The TOEs kick back

Sun Microsystems stepped on plenty of TOEs (read TCP offload engines) when the company announced its Neptune networking card and ASIC. Now the TOE makers are kicking back.

Sun wants to ship the TCP processing job off to its multi-threaded server CPUs. That's a mistake, said Larry Boucher, chief executive of TOE designer Alacritech in a missive he fired off to me today.

Sun's approach "is an excellent way to insure that they can make inefficient use of a large system," wrote Boucher (emphasis mine). "TCP protocol processing can be done much more efficiently in a special purpose processor," he wrote.

Here's Boucher's reasoning:

A TOE implementation is hardly half the cost of the CPU itself…By replacing the conventional NIC (or Sun’s new load-distributing NIC) with a TOE NIC, you can effectively double the capacity of the computer. Since a typical high-end computer of this sort is frequently, even in the PC domain, over $25,000, this makes TOE a pretty compelling proposition.

Makes sense to me. Over to you Sun.

RIO not grand, but flowing

Thanks to Chris Kissel from In-Stat for sharing with me a copy of his recent report "PCI Express, The Journey to 2.0." It painted a picture of the serial RapidIO (RIO) interconnect as a small, but still significant player in embedded systems.

The report confirms what we all know, Express is the big juggernaut in chip-to-chip links, driven by its use in Intel chip sets. About 225 million of the 244 million systems that used Express in 2006 were PCs or servers. That will rise to about 347 million of a total of 440 million systems using Express by 2010, Kissel predicts.

By contrast, the RapidIO interconnect spawned by PowerPC backers including Freescale Semiconductor was used in just 433,000 total systems last year. In-Stat projects RIO will appear in about 1.2 million total systems by 2010.

For its part, HyperTransport looks like a lesser Express. It rides the coat tails of the AMD Opteron and Athlon CPUs, appearing in about 50 million systems last year and about 74 in 2010, at least two-thirds of them PCs or servers.

The interesting bit for me was to see projections that RIO, while small, will be very strong in defense, medical and WCDMA base stations. Kissel projects the interconnect will be used in 35, 55 and 90 percent respectively of systems in those sectors. The raw numbers of those systems aren't huge, but the strong uptake bodes well for the long life of the technology, apparently thanks to its unique robustness features.

So I'll take back, or at least hedge my earlier comment that RIO will go the way of the Advanced Switching Interconnect. Barring unforeseen dry spells, it may flow on for the foreseeable future as one of the smallest of the interconnect rivers.

Green Grid needs growth

The Green Grid, an ad hoc consortium of eleven companies formally announcing itself Monday, needs members in networking, storage and other disciplines if it is to have the clout it needs to set standards for lowering power consumption in the data center. It also needs to get more transparent about what's on its agenda.


I applaud the group of eleven that formally debuts Feb. 26, but it's way too computer centric. No data center initiative will be successful without strong buy in from the likes of Brocade, Cisco, EMC, Juniper and their ilk.


The good news is the group's leadership includes seasoned pros like Tony Pierce, the former head of the PCI SIG, and Jim Pappas who helped lead more industry initiatives than you could shake a stick at including USB and Infiniband. And they know they also need strong participation from data-center users to do their work.


But what is their work? I found out they will take up the debate of whether data centers ought to switch from AC to DC power to eliminate losses in power conversion only because I heard about this from others and specifically asked Grid members. Otherwise the consortium is staying mum about what particulars they plan to put up for debate.


That's a mistake. There are plenty of issues that need to be worked out and the group ought to throw them out for public debate as one way to enlist more people to grapple with them.


The Green Grid is using the model of USB in which a small group of experts quietly defines 90 percent of a solution, then opens it up for the industry to provide feedback on the last ten percent. The data center power issues are too broad and involve too many kinds of systems and technologies for that approach.



What's needed here is more of an open-source software model where the whole community gets involved in tackling the many pieces of this complex problem. The role of the committee is then to codify the most viable pieces of the solution that emerge.

Pass the patents, please

If you are into Wi-Fi or WiMax get ready to start writing some checks to Mosaid Technologies Inc. in Ottawa. The company just bought Agere's patents in these areas for as much as $70 million or twenty percent of the royalties on an ongoing basis, calculated at roughly $5 million a year.

Excuse me for farting in church, but I thought the patent business was set up to reward and encourage innovators to keep doing their good work. Deals like this—and whole businesses based on them by companies like Patriot Scientific—make me feel like lots of people are getting ripped off along the line.

Millions and millions of dollars a year with a minority, if any, of it going to the actual inventors. Makes me ill.

