Blessay: Comparing Merchant and Custom Silicon

Merchant Silicon is an marketing term used to describe the use of “off the shelf” chip components to create a networking product and commonly used by company that design their own silicon chips when explaining that their process is better and more efficient.

The networking industry has two types of silicon: ‘custom or in house’ or ‘merchant or off the shelf’. The ‘in house silicon’ argument goes something like this :

We have our specialist engineers who can custom design an ASIC and supporting chips that are completely focussed on the features needed to deliver Product X. As a result, our designs are purposefully built to be faster, better etc. Our prices are more expensive because our products are faster and more focussed and because our costs are higher.

The ‘merchant silicon’ argument goes something like this:

The levels of expertise required to design and build a silicon chip is very high. It requires specific skills, tools, and high levels of expertise. If your core business is producing a networking device, then having your own silicon development is not core business. By relying on companies who specialise in silicon design and manufacture, we can focus on software and integration to deliver new and better features at a cheaper price.

The Product

When making routers and switches, only a very few components are custom designed as most components are ‘off the shelf’ – for example, the GBIC, the ethernet PHY chips; the management boards always use industry standard CPU such Intel x86 processors, or sometimes i960 network processors and use off the shelf memory and flash.

The custom silicon is usually specific functions such as the frame switching chip, the backplane forwarding (aggregation), specialised packet handling etc.

There are vary large companies who have extended their business from manufacturing ethernet network card components to silicon for ethernet switches. For example, Broadcom makes chips for 24 and 48 port switches as standard components, and are rumoured to make complete Ethernet switches for Google and Facebook data centres or are also OEM’d by other companies.

Companies such as Fulcrum Microsystems make fast, feature rich, silicon engines that are comparable with the largest networking vendors. This is the rough basis of products for startups networking vendors such as Arista networks and others.

HP Networking’s ProCurve division appears to have been attempting to position themselves as a silicon vendor by supporting the OpenFlow software initiative (although current position is unclear).

The Shrinking Skill Base

In the last five years, the rise of the Intel x86 CPU has seen many networking companies move to using standard Intel motherboards. For example:

  • Fortinet mostly uses a standard Intel board with some limited custom silicon for security processing
  • I believe that F5 uses a standard Intel chipset for management, a off the shelf ethernet chipset for interfaces, off the shelf crypto chip for SSL, and just a few custom chips for forwarding.
  • Cisco ASA is an Intel Celeron board with a couple of off the shelf crypto chips for VPN / SSL acceleration.
  • Bluecoat proxy servers are Intel motherboards in standard cases.
  • CheckPoint uses Intel x86 CPUs and some off the shelf silicon for improved forwarding performing in the newer platforms.

As companies like these turn to standard components (sometimes never using any custom silicon), the skills base gets smaller the cost of manufacturing tends to rise. While this has been offset by improved software tools over the last decade there are less and less people designing chips.

And the price of building a large scale silicon fab costs in the billions. While there are a number of manufacturing plants, very few companies can afford to own the fab and choose the outsource the actual production to these companies.

The Risk of Failure

Companies who design their own silicon are taking a significant risk. In the modern business environment of risk free profits, there are limited risk mitigations for a failed chip development. The current fashion of MBA graduates to protect themselves by outsourcing or using proven technology means that more and merchant silicon will be the norm.

Is There an Impact ?

Custom silicon has been dying for a number of years, in my view, and merchant silicon has been working fine. It’s proven market effect in the server and smartphone marketplace so it’s likely to happen in networking as well.

What is the impact on customer ? Not much. Routers will route, switches will switch, and processing will occur as it always has. In the end, its more likely that silicon will get faster more quickly than our performance needs will rise. Already, merchant silicon companies are announcing products at or in excess of those used by Brocade and Cisco today.

The only concern is that networking adopts a feature of the storage market where all products must pay protection money be validated because of failure in the standards robustness. There is enough variance in FC products that they must be validated for use. It’s look to be a good money spinner in the storage industry and the temptation for big networking companies to follow a similar path will be strong.

Better, Worse or Indifferent.

