Friday, March 30, 2007

CEE-ing into Ethernet silicon


Thanks to Patricia Thaler, chair of the IEEE 802.1au group on congestion management, for helping me peer a little deeper into the silicon horizon for Convergence Enhanced Ethernet. The new twist on Ethernet for the data center just emerged into the public this week. Based on discussion with Thaler it seems reasonable that a CEE spec might not get fleshed out until 2008 and chips might not hit until 2009.

Thaler hopes the .1au group picks one of four proposals for congestion management at a July meeting and develops a fairly solid spec by November. "It’s a reasonably complex problem," Thaler said, with a five-person simulation group meeting weekly to iron out issues.

The .1au effort could pave the way for work on specs for priority scheduling and priority-based pause by 2008, she estimated. Most Ethernet silicon already has priority scheduling in the form of weighted round robin schemes, but new hardware support will be needed for congestion management and priority-based pause, she added.

Separately, Thaler expressed concern that IBM seems to be breaking ranks with other iSCSI advocates with its plan to ask the T11 group on Wednesday to start work on Fibre Channel over CEE. Personally, my take is IBM chief engineer Renato Recio is trying to play both sides against the middle to get Fibre Channel and iSCSI going over CEE so in the end its storage on Ethernet no matter who wins.

"I don’t think its time to give up on iSCSI, it's promising," said Thaler, a senior principal engineer at Broadcom. "I don’t think you get things done by changing when things are going slow to something else. In Broadcom we are strong supporters of iSCSI," Thaler told me.

Wednesday, March 28, 2007

CEE: Making waves in the data center


It was only three days ago I said the data center, like the digital home, will be a polyglot environment with many networks co-existing. This afternoon, Renato Recio, a chief engineer in IBM's server networking division, told me about a new effort to bring many of the capabilities of Fibre Channel and Infiniband on to Ethernet. It's called Convergence Enhanced Ethernet (CEE, pronounced "sea").

Renato said he has been working on this for about a year with colleagues from Broadcom, Cisco, other IBMers and Intel among others. "It's all the same people who brought you Ethernet," Renato said. The proponents understand where past efforts such as Infiniband and iSCSI have fallen short, so they at least have a shot at not repeating the same mistakes. That doesn't mean they won't make new and interesting blunders all their own.

My faithful blog readers can have a sneak peak at CEE if you go here. For more info, see the short story I just posted to the EE Times site. There should be more in EE Times print on Monday. If you have any comment on CEE, please post here or send your thoughts to me at rbmerrit@cmp.com

Get off the bus

It's time for today's on-chip bus to give way to the distributed mesh. That's the word from Anant Agarwal, professor at the Massachusetts Institute of Technology. You can read about it in an EE Times report by my colleague Richard Goering on Agarwal's keynote at the Multicore Expo yesterday.

The gist of Agarwal's point is that distributed meshes will replace busses and rings. Meshes help scale bandwidth as new cores are added to a chip, they can be 80-90 percent more power efficient than busses for 16 cores, and they can be simple to layout, he said. Agarwal was in part touting his Raw architecture, a tiled multicore approach using no centralized resources.

So for the Intel's and AMD's of the world still cranking out front-side bus, HyperTransport and CSI designs, now here this: "The bus-based multicore system will fade in the next year or two," said Agarwal.

I would easily add a couple more years to that prediction for the cost-sensitive PC laggards likely to ride this bus all the way back to the station.

Tuesday, March 27, 2007

The case for Firewire


You don't hear much about Firewire these days, especially as a preferred home network. So this tutorial by two Texas Instruments engineers on why the High-Definition Audio-Video Network Alliance (HANA) group chose 1394 rather than Ethernet as the initial basis of its home networking solution caught my eye. Not surprisingly, the crux of the issue is getting better quality of service.

Everything I have heard suggests the digital home will be a polyglot environment with plenty of different networks backed by various PC, consumer and service provider interests. Personally, I doubt 1394 will have a very big role to play, but it will have a role.

