Andrew Schmitt | March 1, 2007, 5:02 PM
  9. I like your pugilistic energy, but carriers have been quietly selling ethernet service for some time now – both point to point and switched, by ethernet switches. If they have to go across a region they typically have to go through an IXC and hence need SONET. Not a big deal to add some line maintenance features or overhead. Don’t have to go to the extent of Sonet. TDM had no higher layers such as TCP.

    Yes Ethernet does not have the TDM jitter control features. But who needs that ?

    I am also not sure where Mr. Chen is coming from on TXCC. From what I can see relying on EoS to pull them out of the revenue hole is a lost cause.

    Posted by JV | March 2, 2007, 2:29 AM
  10. JV – just kidding BTW. People who know me know I have passionate views about the role of EoS.

    The fact is that Ethernet as a layer 1 transport technology offers little cost benefit and no performance benefit. TDM transport allows Ethernet and legacy TDM to co-exist. The carriers will not rid themselves of TDM anytime soon, primarily because they have no desire to build a heterogeneous transport network.

    Cable guys, without a single TDM container to speak of, have a different outlook.

    So, both will play a role, Ethernet in greenfield networks and as remote DSLAM connectivity. But EoS will migrate the installed base and comprise the majority of capacity additions in the transport network. This is exactly what BT is doing with 21CN.

    Posted by Andrew Schmitt | March 2, 2007, 8:45 AM
  11. Agreed and we need a sense of humour in this biz.
    I was told directly by a major ILEC that they are selling all the point to point ethernet that they can and the only time they need to wrap sonet is when they have to cross a region and go via and IXC who can only take it sonet wrapped.

    I agree that they are using their legacy switches and thus sonet will continue. However, as you said any new capacity is going to be packet based. I cannot see ATM based Sonet switches being deployed anymore.

    But access deployments are basically ethernet or packet – both VDSL and PON so when this traffic is aggregated, it is in the OC-48/OC-192 range and so EoS @ lower rates is a very small market.

    Posted by JV | March 2, 2007, 10:41 AM
  12. JV,

    I am not a technologist but it is my feeling that if Ethernet becomes as all pervasive as it appears to be coming, there will be room for many flavors. Any way you slice it, it’s not going to be all native Ethernet or all EoS. These are big markets and there’s room for a lot of people to succeed.

    As Andrew points out, EoS will play a big role in 21CN. On their third quarter conference call, the company noted they anticipated in excess of $100mm in revenues from this deployment, with the most important chip being the Etherphast-48, which is an EoS device. (As you correctly point out, deployments of EoS at OC-3 have been disappointing.)

    In a white paper on their Geostream platform (the one that BT is using), Fujitsu writes,

    “One of the fundamental principals of a Next-Generation Network (NGN) is the ability to offer services independent from the underlying delivery technology, and in themain this means convering on IP as the common network layer protocol for everything.

    “However it is not possible, except in pure greenfield applications, to simply remove and replace all the existing end-user terminals in the network–the cost of doing so would far outweigh the benefits gained by moving to an NGN. This creates a requirement to maintain support for legacy access technology and to convert to IP as close to the edge of the network as possible, as defined by network economics.”

    In either case, I’ll let Andrew do my fighting for me, because he is far better equipped. He is also an avid cyclist so I know he is in shape. The point of my post was really to address what I think are some glaring valuation discrepancies between public market valuations and potential startup valuations. $100 million in revenues for TXCC is a lot for them, even if it is over four years or so. To think that it stops there presumes that BT is a lone wolf in the wilderness. As I am sure you know, BT has started their 21CN consulting division which is, at least according to them, being well received. It’s hard to imagine that there won’t be some business coming from there. If you are a venture capitalist, these are home run developments, yet thte public market is ignoring them.

    As for the TranSwitch product line, like I said, I’m no technologist, but they have found a good home in a number of products that will be used in 21CN. From my perspective, it seems to me that the issues that BT is tackling will be faced by other carriers, and that bodes well for future business possibilities.

    Finally, TXCC isn’t all EoS. Their Envoy line addresses carrier Ethernet requirements, and their CE2 product is designed into ADVA’s FSP150, which could see meaningful revenues as it is deployed in 21CN. They are also coming out with a GPON product line. It’s hard to evaluate these products because they haven’t been announced, but the market is treating the matter as a foregone conclusion that the R&D was a waste of time.