Carriers cripple Bluetooth

Verizon typically requires handset makers disable Bluetooth's file transfer capabilities on their handsets, according to a petition Skype filed with the FCC yesterday. That's because the carrier wants its users to download camera phone pictures to a paid Verizon Web site rather than directly to their local PCs, the petition said.

See more stories about these restrictions here and here.

The crazy restriction is just of a laundry list from US cellular service providers who exercise an iron grip on their nets. Skype's FCC petition and an opinion paper from Columbia University law professor Tim Wu are bringing these practices to light and calling for more open cellular nets.

If you know of any similar practice, this is a good time to chime in. You can make a posting here, drop me an email at rbmerrit@cmp.com or stay tuned for when the FCC puts the issue up for public comment. I will post a link to that public commente site here when it becomes available.

Interconnects 2.0, the beta

OK, I admit it. Google dragged me kicking and screaming into their 2.0 version of Blogger this week, but I like it.

I finally got my ugly mug shot pushed down lower on the page, an easier way to access my archives and an update of my hot links. And it was easy. The only downside is the point size on the typeface seems to have shrunk.

I'd love to hear any feedback from readers about the changes and any improvements you would like to see in this blog whether it is in style, presentation or content. I don't pretend to be perfect, but I can change.

Stepping on TOEs

Like Intel, Sun Microsystems is taking an approach that uses its microprocessors to handle the growing processing requirements of Ethernet. Sun's Neptune card and chip, officially launched today, eschews the use of TCP offload engines (TOEs) favored in the iWarp approach promoted by the RDMA Consortium. Intel announced some time ago its IOAT approach which acclerates TCP processing thanks to some basic features in its Xeon CPUs and chip sets, also side stepping TOEs.

With Sun and Intel both rolling out their own approaches that rely on unique features in their CPUs, the market for chips using the TOE approach will be significantly reduced. It's clear many companies see the transition to 10G as a time to differentiate themselves. What’s unclear is exactly how much pain the industry will go through before we get this sorted out.

Rapid I/O dives into backplane

Check out the story by Loring Wirbel in the print pages of this week’s EE Times if you want to get an update on Rapid I/O. The upshot is the group appears to be re-focusing on backplanes--as opposed to chip-to-chip interconnects--as a primary application.

That feels like a big shift to me, and I think I understand it. Back when it got started, RIO was essentially designed as the PowerPC native parallel processor bus, like HyperTranport on the AMD Opteron. But these days PCI Express is scooping up all the chip-to-chip traffic.

You might note the new dual-core Power PCs from P.A. Semi use Express, not RIO. It will be interesting to see what interface the PowerPC chips coming from AMCC this summer will use, given they were designed under the leadership of Dan Bouvier who drove much of the RIO and embedded PowerPC work while he was at Motorola Semi.

Will RIO find a home in backplanes? A cottage maybe. But just as Express is taking everybody’s attention in chip-to-chip links, so Ethernet is winning most design wins in backplanes, especially with all the interest in the IEEE backplane Ethernet standard.

My net on this is that RIO could go the way of the Advanced Switching Interconnect in a year or two.

The medical dilemma

My heart goes out to the hearty few of you who click on my blog on the weekend. Hopefully I am not talking to a couple dozen automated Web crawlers!

Whoever you are, here's a little tidbit picked up at ISSCC this week.

Mingcui Zhou (left), a graduate student at the University of California at Santa Cruz, presented a paper on a 2 Mbit/s wireless transceiver for medical systems using a 20 MHz frequency. The 1.7x 2.6 mm device handles most of the signal sampling and processing in the analog domain to save power. It consumes 6.2mW, but has not yet been tested for meeting FCC regulations on that frequency.

Two Mbits may not seem like much for us Wi-Fi notebook mavens, but in the medical world, it's huge. Her advisor Wentai Liu said he needs 3 Mbits/s for the 1,024-electrode artificial retina he wants to design. Fat chance of getting it anytime soon.

Some implantable devices must last for years on the smallest battery possible, while others devices use what energy can be scavenged from RF signals. The implants themselves cannot generate more than one degree Celsius of heat for fear of burning body tissues.

The constrictions mean most systems are lucky to get data rates of 700 Kbits/second between an implant and an external programming device. A University of Southern California biomed professor, Gerald Loeb, speaking at ISSCC said he only gets about 120 Kbits/s on the toothpick-sized muscle stimulators he designs for paralyzed and impaired patients.

So if you want a real engineering challenge that makes a difference in people's lives, this is the place for you.