Monday, March 26, 2007

IBM optics hit 160 Gbits/s

Optical chip-to-chip links on a printed circuit board are not yet ready for prime time, but IBM researchers are already cranking up the speed limits. IBM's T. J. Watson Research Center will present a paper on March 29 at the 2007 Optical Fiber Conference in Anaheim of a "160-Gb/s, 16-Channel Full-Duplex, Single-Chip CMOS Optical Transceiver,” the company announced today. It claims the prototype device is eight times faster than today's transceivers although it may not be ready for commercial use for three to five years.

Few details of the paper are available yet. But IBM did say the chip set includes indium phosphide and gallium arsenide optical components as well as a CMOS transceiver in a 3.25 by 5.25 millimeter package. A Reuters report now on EE Times said the device measures 17 square millimeters and dissipates 2.5W.

This chipset is designed to enable low cost optics by attaching to an optical printed circuit board employing densely spaced polymer waveguide channels, IBM said. DARPA helped fund the research as part of its Chip to Chip Optical Interconnect (C2OI) program.

Sunday, March 25, 2007

Infiniband stokes the fire

Mellanox throws another log on the Infiniband fire today, announcing it is sampling some of the ConnectX products it announced back in November. Not much new here other than the marketing department's efforts to keep the flames well stoked.


As you may recall, ConnectX is a family of ASICs and adapter cards that can mix and match Infiniband at 10, 20 and 40 Gbits/s with Ethernet at 1 and 10Gbits/s. And as you may recall my take on the data center has repeatedly been that it will be a mixed environment of mainly Ethernet with a goodly share of Fibre Channel and Infiniband. I expect to see that mix in box-to-box interconnects as well as links inside the server blade.

Today's news from Mellanox doesn't change that picture. It just fills in some details of adapter card pricing--$165 to $479 in high volumes--and availability--sampling now with production ranging from next month to next year depending on the configuration of cards.

Thursday, March 22, 2007

Parallel tracks: HDMI and DisplayPort


I finally had the chance to catch up with one of my long time good contacts, Phil Eisler, now vice president of the chip set division at AMD. I wanted to know how Phil sees the HDMI vs. DisplayPort battle playing out.

"I think the two will co-exist. I see DisplayPort as a replacement for DVI and HDMI as a replacement for TV-out," Phil said.

HDMI is already well entrenched in TVs and other consumer gear, he explained, while AMD and other companies won't be in production with their first DisplayPort devices until late this year. The transition to HDMI that's happening in TVs now will be echoed by a similar transition from DVI and VGA to DisplayPort in PCs in 2009, he added.

There's not much of a winner-takes-all scenario here.

"PC companies see DisplayPort as a bridge to CE systems, but I don’t know if that will happen. I haven’t heard of any CE companies picking up DisplayPort, and that will have to happen for this to be true," he told me.

"The history of standards going in that [PC to CE] direction has not been strong," Phil said.

Sounds like a solid, objective analysis to me. Thanks Phil!

Leaping to 100Gbits

The folks over at HPC Wire have posted an interview with Grid-X, a startup working on a 10Gbit Ethernet offload engine that they claim will be able to handle 100Gbit E.

This is what you can safely call an early-stage development given the IEEE group exploring high-speed Ethernet has not even officially set a target speed yet. What's more there's quite a debate raging about whether TCP offload even makes sense anymore given what Sun is doing in its Niagara2 processor and Intel is planning with I/O accleration in its CPUs.

Nevertheless, the startup is worth watching, especially for the HPC crowd thinking about how to link their high-end systems directly into a Globus-based Grid network.

Wednesday, March 21, 2007

Vinod keeps the faith--and plugs ultrawideband

Veteran venture capitalist Vinod Khosla gave a big plug for wireless USB in his keynote at the VON conference today. The ultrawideband technology will forge Gbit/second I/O links between tomorrow's cellular handsets and MP3 players, cameraphones and more, said Khosla in a talk entitled, "The device that used to be your phone."