    From an investment perspective, we’ve gone 180 degrees. In 2000, valuations were so absurd that they discounted the hereafter. Today, valuations are similarly absurd, but in the other direction. In late 2000, TXCC was valued at $6 billion on the basis $55mm of quarterly revenues, even as revenues were about to fall off a cliff. Today, TXCC is valued at $150mm on the basis of $9mm in quarterly revenues, with the likelihood of a significant ramp in revenues. Reversion to the mean is a powerful concept.

    Posted by Herb Chen | March 2, 2007, 11:30 AM
  13. Herb,

    No wonder Andrew was willing to step outside – he wanted to do the fighting :)

    Just got back in and read your note and agree with many points. Will put some additional thoughts down later.

    Posted by JV | March 2, 2007, 2:44 PM
  14. OK on TXCC here goes

    1) EoS we have discussed. We agree except for the following – that at OC-48 and higher rates, TXCC is out as I believe the larger competitors have products out earlier that have captured the sockets.

    2) PON – in Asia the action right now is in Japan and from the PMCS conf call, they appear to have that business locked. I assume that there are others hanging at the door, trying to take a shot at the business. Basically if you don’t have a GEPON solution now, I don’t believe how TXCC can play. Plus you got PON focused startups that are doing nothing but that.

    3) The volume is in the ONU. On the ONU, the only way to cost reduce the box is to integrate all the gateway functionality into one chip. These are PON MAC, VoIP (absolutely necessary), VDSL and media processing routing/switching s/w all in one chip. All major comm semi players are attacking this space and the winners will have an integrated chip in my opinion.

    4) If you focus on the OLT, there is not as much price pressure but the volumes also drop by a factor of 64 or 256 to 1. Not as many will attack this space.

    5) Now for GEPON vs GPON. The one and only big NA deal is Verizon and if you ar not into one of thir ONU suppliers, then you are out. Probably similar story in Europe.

    I am an investor in TXCC also, but a small fry, not in your league. However, I am not sure they can compete in PON based on the above. So that takes them out of EoS and now PON if you agree with my reasoning. I am sure there is more there than what I see but this is my simple minded analysis.

    Posted by JV | March 2, 2007, 6:32 PM
  15. JV,

    1.) I don’t believe at all that TXCC is out at OC-48. After all, the Etherphast-48 is playing a major role in what is, as far as I know, the largest EoS deployment. The chip has also won key slots with Alcatel for the 1692/1696/1626 platforms which are, as you know, significant platforms in the industry. The company has noted the potential for “several million” in revenue per quarter as these ramp to full production. Andrew might be able to comment further on this chip.

    2.) As for GPON, it isn’t really part of the investment story yet. I do not get the impression that timing will be an issue here, in the sense that if they get the product out on schedule, there will be plenty of slots to vie for. I’ve heard the solution is a goood one, but we’ll see.

    This thread started out as a general thread on the comm IC industry. Who do you think is best positioned from both a product and investment standpoint?

    Posted by Herb Chen | March 3, 2007, 10:50 AM
  16. thanks for the incredibly fertile discussion ;)
    reading your posts i realize there is so much to learn. can’t wait to hear JV’s answer regarding best positioned companies. Herb, you sound like an extremely savvy investor as well. besides TXCC, which other companies do u like?


    Posted by ohad | March 3, 2007, 2:45 PM
  17. Hi,
    It looks like a post of mine did not make it. Here are some more thoughts. However, I have made my share of mistakes, so I don’t think I am a savvy investor though.

    The comm IC companies are


    Of these clearly, in my opinion, BRCM and MRVL are the best. However, they are value stocks and now may not be growth plays. Agere has thrown in the towel and is going to LSI.

    EXAR is for all practical purposes being valued on cash. They just sat on cash for a few years and did nothing. They have no business segments with any potential. They don’t deserve to be a public entity nor a company. They could have morphed into a good analog company but blew it.

    I would not touch VTSS equity given their current situation.

    1) The problem with companies like PMCS and AMCC are that they are still dependent on legacy telecom / WAN for a big portion of their revenue. They continue to serve this market which has shrunk and where there is an over supply of suppliers and products. One can also argue that these “products” are really “ASICs” in the guise of being products. These products are complex products that take a significant amount of R&D to serve small markets albeit with high margin. Such products used to be custom ASICs which seemed like they could be standard products in the late 90s. The reality is that they are not.

    2) The only place with any semblance of standard products and volume is in the Access area such as PON and the emerging WiMaX. But where there is volume, there is competition, and these companies are not geared to crunch costs on a die as traditional semiconductor companies are capable of doing.

