Slashdot Mirror
Is Hyperchip Hype?
Peter Galbraith writes "There was an interview on CBC (here in Canada) last evening about
Hyperchip, a Montreal-based
company that are working on a new type of router that would scale up 1000 times in traffic (so wouldn't be obsolete in less than a year) and would pass packets to their destination in a few hops instead of a dozen or more. Any experts out there think it's hype? Or real?"
The explanation on Hyperchip's "technology" page is pretty thin, but considering they just raised $70 million, I hope they've given more convincing details to their investors.
I was at a Pub one evening (I live in Montreal) and I happened to meet their sales manager... ms. Jen Goldfinch. Although I had seen the Hyperchip building on many occasions, I had never inquired as to what they do. After meeting this woman, I was given the impression that their routers are actually in use by some of the big players in the digital pipelines game. She was actually pretty clear on that, although I can't seem to find any exact information concerning their customers on their website. Perhaps some questions to nortel, and qwest folks might clear this up. The only thing that make me dubious about her claim of widespread adoption, would be that if their products are so much "better" (for the lack of a better word) than the competitions, then why is abilene using cisco products? Unfortunately I don't have that kind of time on my hands.
It is true; that's why link state routing protocols like OSPF take the speed of the link into account.
What hyperchip gyus did is they build *non-blocking* switch that can scale to 64000 ports and allows to use a variety of OC interfaces. This way instead of using multiple smaller switches to interconnect the same amount of ports (take into account that the links between the switches will be a bottleneck in this architecture) one can use just one switch from hyperchip, which is probably capable of wire-speed switching (otherwise it doesn't have to be non-blocking).
The backplane of this switch seems to be quite fast; so by using it 1) the number of hops gets less; 2)the latency is cut; 3)the space in the NOC is saved.
Could be quite interesting for big guys.
http://www.hyperchip.com/technology_faq.html
Maybe contact IBM and see if they've been working with HyperChip as per this FAQ.
they claim to have 41 patents issued... I found 3:
I/O and memory bus system for DFPS and units with two or multi-dimensional programmable cell architectures
Efficient direct replacement cell fault tolerant architecture
Fault tolerant data processing system fabricated on a monolithic substrate
From these it appears they are fabricating wafers with lots of semi-independent processing nodes, which are tolerant of failures of some of the nodes (and can therefore take into account chip production glitches on part of the wafer).
This could give them a potentially large performance advantage, if they can do it right.
Why do you care what the number of hops is? When you can do line-rate forwarding of packets (as most modern switch/routers can), it's irrelevant. Just because your traceroute shows "2, maybe 3 routers" doesn't mean that's all it's traversing (could be going through a LSP); in fact, I'll put money that you can't show me a trace across the country between two endstations that doesn't go over at least 4-6 routers.
BGP is not a Cisco protocol, it's an IETF draft standard (see RFC 1771); every router worth its salt for the last 7-10 years has supported it.
Also, asymmetrical routes are not necessarily bad; there's load-balancing, administrative weighting...
Lightreading's article, Hyperchip Hypes Its Hardware, claims that Hyperchip stands out in 2 ways:
1) "It's aiming to create something much more than a bigger, faster, box. It's aiming to create the Internet equivalent of a Class 5 telephone switch, something that would sit at the edge of optical backbones and handle IP connections to tens of thousands of users. Hyperchip's developments would potentially replace entire ISP POPs (points of presence) and would have an aggregate capacity measured in - get this - petabits a second."
2) "Hyperchip is addressing this requirement in a totally different (some would say bizarre) way. It's devoted most of its efforts into adapting supercomputer hardware to deliver the scalability it requires. Software - considered the key to success by most terabit router vendors and users - seems to be of secondary importance to the Montreal based startup."
The article says that trials will start at the end of the year. That should prove interesting...
I used to work for Hyperchip. I don't believe I'm violating my confidentiality agreement by revealing that the name the founder Richard Norman originally wanted was "HyperCorp" but www.hypercorp.com was taken so he settled for Hyperchip instead. The company was founded on the strength of Richard's patents in wafer scale integration and the original business plan included a lot of stuff that was accurately described by the term "hyperchip".
That makes their product different. With Cisco gear, you have a very, very small number of high speed ports available to you in one chassis (and several of these "capabilities" are eaten up when you go with redundant solutions).
The idea behind Hyperchip, that is a supposedly better implementation than Avichi's that preceeded it, is to have a packet switching backplane that is expandable to multiple bays, as opposed to the tiny boxes such as the cisco 12000. Since it is a common backplane, there are fewer "hops".
The real limit is power. The Avici systems used over 400 amps of three phase power per bay, and (I believe) scaled to 16 bays, each one capable of running over 60 OC-192s at line speed.
Hyperchip's unit looks better.
Right now, their just trying to figure out how to market it, and how they can add services inside the box that you wouldn't get otherwise. Think of all the stuff you would like to do to streams of that size, but just can't. Also, think of what to do with packets that are going to full pipes. At OC-192 speeds, you can't hold on to packets. There isn't enough time to put it in to memory.
PS OC-192 can carry approximately 10 Gb/s. Or, over 1.2GB/s (this isn't ethernet, it's sonet) (but will have 10 Gb ethernet interfaces, but they just can't carry as much)
At 10 Gb/s, a 1600 byte packet (for those ethernet fans out there) is on the wire (going across a fixed point)for 160 nanoseconds.
Open foot, insert mouth.
CISCO actually preaches in their advanced networking design (one of the things needed to get a CCIE) that all intelligence should be moved outside the core. The core exists only to switch they scream time and time again. They are right.
Examing packets is damn expensive, you don't want that in your core layer at all. You want it moved out as far as possible to move the possible bottleneck as far towards the end user as possible. A well designed network does its job well without having any fricken idea whats going on other at any high levels.
Conversely the upside of this belief for vendors is it helps to sell more equipment since you need more layers to properly shield the core from having to examing the packets.
--- I do not moderate.
Multishelf is the new design. Juniper's next gen product, Gibson is multishelf. Alcatel's (not that anyone gives a shit about their IP products) next gen product, the 777, is multishelf. Cisco's and any other core router vendor that still has money is are also building multishelf systems.
The Hyperchip gear is not revelutionary, it simply can control a lot of shelves and ports from a single vantage port. This doesn't mean that single pipes are going to be running at millions of megabits, it means that you can switch large amounts of traffic across your backplane. This isn't a big deal, its just more then anyone needs right now. The rest of the companies will come out with products that can be daisy chained all day long also, but are busy making in between type products to generate revenue in the mean time.
This also has nothing to do with DWDM, because again, we aren't talking about bigger pipes. Just bigger boxes to plug more pipes into. The current pipes are OC192/10Gig-E. They are essentially the same thing and work just fine on networks today. The next step is OC768/40Gig. Currently most core routers are being made with 10 Gb/s slots. The next gen of routers will be made with 40 Gb/s slots. This is the landscape for the next 3 years. This is also perfectly natural and makes sense to anyone that has a clue. Now the statement that its 1000 times "faster" is just marketing and hype. But people that actually buy boxes like this don't give a shit about hype. They put boxes in their labs, test the hell out of them, and if they like them, they will by 100 million dollars worth of them. Simple as that. Let the ignorance continue punks.