Tuesday, April 22, 2008

Still more on 10G

More 10G news flowed this week in the after glow of the Storage Networking World conference where FCoE had a big coming out party.

I was interested to see 10GBase-KR get a boost with Fujitsu's new 10GE switch chip, but analysts were quick to point out there are plenty of signal integrity problems ahead getting today's 3 Gbit backplanes to 10G.

Good also to see Aquantia joining SolarFlare with a 5.5W 10GBase-T transceiver. It will be interesting to see how long Broadcom and Marvell stand on the sidelines waiting for a market to develop here.

On the server side, Mellanox is promising new firmware so it can carry a wealth of protocols on top of its ConnectX cards that handle all sort of net and storage functions on Infiniband and 10GE. It marked the first time I had heard Mellanox use the increasingly widespread term from Cisco—Data Center Ethernet—referring to the lossless version of the net still going through the standards process.

5 comments:

RapidIO Executive Director said...

Moving to 10G per pair, may be needed in a small set of focused applications, BUT today and into the next several years the migration steps of 1G, 2.5G, 3.125G in X1 and X4 links covers the majority of applications, and with 5G and 6.25G also already defined and AFFORDABLE, no rush to 10G/pair. If the underlying protocol is efficiently using the bandwidth, w QoS, systems with backplanes with 10Bbaud of data throughput -like X4, 3.125G RapidIO will continue to dominate embedded applications that are Power wise and bandwidth needy.
At 100mW per channel (3.125G) VS 1W - 2W for 10G KR pair, the 10Gbaud data on 4 pairs at 3.125G looks very good. And since all major International Standards support X1 and X4 in backplanes, connectors and form factors/boards. Save your Watts, and Dollars for more processors, more end points and more users per channel. Match your speed to your need!

Rick Merritt said...

I hear Watt you are saying, this is the era when power concerns are primary!

Anonymous said...

It is a stretch to assert interconnect power savings will enable an additional processor (x86 not one of the low power models such as just purchased by Apple) to be provisioned within an enclosure. Perhaps summing up multiple enclosures one might obtain enough power to provision such a processor but not within a single enclosure.

As for having the blessing of international standards for any technology, that is a nice thing to have but for many, it is no longer relevant. The traditional standards bodies such as the IEEE, IETF, PICMG, ANSI, ..... all move at incredibly slow rates relative to market need and they are fraught with politics that can lead to compromises that diminish the impact of the technology being put forth. This process serves a purpose to weed out bad ideas but the speed to deliver a specification is often mind boggling slow. This is why more and more companies are working outside of standards bodies and then bringing in near complete specifications for approval. While approval is deemed important, it will not stop anyone from execution. A recent example is the Ethernet per priority pause specified in a public document by a large network company. Within the IEEE, pause is seen as a terrible feature of Ethernet and is often turned off to solve interoperability problems. Many see per priority pause is taking a bad idea and making it that much worse even if that isn't the case. The debate has gone on for a good chunk of time within the IEEE and while they debate, the industry is moving to execute making the standard largely a nice touch but largely non-interesting.

The same can be said about technology wars and their marketing proponents. Any technology may be ideal for one set of applications while poor for others. When marketing attempts to force the "ideal" or "best" nomenclature it does a disservice to the technology as well as the industry. Picking single dimensions such as power per lane, which is important, is also a disservice. The technology combined with the industry enablement and adoption need to be examined as a whole and not a set of selective comparisons. The reason Ethernet hasn't taken over everything is because people see the technical and business differences that lead to other technologies being preferred. It isn't that Ethernet could not be used for everything but because it should not be for solid logic. Power is something Ethernet is trying to address while carrying forward a complex legacy and future interoperability set of requirements. Whether it is successful in penetrating more usage models is a nice debate but in today's tactically focused economy, it remains just a debate and companies should focus on solving practical problems in innovative ways rather than trying to create grand unified arguments for or against other technologies.

RapidIO Executive Director said...

While some standard bodies move at a slow rate, I would say it is true of the majority. PCI, PICMG, VITA, VESA, OBSAI, CPRI and RapidIO are all fast moving and effective, members drive these spec's to match the need. Where the need is great, the parties come together and quickly define and approve. As a active player in the standards development and a founder of a few of these organizations, I would say the companies are not moving away from standards bodies but moving to them. The need for the development of a strong and broad Eco-System of common solutions has never been greater, the days of large custom systems design are gone, it takes too long and cost too much. And if your a large 85% + market share company you can drive the standard by defacto, as in your example. But otherwise with the right mix of users (OEM) and suppliers (Semi, SW, Test, etc) standards groups can be fast and effective.

Anonymous said...

I agree that standards are critical and the standards bodies should be proud of what they can accomplish. But the reality is the draft specs are created by one or a handful of companies who then seek its approval by the broader set with as minimal of change as possible.

One other thing to keep in mind is speed is always relative. For some, the standards bodies move fast; others they are at a crawl. What hurts more is when people execute ahead of the standards creating interoperablity problems for those who come a bit later and execute to the final specification. The technology and customers suffer alike. I'd never advocate waiting until the very end to do anything but prefer to see more emphasis on the beta approach where there is time to bake the implementation and validate the specifications are correct before calling it final. This is what some standards bodies require and it works well for the industry and for customers alike.

 
interconnects