    3) The reality is that these companies should be viewed as “chip companies” not “comm chip companies” in which case, they are woefully non-competitive.

    4) Each company may have some niche area of competitive advantage or IP, but may be too much of a niche to be significant.

    5) each have made attempts at diversification. But continue to spend in legacy.

    Having said all this, I think PMCS is generally considered the best of the lot, if you exclude BRCM and MRVL.

    But here is the dilemma, they are valued beyond perfection.
    Sorry I could not help and come up with a silver bullet.

    Posted by JV | March 5, 2007, 10:56 PM
  18. I guess that situation characterizes the vast majority of markets today.
    an 80/20 market where the dominant players aren’t growing fast enough while non of the small/pure-play players seem to be able to overthrow the king, or perhaps build an entirely new kingdom.

    Seems like a small venodr can take the lead only if it foresees fundamental market changes and get there way before the big guys realize what is going on. A good example for that might be SNDK or AAPL.
    Are there comm chip niches that might emerge as main-stream segments with high entry barriers, but for now, the giants don’t pay attention to?
    What about the NPU segment? Looks like ezchip’s efforts all these years when nobody was paying attention, has positioned them to capture a huge chunk of that market

    Posted by ohad | March 6, 2007, 1:27 AM
  19. I am sorry for the posts getting swallowed. This site gets 500 spam comments a day and the spam filter (Akismet for those who care) does an outstading job. There are always false positives and I can dig them out if notified.

    Posting many links in a comment is a sure fire way to get it flagged as spam.

    I need to change the commenting system so people become aware of this.

    Posted by Andrew Schmitt | March 6, 2007, 10:24 AM
  20. I am not thrilled about the NPU segment. It is a small market that is not growing exponentially. One reason is that embedded CPUs are being given instruction extensions and hardware add-ons to do the same job, specifically in the volume apps at the access and client points. These are Mips, PowerPC and ARM based processors. Once you go high end to 10G or so the market shrinks to to very small and the cost to make a chip escalates = small market high R&D. Same siutation for switch fabrics, which someone like ezchip does and does well. Will these blossom into large markets with standard product offerings ? this is questionable.

    TXCC just had a rif announcement. This does not seem to bode well. any comments ?

    Posted by JV | March 6, 2007, 5:27 PM
  21. As I said, my technical understanding in this matter is virtually non-existant. However, I find it very likely that the trend we’re seeing of moving the intelligence and decision making from the end-user boxes to the core network should cause a huge surge in the amount of trafiic that has to be processed. So in the not so dsitant future, CESRs should be capable to handle so much data that CSCO and its peers would have put extremely strong processors in every switch and router.
    Are you actually saying that the market 10G and 100G won’t be that big or that there are simply other alternatives to NPUs in 10G/100G boxes?


    Posted by b999200 | March 7, 2007, 3:47 AM
  22. the recent remark by b999200 is mine. I used that name for my first posts here but decided to change it since most people here post with their real names.

    Posted by ohad | March 7, 2007, 6:02 AM
  23. The processing power is needed, but I am not sure you need “network processors”. Freescale, PMCS, AMCC, IDT, BRCM all make processors that can do the job and can be adapted with any extensions bot h/w and s/w

    I question whether there is an NPU market. There are processors that are being adapted to tasks such as L2/L3 header peelbacks and look-ups and forwarding. processors to assist in encryption / decryption. processors with DSP extensions for VoIP and …..

    Posted by JV | March 7, 2007, 12:56 PM
  24. As an adjunct ot he above post the reason why the mainstream players are not doing the high end today is not because they are not capable of addressing that segment. It is because that segment is too small. 5 years from now they will. There is no real new “stuff” needed. Even if you came up with a new elegant way of inspecting that traffic, you still don’t have a big market. There is such a ting as being ahead of the market. From that standpoint, companies such as TXCC are actually doing the right thing by focusing on the access end of the wan. Whether they will be successful remains to be seen.
    Thoughts ??

    Posted by JV | March 7, 2007, 1:00 PM
  25. I do not like NPUs.
    I do not like them in a box.
    I do not like them with a fox.
    I do not like them in a house.
    I do not like them with a mouse.
    I do not like them here or there.
    I do not like them anywhere.
    I do not like NPUs.
    I do not like them, Sam-I-am.

