Foundry Strikes at the Core

COMMENTS

Newest First | Oldest First | Threaded View

<< < Page 3 / 13 > >>

NetDiva 12/5/2012 | 3:23:24 AM

re: Foundry Strikes at the Core The only way to make a real router cheap is to design and build it in India or China. The real cheap core router will come from Huawei, ZTE or someone like that... certainly not Foundry.
-----------------------------------------

OR by using third party off-the-shelf components on 'Foundry's Direct Routing' boards.

Reply | Post Message | MESSAGES LIST | START A BOARD

DocGonzo 12/5/2012 | 3:23:23 AM

re: Foundry Strikes at the Core It is admirable for a company to try to turn the tables on the high-margin dominants in this space. However, Foundry is not the company to get it done. They may win a couple of smallish customers, but no significant SP is going to pervasively deploy this from them. They do not have the architecture, routing expertise, support infrastructure nor sales relationships to make it work. There are many reasons why the core market is distributed between two dominant vendors. IMO, this bubble in the pool will not change anything and will quickly pass in the wind.

Doc

Reply | Post Message | MESSAGES LIST | START A BOARD

tmc1 12/5/2012 | 3:23:21 AM

re: Foundry Strikes at the Core NetDiva,

Using 3rd party chips gives you faster development (assuming they work as advertised and are not rife with bugs) and lower cost in your internal R&D team with perhaps access to better expertise (this is disputable with many of the chip pushers out there).

It does not however, give you lower cost on a high-end product like a router. It actually gives you higher cost on most products (not hubs, simple switches, basic fabrics though), worse density and integration headaches. Routers already have an abundance of 3rd party chips (FPGAs, MACs, PHYs, Memory, etc.) and the more of these you use the more the COGS go UP. You are better off doing Forwarding and Traffic Management chips in-house IF you have/can get the expertise and JNPR, CSCO and many others follow this plan.

Reply | Post Message | MESSAGES LIST | START A BOARD

arch_1 12/5/2012 | 3:23:19 AM

re: Foundry Strikes at the Core tmc1 objects to my "simple core" concept. Clearly, I did not make myself clear.

As tmc1 points out, the control plane can be horrendously complex, with the protocol du jour, cisco bug compatability, and several million lines of source code. No, the control plane is not simple by any means. Unless you implement the important control-plane protocols, you do not have a core router.

However, all of this hideously complex control plane logic can be implemented in software on a general-purpose CPU. The development cost of this software is high, but unit cost is near zero. CPU cycles needed in the control plane are very nearly insensitive to the speed of the interfaces, and are only weakly coupled to the number of interfaces on the router (depending on the particular protocol.) The control plane operates on control-plane packets, which are a trivial percentage of the packets.

By contrast, the data plane (or forwarding plane) operates on every packet. This is where a router (or any other piece of networking equipment that operates at L2 or above) must process each packet at line rate. This is the level which must be kept simple to keep core prices down.

If your forwarding plane can do IP address lookups at line rate and can do basic priority queueing, you win. If your customers insist on any additional data-plane feature in the core, the costs will increase dramatically.

Control-plane complexity drives NRE and time to market. Data-plane complexity drives unit costs.

There is an alternative to going overseas to drive down control-plane NRE. We can instead use Open Source software. Collaboration on the control plane via open source is a whole lot better than abandoning the market to low-cost off-shore development.

Reply | Post Message | MESSAGES LIST | START A BOARD

OldPOTS 12/5/2012 | 3:23:18 AM

re: Foundry Strikes at the Core I would add some more detail to arch_1 very good explination.

Before forwarding a packet, some processing of the packet must be done to determine whether the packet is a control or data packet, as part of that loose coupling.

Also the control processor can be multiple processors to handle the complexity by performing various functions, like one each for Ethernet, IP (types), and MPLS. The control packet interrogation/separation process could direct each packet to a particular processor based on functional/architectural design.

I believe this is incorporated into most Tier 1 and even now Tier 2 routers/switches.

OldPOTS

Reply | Post Message | MESSAGES LIST | START A BOARD

optoslob 12/5/2012 | 3:23:16 AM

re: Foundry Strikes at the Core tcm1 wrote
"....It actually gives you higher cost on most products (not hubs, simple switches, basic fabrics though), worse density and integration headaches. Routers already have an abundance of 3rd party chips (FPGAs, MACs, PHYs, Memory, etc.) and the more of these you use the more the COGS go UP. You are better off doing Forwarding and Traffic Management chips in-house...."