The “custom silicon is better” argument has been comprehensively disproved by Intel, AMD and ARM. Once upon a time, CPU’s for computers were made by many companies and yet the industry has converged on just two core CPUs and half a dozen supporting chipsets. ARM is offering opportunities for servers in the future and dominating the portable electronics market, Intel owns everything else and AMD tries to keep up. If custom silicon could do it better , faster or cheaper then we would still have many vendors of CPUs.

It’s proven that companies who focus on making silicon chips, can do it faster and more efficiently than non-specific companies who attempt to develop their own. Today, Brocade and Cisco make their own silicon for most of their products in an attempt to differentiate themselves from competitors in their high end products but it’s questionable whether that approach has a long term future. It’s likely that high rates of forwarding performance can be delivered using Non Blocking Spanning Switch designs based on Fat Tree CLOS designs can readily be delivered by clusters of chips instead of custom designs.

I have the view that Merchant Silicon will dominate eventually, and physical networking products will become commodities that differentiate by software features and accessories – not unlike the “Intel server” industry (you should get the irony in that statement). As a result, any argument between “which is better – merchant or custom” is just matter of when you ask the question.

One interesting feature is that John Chambers continue to publicly state that custom silicon is their future. The are parallels with Sun Microsystems who continued to make their own processors in the face of an entire market shift, and that doesn’t appear to have worked out very well. In this another wrong footed innovation from Cisco ? Time will tell.

  • DGentry

    Add Extreme Networks to the list of companies where proprietary ASICs didn’t equate to a sustainable advantage. In the late 1990s/early 2000s Extreme’s ASICs had some big advantages, mainly in the amount of buffering they could provide compared to BRCM. By the mid 2000s those ASICs had turned into a disadvantage, when the next generation was late and expensive. By the late 2000s they were using mostly merchant silicon… but probably too late.

    • Greg Ferro

      The fact that they are still in business tells you something about the success of the model.

  • JBL

    What if there is an appropriate time/place/product to roll your own silicon and another area where off-the-shelf just makes more sense? Why does it have to be death to custom ASICs? There may always be an application where custom ASICs are better suited for a particular job, no?

    • Greg Ferro

      I’ll use a car metaphor. Consider custom silicon as F1 (assuming that you believe that custom silicon is better, which isn’t a certainty), only a few people could afford to buy any racing car, and the maintenance costs are phenomenal. Everything is custom parts, and requires a lot of specific skills and planning. For day to day use, conventional cars work and offer good capex and opex.

      For custom silicon, it’s hard to do it well, requires a lot of management and control to keep running. If Cisco’s business was going well, and new products with useful features where coming to market, we would know that Cisco is doing a good job. However, Cisco appears to have lost it’s focus and control and we are seeing rumours that they will release merchant silicon equipment in the near future.

      So, no, custom ASICs are not a guarantee of excellence, not is merchant silicon. But Merchant silicon has less chance of outright failure and that’s a better business proposition overall.

  • Oliver Gorwits

    Custom silicon can sometimes provide a “kitchen sink” of features which allow the product to be useful beyond its original planning. For example you might find MPLS missing from the merchant silicon based products, but it could be useful the day you need it in your enterprise backbone.

    However I think it’s the long dev cycles of custom silicon which might do them in, in the end. Cisco’s Nexus has 10 year old designs which are super power hungry compared to modern, two year old merchant designs. That affects both the bottom line, and PR – two things a customer pays attention to.

    • Greg Ferro

      Modern silicon can be extensively programmed to support new features. I don’t think the limitations exist in the same way as silicon design and manufacture has moved ahead. Software can enhance the functions of the chipsets much more than in previous years.

      And the focus could allow them to progress more quickly to new designs. Yes, the Nexus looks old but only when compared to merchant silicon choices.

  • Pavel Bykov

    I don’t think Intel/AMD argument applies here. You see, ASICs have not followed CPU’s path. In CPU market, we have dominant forces that make advances before anyone else (e.g. Intel/AMD/Arm trio you mentioned, and others). But that’s not the case in ASIC and fabric department. Check out the architecture of Cisco Nexus 7000, Juniper QFabric, Cisco CRS1/3, and others. And what about 100G interfaces? The truth is, the generic ASIC makers are always catching up, copying, and are not innovating quite fast enough. Sure, I do agree in favor of generic ASICs once the technology is spread broadly, like in low-mid range switching. But 100G switching, complex non-blocking architectures, decentralized failure-resistant self-healing multi-terabit hardware arrays are still beyond wild dreams of generic ASIC makers.