    Posted by Andrew Schmitt | March 7, 2007, 1:26 PM
  26. Andrew – you are truly gifted.
    This is pure poetry :)
    And I thought you were a fighter but the pen is mightier than the fist

    Posted by JV | March 7, 2007, 1:39 PM
  27. More reasons

    – These “NPUs” in some cases have a “custom” attribute to them. Some real, some imagined. So some of them tend to also roll their own. Cisco and Juniper have large ASIC groups that do this. Anyone can license a multicore embedded CPU and add their own h/w and s/w extensions. In particular, tensilica allows you to add instructions that get mapped into h/w, making it very efficient. I believe that Mips and ARM also allow you to do that now. Although tensilica was almost push-button.

    More comments ??

    Posted by JV | March 7, 2007, 2:11 PM
  28. Thanks for the replies. So how can you explain the fact CESR vendors are shipping or developing products powered by NPUs?
    JNPR already launched several ethernet boxes that have NPUs and , ZTE and Huawei are integrating Xelerated’s NPUs for metro ethernet and most importantly, MRVL and ezchip are co-developing NPU solutions to be shipped to CSCO next year. In other words, CSCO literally told them they are interested in NPUs…

    Posted by ohad | March 7, 2007, 5:13 PM
  29. btw – Cisco does not care if the supplier makes money or not. Even if they plan to use only 5, they instruct suppliers to keep going. Xelerated is struggling to go anywhere. Do an ROI calculation on the $$ invested and the return to date (after about 6 or 7 years). Same with EZchip. This goes back to Herb’s comment about VC investments. If Cisco and Juniper can get a foolish investor to fund their R&D, they will take it and they have been taking it for some time now. Especially in the “NPU” area.

    Posted by JV | March 7, 2007, 5:59 PM
  30. I must admit I never saw that possibility as a viable one. I am under the impression that in the coming years we should see a gradual migration in the CESR business towards NPUs, but nothing is guaranteed of course.
    As to R&D investment going down the drain, it wouldn’t be fair to judge the NPU companies at the present, only 1-2 years before the expected(!!!) inflection point. If you look at ezchip, for example, they have been burning on average 10$M per year in the last 3 years- not bad for a company in their situation. The R&D money that has been poured into it by LNOP and others doesn’t seem to be too high, I even think that overall the number is less than 100$ million.
    Compare it to companies like TSXX , VTSS , BKHM which are part of active viable markets, unlike the NPU market. Those companies may be much more diverse in terms of product offering but as stated earlier, none of the players in their field has a clear advantage over the other competitors.

    In summary, I wouldn’t bet the farm on NPU companies like LNOP, but I think the risk/reward ratio is pretty good. Though I am not a subscriber, I think George Gilder disocvered a true winner here. Time will tell

    Posted by ohad | March 8, 2007, 8:19 AM
  31. One thing I do know:

    If NPUs are successful, they will be based on open instruction set architectures. Most NPUs today are not. This is why the latest crop of RISC based multi-core processor startups is interesting.

    Posted by Andrew Schmitt | March 8, 2007, 8:24 AM
  32. I agree with Andrew. The most interesting development is the crop of RISC based multi-core processors. Intel demo’d an 80-core as an example of what is to come, but mainstream will be 4-16 cores imho.

    They can be used as independent processors doing a dedicated task (TCP header /CRC …) or a particular job type can be split out among processors.

    And the challenge is going to be memory and / or bus and / or interprocessor connect bandwidth. Gettting in and out of contect memory for example.

    You can get these from MIPs, ARM, PowerPC and so on. The Cavium high end is really a 16 core Mips engine with a crossbar and some h/w assist from what I can gather.

    Then with an open instruction set and a complier, the action is really in s/w.

    The consequence is that therefore there is no such thing as an NPU. These are RISC CPUs.

    Posted by JV | March 9, 2007, 8:30 PM
  33. Rough rule of thumb – one 1GHz RISC CPU can offload a 1GE TCP/IP stream, without any specialized h/w.

    Imagine having 4 or 8 or 16

    Posted by JV | March 9, 2007, 8:35 PM
  34. Exactly… the action is really in the software. Well put.

    If you look back, the hard problems always move to software, if in exchange a common, cheap hardware platform becomes available. It happened with computing, it’s happening with storage, and it will happen with networking. Fact.

    Posted by Andrew Schmitt | March 10, 2007, 8:48 AM
  35. Amen !

    I suspect that a lot of specialized processing h/w can be eliminated if these things were easier to program. I know that users are frustrated on this particular aspect.

    Posted by JV | March 12, 2007, 7:12 PM