On one level it was just a shameless commercial for another one of his startups, Artimi, working on low power wireless chips. Even the uber-weathy have no shame about using their fame to provide free commercials. How do you think they got that way?

But on another level, I think Khosla is just all fired up about the bright future of technology and he gets others juiced, too. Perhaps the elder statesman of high tech investing shares the same affliction as his old buddy at Sun Microsystems, Andy Bechtolsheim. Andy, I am told, sometimes has trouble sleeping at night because he is so excited, thinking about where high tech is headed. Having interviewed him a few times, I can believe it.

Quoth Khosla: "Network growth has never stopped. It didn’t even stop in 2001 for the dotcom bust, so don’t count on it stopping anytime soon."

And they all said Amen.

Sunday, March 18, 2007

Express stalled, KR A-OK in backplanes

Monday's top interconnect news is that IDT is pushing forward its work on the defunct Advanced Switching Interconnect to deliver a chip set aimed at putting PCI Express in the backplane of computer and comms systems. The analysis is it doesn't stand much of a chance. Details on both the news and analysis should be here by mid-day Monday.

In short, top technologists I talked to at Alacatel-Lucent, HP and Continuous Computing said Ethernet is and will be the big winner in the backplane for the long haul. RapidIO has a small supporting role to play, and Infiniband may gain a slice of the business, at least in the medium term.

The RapidIO folks have posted a detailed 50-page white paper about why they have a technical edge over Ethernet. That will give them some traction for systems that are willing to pay more for extra performance and features.

The big take away: Watch out for the 10Gbit serial Ethernet standard called KR. It could be huge some day.

Friday, March 16, 2007

Digital media on trial


Next week I will be keeping an eye on DVD CCA vs. Kaleidescape case. I previewed the case in an EE Times story today. Pray for me as I wade through the legal verbiage and protocol to see what emerges of importance about policy in the era of digital media. There may even be some tidbits about copy protection, security and interconnects once they delve into the details of the Kaleidescape systems.

As always, if you have insights or opinions about the case or its principals, fire away here or at rbmerrit@cmp.com


On Monday, look for news and analysis about PCI Express in the backplane (it don't look good folks) from Hewlett-Packard, AlcaLu and others. There will also be word why the 10G Ethernet KR spec is super-hot in today's multicore era, and a reminder that Mellanox is on deck to ship its QDR 40Gbit Infiniband products sometime this year.

Here's hoping you make lots of great connections this weekend!

Thursday, March 15, 2007

USB trumps ExpressCard

Despite—or maybe because of—the news about ExpressCard coming from Cebit and Verizon Wireless today, I have to say the ExpressCard seems to be playing a losing hand in innovation compared to USB and good old system integration.

Novatel Wireless announced today its V740 card is available through Verizon. The card offers 600 Kbits/s to 1.4 Mbit/s downstream and 500-800 Kbits upstream for a cost of $179 if you buy a two-year contract with Verizon, or $299 with a one-year contract. If the relatively slow speeds didn't discourage you, the high price and carrier controls ought to.

In any case, notebook makers are integrating cellular as a top priority, so these cards may soon go the way of their Wi-Fi cousins. But it gets worse.


At Cebit this week the PCMCIA folks are demonstrating some of the 127 products now wearing the ExpressCard logo--including computers and readers. Frankly, outside of digital TV tuners, most are cards are for functions a self-respecting notebook should have integrated—e.g. Wi-Fi, Gbit Ethernet and 1394.

USB has already filled any need for flash-card readers, USB expansion hubs and more. That will only increase when the USB folks start talking in a year or so about a USB 3.0 to leverage the 5GHz Express inside the box. The work is already well under way inside Intel.

Nothing ever dies in the PC, it just slowly fades away. ExpressCard won't fold its hand anytime soon, but its pile of chips is visibly shrinking.

Wednesday, March 14, 2007

I know where they buried ASI

OK, no big expose, just an interesting note from a call with Chuck Byers at Alca-Lu today. Turns out the intellectual and financial assets of the Advanced Switching Interconnect group were taken over by the PICMG, the PCI Industrial Computer Manufacturers Group.