    I remember the days of x86/RISC/SPARC/Alpha, and clearly x86 was winning, but it was because they were ahead. Generic ASIC makers are not ahead, they are way behind. Cheaper, but behind. And also I wouldn’t underestimate interoperability nightmares.

    Once Internet will be standardized an innovated by the generic manufacturer(s), and they’d be near the top of the following list, that’s the day the tides will change:

    But for now, they’re waiting for others to innovate.

    • Greg Ferro

      I think the problem with your logic is that you assume that manufacturers such as Fulcrum will not be able to compete and move ahead. I believe that a company focussed on making ASICs will develop, innovate and progress faster than a company that has dozens of business units all fighting each other to develop a coherent strategy.

      Also, remember that Cisco outsources many of the chip development processes anyway. Testing, design validation, audit, specialist design are just some of the areas that Cisco / Juniper etc outsource to dedicated small companies. And these smaller companies are equally available to the bigger merchant silicon companies too.

      As Cisco has gotten bigger, it is less able to progress. And now that management focus is on user-centric products such as Linksys, Cius, IP Telephony this business will suffer.


  • Pavel Bykov

    I wasn’t talking about potential or capabilities, I was talking about current situation. I agree with you that a focused company are able to move ahead, but they are not. Not yet. x86/Intel has gained a lead by being there before any one else.
    No comment about Cisco/MSFT bursting with their own glut. They are way overextending themselves. Together they had $1T valuation 11 years ago, nearly 3% of the world economy. Now together they have $300B, 70% down, 0.6% of the world economy.

  • Pingback: AeroHive, HP, ‘Big Boner’ AP’s and Wireless LAN Controllers()

  • Pingback: Aerohive | In Search of Tech()

  • Douglas Gourlay

    Greg, very valid points made on here. There is another thing to consider though, and its which process does a vendor use when making their own silicon versus a merchant silicon vendor.

    Every merchant silicon supplier uses a full custom process. This means they write the logic (verilog), they hand place the gates onto the substrate, do the metal-layer routing, and ensure the timing of the entire system is laid out so that each function of the pipeline is within one clock cycle of the next stage. This means less wait-state, faster chips, lower power per unit of work performed, etc.

    This full custom process lets these vendors use the latest in 40nm technology, go to 20mm per side on the die, and clock their chips at around 1.5Ghz.

    By contrast almost every ASIC implementation coming out of the single vendors is a traditional ASIC flow. This means they write the logic/verilog, but then outsource the timing, design, packaging, metal-layer routing, to an automated tool from IBM, Cadence, LSI, etc.

    This process is faster, and more fool-proof, but doesn’t always work with the most advanced processes, usually the yield curve melts north of 17mm on a side, and clock rates are usually about 500-600Mhz.

    Net-net – you can get 4x the transistors, 40% more space to put them, and clock them 3x faster on most full custom systems versus ASIC flow.

    This is exactly why you see Broadcom shipping their Trident system with 64 ports of 10Gb Ethernet on a single-chip versus a single-vendor system which takes 15-20 chips to achieve the same in a 1RU, or since this is the same density of wirespeed L2/L3 ports as, say a Nexus 7010 that uses over 240 pieces of silicon to accomplish the same density of wirespeed 10Gb L2/L3 ports.

  • Pingback: EOL/EOS for the Cisco NAC Network Module (Clean Access) – Cisco Systems()

  • Anonoymous Coward

    Greg, i think you may need to stop swallowing every line that HP serves to you. Rmember that about 5minutes before buying h3c HP were touting the benefis of of their provision ASICs as the best thing since sliced bread, now that they have H3C its all about merchant silicon. Im not arguing that one is better than the other, to be honest, ithink this commentary is irelevant, ultimatley all that maters is the end product; features, functionality, total vision…. Its the reult that matter, not how you get there.

    • Ferro Greg

      I agree – on one hand they are say how fantastic their product development is, then they produce a merchant silicon switch. Cisco does the same thing.