A quick serving of stew

I've been tied up at ISSCC for a couple days, but there is plenty of interconnect news cooking. Here's a quick sampling to keep you nourished.

For those of you living in the data center, stay tuned next week for fresh details about Sun's Neptune Ethernet design, and there is word from the Green Grid on the horizon. For those of you into biomed, I have a brain dump coming later this week from ISSCC where I grew a few new cells.

Loring Wirbel has a strong update on RapidIO you will hear more about on Monday. Sounds like the group is turning up the volume on use in the backplane, taking on Ethernet and talking up new flow-control and Quality of Service features coming in their version 2.0. Are they recognizing the PCI Express juggernaut will take the lion's share of the chip-to-chip links? What about all those DSP farms in cellular base stations they used to talk about? More TK.

Consumer gurus, check out the latest about HDMI trying to make in-roads into the PC. They are crowing now, but I suspect they may go quiet once DisplayPort surges forward late this year.

BTW, the HDMI folks say I am wrong about the royalties mentioned in the article. The now-defunct UDI would have had no extra royalities, they say, beyond the HDMI charges that just went to administer the licensing group, not provide any signficant benefit to Silicon Image.

Now, I have to dash off to an interview with one of my favorite subjects, Jeff Hawkins, creator of the Palm Pilot and Treo. Lately, Jeff have been living in the heady world of brain research.

Stay tuned!

Swing low, high powered I/Os

That's the new song Rambus wants to teach the interconnect world with its ISSCC paper coming on Wednesday. Their 6Gbit serdes delivers 2.2 mW/Gbit/second based on a voltage swing of less than 200 milliVolts.

Many interconnect standards today specify 1V swings to make it easier to build the transceivers. But that spec is too loose as we enter the era when microprocessors are power bound.

Everybody ready to line up for the new limbo dance?

Inside the Google server

Who is Google Platforms and why do they need deep expertise in board design, backplanes and high-speed signaling?

A job posting on an email list for signal integrity engineers Friday (Feb. 9) piqued my interest. It said Google Platforms, which describes itself as a "group of elite hardware and systems software engineers," has "many openings for senior and junior SI engineers." Applicants should be able to "generate requirements for placement and routing based on simulations," the ad said.

The New York Times had a good report last year on how Google designs its own servers to save cost, though it did not get any details about those designs. I know Google sells two Web search appliances, but these products don't require the kinds of high-speed backplane design skills that Google is seeking in its new ad.

According to the posting, Google is looking for engineers who can use "time and frequency domain tools for simulation of high speed serial links (chip to chip and across backplanes)." Applicants also need to model "physical structures such as transmission lines, vias and connectors."

The jobs require "solid understanding of SerDes design, PLL design, LVDS, SSTL, HSTL, CML, and other high performance I/O technologies." Engineers need to be familiar with tools such as HSPICE, Ansoft HFSS, CST Microwave studio and Spectraquest.

I am trying to get an interview with the Googlers, but they really don't want to talk about their hardware engineering work. So if you know something about Google's server designs or the Google Platforms group, drop a posting here or contact me at rbmerrit@cmp.com

An Rx for Bluetooth and USB

The Continua Health Alliance, an ad hoc group of high tech companies trying to enable networked consumer devices, is gearing up to announce a suite of technical specifications that will include some tweaks for medical devices to at least two mainstream interconnects.

Continua already has members working on a Bluetooth profile for medical devices as part of the Bluetooth Special Interest Group. Other members have been meeting on an ad hoc basis for several weeks to consider protocol changes for USB that may be needed for health care applications. The group expects to present its views to the USB Implementers Forum soon.

Separately there's lots of work afoot driven by Continua members and others to expand the IEEE 1073 spec, now called 11073, that defines a wired and wireless communications transport for medical systems. The standard is a cornerstone of the Integrating the Healthcare Enterprise project which seeks to create links among a broad array of medical systems. Vendors including Draeger, GE, Philips and Siemens will demonstrate at a medical conference in New Orleans in late February real-time multi-vendor medical device interoperability using the standard.

Launched last June, Continua expects to reveal by the end of April a comprehensive suite of interfaces it will recommend for networked consumer devices. The group then will write interoperability guidelines for implementing those interfaces in medical systems. The work is expected to be complete before the end of the year.

Payday for Mellanox

There's something to be said for persistence and hard work. In fact, there are about 102 million things you can say for it today. That's the amount Mellanox Technologies raised in its IPO that saw its stock price climb to $17, according to an EE Times story this morning.

