x
Comms chips

Tau Touts Cheap Switch Fabric

Startup Tau Networks Inc. is hoping to make a splash at the Network Processors Conference West tomorrow, when it unveils its T64 packet switch fabric (see Tau Delivers Switch Fabric).

Aimed at enterprise and metro networks, the T64 switch fabric chipset scales linearly from 2.5 to 640 Gbit/s of full-duplex bandwidth, giving it a maximum capacity of 64 10-Gbit/s ports. It comprises two chips: the TL2410 line interface on the line card, and the TX6410 crossbar/scheduler on the switch card. No surprises there.

What may come as a surprise to some is Tau's marketing of the chip. It's making a huge song and dance about the price.

According to Tau, the main reason systems vendors continue to develop their own switch fabric solutions rather than buying from merchant silicon vendors is simply because off-the-shelf parts are too expensive. Performance may have been an issue in the past, but that certainly isn't the case any more. Price is now the main concern.

That's where Tau believes it can make an impact. "Charging $500 to $600 a chip is a non-starter," says the startup's VP of marketing Dave Gampell. "[Our design] allows us to set the price where it needs to be." Tau is projecting a price of $250 to $350 per 10-Gbit/s port, depending on the system configuration.

One reason switch chips are so expensive, he contends, is because they are physically very large. Large chips mean fewer chips per wafer, which doesn't lead to good economies of scale. "The wisdom in the semiconductor industry is," says Gampell, "You could sell 1,000 chips at $1,000 apiece for a total of $1 million, but you'd be better off selling 10,000 chips at $100." Even though the cheaper chips bring in the same total amount of revenue, the margin on manufacturing them is higher.

To this end, Tau claims to have shrunk its chips and their packaging. For starters, it's optimised the layout of the traces inside the chip by custom designing the layout where necessary. Place-and-route tools, which automate this process, don't always come up with the most compact solution, Gampell contends. Tau also claims to have shrunk the BGA (ball grid array) packages by using fewer solder bump connections, but without sacrificing any features.

Analysts don't buy the story completely. "I'm surprised they promoted cost so significantly," says Jeremy Donovan, chief analyst at Gartner/Dataquest. "In my view, there are only three ways to cut prices: large volumes, efficient design, and low margins."

Tau certainly isn't selling large volumes at present, and Donovan doesn't believe that anyone can make the design so efficient it will cut price by 50 percent. "From where I sit, the only way Tau can achieve very low prices is with low margins," he says.

But, he adds, there are so many switch fabric vendors around today, that price may be one of the only ways left to compete successfully.

Putting price on one side, however, analysts do seem to be impressed with the features of Tau's chip. "It's not lagging behind in anything: power, price, or performance," says Donovan.

He particularly likes the number of classes of flows it can support. More flows translates as an improved ability to regulate and monitor quality of service to customers.

Tau claims that the T64 can monitor flows down to the OC3 (155 Mbit/s) level, and offers 16 levels of service per flow. Doing the math, that means each TX6410 crossbar/scheduler on the switch card can support 65,536 flows. For comparison, Agere Systems (NYSE: AGR/A) says its PI-40 chipset can support flows down to the OC12 (622 Mbit/s) level, meaning it supports only one quarter the total number of flows or less (see Agere Announces Products, Contracts).

Jag Bolaria, senior analyst at The Linley Group concurs with Donovan's view, adding that another strength of Tau's switch fabric is the scheduling algorithms, which eliminate the need for a lot of external memory -- another big cost saver for the customer.

"You might imagine that with all these flows, we need a huge amount of memory," says Tau's Gampell. Rather than holding one queue per flow, which is the usual way, Tau has a developed patent-pending queue management scheme that allows it to "juggle" queues without using much memory.

Furthermore, he says, if a flow gets blocked, the crossbar doesn't need memory to store the buildup of data. "Our logic was that the network processor also needs loads of memory to function, so why don't we use that." If one flow is misbehaving, then the crossbar simply sends a message back to the network processor to say "don't send us any more data."

Working against Tau is the fact that other switch fabrics are more established. Agere's PI-40, for instance, is in volume production right now, the company claims. Tau's chips, on the other hand, have only just come back from the foundry, and are still being tested.

Founded in October 2000, Tau Networks has been funded entirely by VenGlobal Capital. Gampell didn't disclose the total amount of funding, but did note that it was in the single-digit millions. "In my 15 years in the semiconductor industry I must say I've never seen so much done with so little."

— Pauline Rigby, Senior Editor, Light Reading
www.lightreading.com
sigint 12/4/2012 | 9:29:26 PM
re: Tau Touts Cheap Switch Fabric Analysts don't buy the story completely. "I'm surprised they promoted cost so significantly," says Jeremy Donovan, chief analyst at Gartner/Dataquest. "In my view, there are only three ways to cut prices: large volumes, efficient design, and low margins."
__________________________________________________

Analysts are incorrigible.

Gampell is accurate in his contention that large die sizes push up chip price.

Large dies mean fewer dies to a wafer. Manufacturing problems occur spordically at random location of a wafer, rendering the die at that location useless. The larger the die, more the wafer area ruined due to even a single localised problem.

Semiconductor yields can fall dramatically as die sizes go up.

Why else would Intel sell the Pentium as two dies (CPU + L2 Cache) in a single package?


All of this, of course, wouldn't make any sense to an Analyst, who says:
_________________________________________________
"From where I sit, the only way Tau can achieve very low prices is with low margins," he says.
_________________________________________________

That's hardly breaking news. Which vendor hasn't been forced to cut margins?

Disproportionate die size bloat is indeed a consequence of inefficient design: legacy of an age when 'time to market' constraints didn't allow human creativity to flourish, when place-and- route tools displaced the human genius.
lightbulb0 12/4/2012 | 9:28:17 PM
re: Tau Touts Cheap Switch Fabric is there single innovation?
fabric_man 12/4/2012 | 9:28:07 PM
re: Tau Touts Cheap Switch Fabric > is there single innovation?

I guess it depends on what you mean by "innovation". I downloaded their product lit from the website. To me, it looks like two major innovations:

1. Very high number of flows in the fabric. More flows (assuming no more cost) helps with traffic management.

2. Very small packages esp. on the line card chip - lowers cost and power. This is helpful for larger systems where line card savings get multiplied by 16x or 32x.

They have various other features too, which admittedly some other vendors also have, although I don't think there is another vendor that can deliver the entire feature set. As the article said, Tau doesn't seem to be lagging in price, performance or power. An innovation in only one of these areas won't help if the other areas are neglected - this is true of specific performance parameters as well. Focusing too much on one "innovation" can kill you in other ways.

Looks like they've done a pretty good job of covering all the important areas.

-- fabric_man
HOME
Sign In
SEARCH
CLOSE
MORE
CLOSE