Since PICMG owns a number of backplane standards including PICMG 3.4, for PCI Express in a backplane, the group needed to refer in its spec to some of the work on the ASI web site.


"Whatever cash ASI had on hand will just about cover the hassle of keeping the material online indefinitely," said Byers. "The group also keeps the CD-ROM with all the ASI specs, but there's no active development planned for the technology," he added.



I had been waiting for some sort of official word on ASI since I heard it was on life support last August, then saw its Web site disappear in December.

There is some interest in reviving some of the ASI methods to get Express in the backplane, and a fair amount of resentment over the failed promises of ASI…but more on that come Monday. Stay tuned.

Monday, March 12, 2007

Will Moto's set-tops go wireless?


I can hardly believe it. Set-top box makers and TV service providers have told me time and again there is no way they will trust premium broadcast TV delivery to wireless anything. They don't want the support calls when glitches interrupt the big game or the HBO special--let alone security hassles from stolen premium content in apartment buildings. I've heard it from AT&T, Comcast and Cisco's Scientific Atlanta division.

But here comes startup Amimon with its variation of 802.11n and Motorola Ventures as a backer in its latest capital round. The quote in the press release is the real eyebrow raiser.

“Our investment reflects our belief that Amimon's solution is well positioned to offer a high quality wireless uncompressed HDTV link between video sources such as digital set-top and high-definition TVs,” said a Moto Ventures spokesman.

For now, I am chalking this up to a lot of creative spin. Moto makes no promise of shipping wireless set-tops of any variety in the press release, so I'll believe it when I see it.

Need a backplane adjustment?


A handful of switch makers are betting engineers who design server blades, storage arrays and ATCA-based comms systems will shift their backplanes from Gbit Ethernet to PCI Express or Infiniband over the next year or two.

If you are one of those systems engineers, I'd love to hear you chime in about whether you are planning such a shift, why and what you see as the main trade offs. You can post a comment here or send me an email at rbmerrit@cmp.com. Do it before end of day Wednesday if you want your thoughts considered for inclusion in the round up I am doing for next-week's print issue of EE Times.

SI meeting in PA

Here's a quick public service announcement for those of you who don't monitor the SI List:

If you live in the Harrisburg, Penns., stomping grounds of Tyco you may be interested in a day-long Signal Integrity Symposium coming up April 10. Speakers will talk on high speed connectors, EMC radiation and other topics. The symposium is co-sponsored by Ben Franklin Technology Partners, the Susquehanna Section of IEEE, the Innovation Transfer Network, Penn State Harrisburg and its School of Science, Engineering and Technology. The cost is $115 and includes dinner. Tektronics, Agilent and others will be on hand at a vendor reception. For more information, phone 717-948-6538 or e-mail blf3@psu.edu.

Friday, March 09, 2007

Switching on the lights


Can't close the week without mentioning Lightfleet Corp. (Camas, Wash.) that claimed this week it will bring free space optics to high performance computers later this year. There have been plenty of claims in this area but little reality to date.

The startup, founded in May 2003, is not providing details of its technology yet. But it did say its Corowave switch essentially includes a laser transmitter and an array of opto-electronic receivers. Each processor in a server is outfitted with the switch, creating an all-to-all broadcast or multicast network with each CPU filtering out what data it needs. The company has one US patent (No. 7,136,419) issued in November.

How that CPUs maintain alignment and how the filtering is done sound like interesting challenges. The company tipped its plans in hopes of raising interest in licensing the switch it intends to put in its own systems in July.

Perhaps this summer the company will shine more light on exactly what it is up to. You can check out somewhat sketchy reports at HPC Wire and the Wall Street Journal. Meanwhile, I welcome posts from anyone who has seen the technology or has observations about the state of the art in free space optics.

Thursday, March 08, 2007

Serial ATA shrinks to grow


Like many things in electronics, smaller is often better. The Serial ATA International Organization announced version 2.6 of the 3 Gbit/second SATA spec this week, now available for download free to members and at a cost of $25 to non-members. Two new small connectors are the main highlights of the revision.