Years ago the whole Infiniband initiative nearly collapsed, taking a handful of startups and a team of architects at Intel with it. But the Mellanox crew "stayed the course," turning out decent products and finding real markets for them. Today Infiniband is accepted as the rising tide in high performance computer clusters.

The Mellanox-ites still have a big selling job to do and lots more products to crank to establish Infiniband in storage. Fibre Channel is well entrenched and 10G Ethernet is finally getting its act together. But these days Mellanox has a little more cash to fuel its ambitions. No doubt some of that will also pay for a well deserved celebration.

Express breaks out of the box

PCI Express, the biggest juggernaut in interconnect technology these days, is officially available as a cabled option for a host of server, consumer and embedded uses. Infiniband, Ethernet and USB proponents don't need to worry—yet.

The new spec is aimed at enabling I/O expansion drawers for servers. OK, that is an Infiniband application. It will also link comms and other embedded systems, where I suppose Ethernet could be used. Other possible uses include a link to external graphics accelerators for those power gamers and a cabled link between notebooks and docking stations. I'm not sure what value either of those last two bring, but they do sound suspiciously like higher bandwidth alternatives to USB for consumer systems.

PCI SIG execs went deliberately slow on this spec to make sure people didn't get confused about the main purpose of Express as a chip-to-chip link. They also didn't want to raise the specter of overlap with links like Infiniband, Ethernet and USB. But now that it's out there people will likely try to push and pull it in all sorts of directions looking for new market niches in typical PC industry fashion.

For today, its just a novelty, running at the 2.5 Gbit/s rates of Express 1.0 at widths of 1, 4, 8 and 16x. Later, we'll see 5 Gbit/s versions and with initiatives like Geneseo launched last fall, who knows all what else. Watch this Express train because it's carrying a lot of freight these days.

Low power age dawns for I/O

Everybody knows that we are in a new age where power consumption issues are numero uno in design, driven mainly by the fact new semiconductor process technologies are not scaling down nearly as much as in the past. That means the Intel-driven go-go years of cranking up the data rates are over and the new mantra is power efficiency.

There's even the global angle. What President Bush calls our "addiction to oil" and what Al Gore now calls the "climate crisis" are driving a new clean tech movement.

So, two news nuggets on this trend line:

First, as promised last Friday, the details about the Lawrence Berkeley Lab's ambitious and multi-pronged power consumption initiative are out. OK, this is probably the most ambitious group of ten researchers I ever ran into, so keep expectations low.

They aspire to set new power standards for large switches and home gateways. They even hope to straighten out problems in the living room so that a Sony TV can shut down Panasonic DVD player when that makes sense. If they get it right, consumers may even see on, off and sleep buttons on their remotes someday. That will require all sorts of infrared and home networking standards to get more power aware in a standard way. I/O meets low power.

I/O meets low power in another big way on Monday when the wraps come of the Rambus paper at ISSCC that I tipped in an earlier blog. Stay tuned.

Meanwhile, if you've got some hot tips about what's coming down at ISSCC I am all ears. Sing to me here or at rbmerrit@cmp.com

Two last dispatches from DesignCon

I learned a lot at last week's conference where I met many of the luminaries in high-speed interconnect. I tried to capture two points in particular in my wrap up of the conference in EE Times this week.


Engineers are calling for a standard approach to signal integrity to replace the various techniques written into separate, often redundant standards such as Infiniband, PCI Express, Serial ATA and so on. The area is ripe for an ad hoc industry "coalition of the willing," as it were, that could quickly hammer something out and pass it on to a standards group.

Second, the move by Cadence to establish an applications programming interface as a way to model high-speed transceivers is gaining momentum. Mentor and SiSoft are now on board and Cadence was trying to recruit test and chip makers late last week.

The road to a standard on both fronts could take until beyond next year's DesignCon, but the momentum is real.

I'd love to hear what you think about these efforts and what your key take-aways were from DesignCon. Post 'em here or drop me a line at rbmerrit@cmp.com

Pinching power everywhere

The researchers over at Lawrence Berkeley National Laboratory are feeling their Wheaties. Turns out the Energy Efficient Ethernet project they helped kick off at the IEEE is just the tip of the iceburg. These folks ideas for new energy-saving specs for everything from network gear to TVs and set-top boxes.

They are also talking about a concept for letting PCs go into a deep sleep state while network gear maintains their presence on the net. And, they even want to get consumers in the act, suggesting tomorrow's remotes might add a sleep button to the ubiquitous on/off control.

What are these ideas and can they pull them off? Check back Monday afternoon for more details. Meanwhile to get a better sense of their thinking, check out spome of the papers they have published on the topic.