The debate over custom and merchant silicon is an old one, but it’s gaining steam driven by developments in software.
- What matters to IT buyers is that the product provides adequate performance, and vendors using custom silicon need to make their case.
There is something of an intellectual debate occurring in networking over the need of custom versus merchant silicon. It’s not a particularly new debate, but the rise of white box switching, an ever increasing number of switching products coming to market using merchant silicon and the increased focus of software for both advanced features as well as tight integration with the rest of the environment is making the debate far more relevant.
On the one hand, vendors like Arista, Extreme, Juniper and others, have made great progress with merchant silicon by adding capabilities via software. Arista’s EOS is a Linux OS with a shell and you can run many Linux-based programs on it. Extreme’s XOS, also Linux based, was one of the first, if not the first (though no one knew it) networking vendors to add scripting capabilities to their OS many years ago. These products work well because there is a clean separation between the forwarding features that are implemented in hardware which perform very fast and management features residing in software that are outside the data plane. Today, there is further competition in the switch space with bare metal switches from ODM’s like Accton and Quanta as well from mainstream vendors such as Dell, HP, and Juniper which are selling white box hardware and partnering with companies like Big Switch and Cumulus Linux.
On the custom ASIC side, vendors like Alcatel-Lucent Enterprise, Cisco, Extreme, HP, and Huawei that have custom ASICs in some of their products in order to perform processing outside the capabilities of typical merchant silicon like application detection and classification, security processing, etc. Even Arista, the poster child of merchant silicon, has a switch model with FPGAs built in because the merchant silicon on the market couldn’t add enough performance to power the applications used in financial and transactional networking. There is still a great deal of benefit from software on these switches, but through custom ASICs and FPGAs, these switches can do much more than just switch frames from one port to another.
I see the decision breaking down like this:
- Merchant silicon products are for you if all you need is fast forwarding and routing of typical standard based protocols. Aspects like Layer 2/3 protocol support, stacking, port density and performance will be uniform across many vendor products and any differentiation will come from the switch OS and other software components. This uniformity in merchant silicon enables things like disaggregation of the switch OS from the hardware.
- Custom silicon is for you if you need everything that merchant silicon provides plus additional processing capabilities at speed such as packet classification or programmability of the hardware in the field.
Thing is, using merchant or custom silicon isn’t an either/or decision. Whether a switch uses merchant silicon, custom silicon, or a mix (it’s almost always a mix) only matters is as much as IT buyers want to be sure the product can do what it needs to do at the required performance levels. If a vendor’s switch uses custom silicon, then IT buyers need to understand the advantages of custom silicon and have the performance claims backed up with data.