Sorry but this is VERY fuzzy thinking. The chip business is about achieving the volume needed to reap the benefits of the inherent Batch processing methods used in IC production.

If you look at any advanced digital chip today it will be implemented in 90nm (or tighter process line widths) the mask costs alone are over $1M and if you are trying to integrate functions like 10Gbps interfaces you will probably have pay the mask cost 2 or 3 times before the silicon is working as per design.

Once you achieve "working silicon" you want to sell it to everyone because that one wafer fab can turn out 30000 wafers per month for a cost of under $2000 per wafer. This represents more bridges and routers than the world could ever want, for a cost of a few dollars each. BUT you must have the volume or the whole equation falls apart.

These days one 12" wafer will have between 1000 and 10000 chips on it, generally the fab wants to run at least 12 wafers per lot and to keep the process / product centered they usually want no less than 1 lot per month. So the minimum volume needed to utilize this IC production flow is between 100K and 1M units per year. Even at these volumes the Mask costs of over $1M per mask set (2 sets minimum per new product) account for between $2 and $20 per chip. If your volume is less than 100K than do the sums yourself, however dont forget to halve or quarter your yield because process drift and the odd bad lot will kill you.

The real problem is to define a chip which enables box makers to assemble commodity chips with their differentiated software while delivering the volume necessary to run a product in an advanced fab.

The US industry needs to figure out how to make this happen OR simply get out of the way, because I can assure you that Huawei is having discussions with Asian chip houses about how to achieve the cost models that they require.

Optoslob

Reply | Post Message | MESSAGES LIST | START A BOARD

tmc1 12/5/2012 | 3:23:12 AM

re: Foundry Strikes at the Core arch_1,

It doesn't matter how cheap the forwarding plane is. If the control plane is not stable and well-implemented it is not going in the core of a Tier 1 network. This rules out many vendors including Foundry. That was my point... and Open Source? You have GOT to be kidding. Stop sniffing glue.

optoslob,

You chip economics are great but what if the chip really doesn't work well, is full of bugs or missing needed features (like most of the off the shelf NPUs). It really doesn't matter how cheap it is. Even though it can be mass-produced in volume that does not matter. What matters is how many they can sell to vendors. The volume is not high enough to support those numbers because not enough are being sold so the price is higher (plus their is markup over cost). The smarter vendors choose to develop their own silicon or partner with someone (cisco CRS with IBM) so they control the performance and features as well as the cost of their own product. This may be more expensive in some cases but it is a significant strategic advantage.

I am talking about how this market really works today and not how it could work.

Reply | Post Message | MESSAGES LIST | START A BOARD

rjmcmahon 12/5/2012 | 3:23:11 AM

re: Foundry Strikes at the Core It doesn't matter how cheap the forwarding plane is.

I agree that the control plane is an essential ingredient but I disagree that an inexpensive forwarding plane doesn't matter. Reducing cost on the forwarding plane, while retaining wire rate feature forwarding peformance, is equally important and seems to be as challenging. I don't think that there are many that can get both pieces of this puzzle solved. (I can't comment on foundry's product as I know nothing about it.)

Maybe not necessary for the core, but ever look at something like Snort rules? Try to get those running wire rate at a per port price that a customer will buy. No small challenge from what I can tell.

Reply | Post Message | MESSAGES LIST | START A BOARD

paolo.franzoi 12/5/2012 | 3:23:11 AM

re: Foundry Strikes at the Core
tmc1,

The world's last great mainframe computer - the Class 5 switch is going the way of the dinosaur.

At some point, control plane software for core IP networks will be very easily done. Maybe not today, but it will not remain a mystical place forever.

Planning for it to be is folly.

seven

Reply | Post Message | MESSAGES LIST | START A BOARD

materialgirl 12/5/2012 | 3:23:10 AM

re: Foundry Strikes at the Core Do you need all of that Class 5 functionality in a "dumb" IP network? I thought the whole point of IP was to get rid of as much network intelligence as possible.

Reply | Post Message | MESSAGES LIST | START A BOARD

<< < Page 3 / 13 > >>