Procket Founder Takes On Servers

One thing about those Procket guys -- they know how to think big.

After taking his lumps against (Nasdaq: CSCO) and (Nasdaq: JNPR) in core routers, former Procket Networks CEO Sharad Mehrotra is accepting a job that doesn't sound much easier: CEO of a company targeting the high-end server franchises of Hewlett-Packard Co. (NYSE: HPQ), IBM Corp. (NYSE: IBM), and Sun Microsystems Inc. (Nasdaq: SUNW).

Dubbed Fabric7 -- which has no known relationship to Wave7 or Blake's 7 -- the 80-employee startup today announced its Q80 and Q160 servers, which combine Advanced Micro Devices's (NYSE: AMD) Opteron processors with proprietary switching technology. The latter gives the boxes speed comparable to high-end Unix systems or even mainframes, despite using off-the-shelf chips, Fabric7 officials claim. (See Fabric7 Unveils Servers.)

Fabric7 has raised $32 million from investors including New Enterprise Associates (NEA) -- where Mehrotra was a partner -- Foundation Capital, Goldman Sachs & Co., Sanmina-SCI Corp. (Nasdaq: SANM), Selby Venture Partners, Vanguard Venture Partners, and Yasuda Enterprise Development Co. Ltd.

Mehrotra's last attempt at this sort of thing ended in disappointment. Procket finished its core router, but a brutal telecom market coupled with Procket's high tab -- its venture funding exceeded $300 million -- doomed the company. In 2004, Procket had to settle for selling pieces of itself to Cisco. (See Cisco to Pay $89M for Procket Assets.)

Mehrotra, one of three Procket founders, was long gone by then. He'd been replaced as CEO in 2001 and had left Procket's board by 2002, having joined NEA. It was there that he developed the ideas that became Fabric7, which launched in May 2002 and began to emerge last year after raising its B round. (See Bigwigs out of Procket? , Execs Cast a Wide Net, and Fabric7 Feasts on $17.5M.)

So, what's to keep Fabric7 from falling into oblivion like Procket?

"It's a very consolidated market. The big vendors are vulnerable in their high-end machines, all for different reasons, so I think we're in a good position to pull this off," Mehrotra says.

And Mehrotra isn't shaking off the lessons from the Procket ride. "A lot of what happened to the last project, certainly some of it had to do with the bubble, but a lot of it didn't," Mehrotra says. One difference from Procket is that Fabric7 is taking advantage of off-the-shelf chips -- not just the AMD chips, but network processors and switch fabrics, too. That's giving the company a faster start at a cheaper price.

"Unlike Procket, we didn't have to spend $150 million on a set of chips," Mehrotra says. "This company was a chance to take all the beneficial ideas about networking and synthesize them to this family of machines at a cost structure that's one-tenth of what we did last time."

Having worked at Sun, Mehrotra comes from a computing background. He sees Fabric7 as a chance to add his networking experience to the mix. The way he reasons it, servers shuffle data from one point to another just as routers do -- so why not apply some new switching technology to the concept of clustering servers?

Using a concept similar to a router's switch fabric, Fabric7 can link servers arbitrarily and "build computing structures completely on the fly," Mehrotra says. This means Fabric7's boxes can be linked together to act as one, an idea similar to the multichassis configurations pursued by Procket and other router vendors.

Different servers could become team players within this distributed environment. One server could handle all connections to a storage-area network, for example, while servers on other sites run the applications that use the SAN. And servers would be able to shift tasks instantly. That's a key point, because it creates a more fluid data center where computing power can be reassigned to particular areas in need.

Fabric7 shares one of Procket's biggest problems, though: a stubbornly difficult market. The customers for high-end servers bear strong resemblance to the buyers of terabit routers, and that's not good news for a startup.

"Big iron, no matter how you look at it -- PBXs, servers, switches -- it's a conservative buy," says Rob Enderle, principal analyst of The Enderle Group, noting that the purchase is so expensive that buyers "want to make sure the company is rock solid." That puts startups at a handicap. On top of that, the sales cycle for high-end servers is long -- just as it is for core routers -- meaning revenues are slow in coming.

A startup's best hope is to partner up with one of the big players, Enderle says. Such pairings can lead to an acquisition -- a happy exit. Or, they could just delay the end: "Eventually, they might backfill you with one of their own products, and then you're done," he says.

Naturally, Mehrotra would like Fabric7 to succeed, but he concedes that the odds are against him, just as they were with Procket. He doesn't seem to let that bother him. "Entrepreneurs do what they do mainly because they like to do these projects," he says. "What drives me are difficult markets with enormous payoffs. That's why you see me doing these projects."

— Craig Matsumoto, Senior Editor, Light Reading

<<   <   Page 2 / 2
rexhaus 12/5/2012 | 2:55:16 AM
re: Procket Founder Takes On Servers Their website says they are using AMD DirectConnect up to 8 chips or 16 cores. It mentions a 32 x 32 crossbar which uses up to 4 ports per processor (10Gbps), which could indeed be off-the-shelf.
At larger sizes, they talk about Ethernet, including 10 GbE, although it's a little unclear whether Gigabit Ethernet is also part of the mix as 10GbE isn't that widely deployed at this time.

At any rate, as Fabric7 presumably knows, all of the above hardware is essentially commoditized. Both the development and sales/business development processes are extremely different than the router world. Perhaps the toughest technical problem is not building and interconnecting the hardware, but software to take advantage of it. Most applications will not be able to take advantage of anything beyond a handful of processors in an SMP anyway, so 16 cores and a crossbar may not mean much (esp. as this is available to anyone else e.g. Sun, or even Dell if they want). They claim an Oracle Architect and ex-Sun COO Ed Zander on their board of advisors (Why would he want to do this? Isn't he busy with more urgent items? Must be a hangover from his days in Private Equity.) So presumably they understand the overwhelming importance of application software, which is completely unlike the router market. They have a serious chicken-and-egg conundrum that the developers won't care if they don't have customers, and the customers won't care if they don't have the applications.

At this pricepoint, this is an extremely tough sell. They appear to have zero sustainable advantage, and it is completely unclear to me at least why they would have even a remote chance of success.

bsd_devil 12/5/2012 | 2:55:10 AM
re: Procket Founder Takes On Servers I don't understand what the big deal is! So they have created a pool of servers such that each can be dedicated to a different task or can be used in cluster.

I believe the same can be achieved with ATCA based servers (if they are available), or with IBM's blade-center.

As far as the switching fabric is concerned, both ATCA and BC are expected to allow a variety of switch fabrics: RapidIO, PCI-X, AS, etc.

Funny, that company's can still get funding for doing mundane things. Just goes to prove how dumb and uninformed VC-s are.

ron202 12/5/2012 | 2:55:05 AM
re: Procket Founder Takes On Servers Funny, that company's can still get funding for doing mundane things. Just goes to prove how dumb and uninformed VC-s are.

we should give them credit that maybe are trying to solve an engineering problem. Main issue (if I remember correcttly ) in a multiprocessor system is cache coherency. Opteron processors are interconnected via HT (hypertransport) and cache coherency is accopmplished via cHT (highly guarderd propietary mechanism). AS offers low latency and RDMA capabilities.

Both IBM blade server and ATCA based blade server are a collection of servers sharing a common chassis and an integrated switch fabric - not really multiprocessor servers (each blade can be multiprocessor).

So I'll not rule out that can be some techical value in what they are trying to accomplish. Good luck to them.
<<   <   Page 2 / 2
Sign In