      Merchant Silicon matters a great deal. It completely changes the way products come to market. Consider when Mini-computers were the thing, each with their own CPU, Memory, Drives etc and compare this today where almost all servers are Intel Merchant Silicon.

      It has significant impacts on how the networking industry operates and that’s what i’m tracking. In a sense, this product is unremarkable except for the Broadcom chipset.

  • Pingback: Proprietary? So what. | In Search of Tech()

  • Mark

    “Sun” SPARC is no different from MIPS, PowerPC, or ARM.  SPARC is a committee based set of specifications which are licensable.  I have a Netgear ReadyNAS which uses a SPARC processor.  Granted, SPARC did not gain the same level of embedded success as MIPS and ARM.

    The one question I have, is: “If merchant silicon is good, why not merchant network operating systems as well?”  The idea the Ethernet switches based on merchant silicon are like PCs is simply not true.  Dell, HP, etc. are not adding value with their own proprietary PC operating systems.  Force10, Arista, etc. are essentially the equivalent of Sun and the other workstation vendors in the mid 1980s, before they developed their own RISC processors.  They were all building computers based on merchant silicon (most based on the Motorola 68000 CPU), and their value add was their operating system.  Ultimately, not just the microprocessor was commoditized (after a period of CPU decommodization), but the workstation and small server operating system was commoditized as well.

    Imagine if Broadcom or Marvell decide not only to provide all of the network ASICs, and reference switch designs, but also a NOS which ran on those reference switches?  Or better yet, an outside company, let’s call it “Microswitch”, developed an OEM NOS for companies who just wanted to build and sell switches, and focus on logistics, not software development.  This may become possible to some extent with externally manged, OpenFlow compatible switches.

    • Etherealmind

      Yes, it’s an interesting train of thought. There are already open source OS’s for these switches – currently intended to deliver minimum function as OpenFlow switch.

      And there are a number of white box switch makers already delivering products in this space and are rumoured to be widely used by very large cloud providers.

      • Harsh Singh

        Cumulas and Pica8 are doing just that aren’t they. The large cloud providers, “MSDCs”, have the technological prowess to deal with the disaggregated solution. The question really is will the broader market be able to adopt this solution.

  • Pingback: Merchant Silicon and Vendor Software – The Hype in 2012 — My Etherealmind()

  • averagejoe

    I hate the fact that many tech companies are outsourcing products to other OEM companies. Yes I know most of the stuff is designed by the primary company, and I realize it’s cheaper for them to have someone else make a product, but man, what are we even buying anymore? I know all laptops are pretty much made by OEMs that most people have never heard of, but I CERTAINLY HOPE that companies like HP, Dell, IBM, etc. are assembling AT LEAST their servers and high end workstation products in house. If it must be overseas, at least make your desktops, servers, and workstations in your own factory. That’s why I like companies like Cisco, EMC, many semiconductor companies, Motorola Solutions, etc. because they assemble and make their products in their own factories. I have an Iomega NAS and it is assembled in USA. Many Motorola Solutions products (e.g., two-way radios) are made in USA or Canada, or at least in their own factories in Malaysia. Companies like Seagate, Western Digital, and Hitachi make their own drives in their own factories in Thailand (well Hitachi storage is now WD).

    Moral of the story: if you’re going to design a product, make it yourself. You don’t have to be completely vertically integrated, but at least put the motherboard, hard drive, processor, and casing together in YOUR OWN plant. This goes for other kinds of tech companies too, like networking, printers, servers, semiconductors, POS, and data storage.

    Not all companies have done this of course. Like the article says, Cisco mostly makes their own silicon, and I’m sure IBM puts together their own stuff. All the hard drive makers have their own plants, and many semiconductor companies are NOT fabless and do have fabrication plants of their own. Lexmark makes their own printers, and HP makes their own colour printers.

    I just hope companies don’t take the cheap way out. I hope they continue to not only design their own products, but make their own as well. Making them in North America would be great (to support my local economy), but if it must be overseas, please, in your own factories! I will pay a little more for peace of mind and also to support companies that pay taxes to my government, which provide amenities for me (roads, hospitals, schools, etc.).