An "internal slimline" cable and connector will help SATA optical drives slot into smaller form factor systems. An "internal micro" aims to slip 1.8inch hard drives into the possibly emerging category of ultra-mobile PCs.

Length is the figure of merit on the new connectors here. The slimline measures 28.7 x 5.4 x 5 mm and the micro is 35 x 4 x 5 mm compared to the existing standard which was 42.73 x 3 x 5 mm.

The spec also includes an enhancement to native command queuing (NCQ) so a host CPU can quickly park the head of the hard drive when a laptop containing the drive is dropped. OEMs have been putting accelerometers in notebooks for a while to make sure heads are parked to protect media before a falling system hits the floor.

NCQ has been a nice feature to bolster performance of SATA drives, but it gets in the way of this particular scenario. It is "unacceptable to wait" for potentially 30 outstanding NCQ commands to complete before parking the head when a notebook is in freefall, a spokesman noted. No kidding!

The big picture here: Serial ATA, far and away the most successful storage interconnect in years, is making its way into the portable world in a big way. There's plenty of curl in this big wave.

Tuesday, March 06, 2007

Intel jumps in RFID


Well, it's just a baby step, really, by a small startup incubated as part of Intel Capital. I have to admit I did not know before reporting this story on EETimes.com (about a new single-chip UHF reader) that Intel Cap routinely nurtures 6-12 internal Intel startups at any given time. Makes sense. The tiny company gets the benefits of the big company name and resources without the big company taking much risk.

I also didn't realize there is a big technology unknown about which RFID technology is suited for the Holy Grail market of replacing the world of bar codes—LW, UHF, HF or some hybrid. Even the analysts don't know what it will be yet, but they think it coulod arrive circa 2010. What do you think?

Monday, March 05, 2007

The x86 sword fight


The hardware.info site did a good job laying out the basics of the Intel versus AMD battle over CPU interconnects in a tutorial on the Geneseo, CSI and Torrenza programs.

There were just a couple points they neglected to hammer home. This is a battle for mindshare, trying to get the most innovative accelerator cores to commit resources to an Intel or AMD ecosystem. Both parties not only want those chips plugging into their motherboards, they want the companies behind them to become partners, delivering silicon blocks to future multi-core CPUs.

The online tutorial correctly concluded that AMD has a time to market advantage in that it already has its cache coherent HyperTransport bus up and running and just needs to iron out licensing terms. Anyone know if that is that done yet?

However, the article fialed to credit Intel with leveraging the PCI Express juggernaut with Geneseo. Express has an order of magnitude more industry development resources behind it. See my earlier post here. What remains to be seen is just how Geneseo and Intel's future processor bus, CSI, split the pie in CPU interconnects.

Roll it all together and one ting is clear: the future of the x86 is all around interconnects and a growing number of companies plugging onto them.

Sunday, March 04, 2007

Intel drives multiple CPU buses


Two tiny details in an Intel server briefing last week gave a peak at some big moves in the X86 king's plans for attacking archrival Advanced Micro Devices.

What Intel suggested in its briefing is that its upcoming four-way Xeon code named Tigerton will sport four dedicated processor buses, each one linking a CPU to a chip set in a four-socket server. This could go quite a way to matching AMD's use of the cache coherent HyperTransport interconnect embedded on its Opteron CPUs to link four processors in a system.

Of course, Intel's Xeon will still lack an embedded memory controller so the Caneland system its Tigerton processors plug into will still have—at least theoretically--lower performance and higher cost than a four-way Opteron server.

It will be interesting to see whether Intel makes any technical tweaks in its processor bus as part of this move. I also wonder whether Intel will be more open about licensing the bus so other chip or system makers can design their own multiprocessing approaches. Perhaps this is a small step toward Intel's long-rumored, HyperTransport-like interconnect known as CSI.

Stay tuned for more details when Tigerton hits this fall.

 
interconnects