System-on-a-Chip (SoC) vs Network Processor (NPU) – What’s the Difference?

This week Broadcom announced their latest SoC a “3G phone on a chip”. It reminded me that I’ve been very hard on Network Processor’s (NPU‘s) for a long time, and that I’ve been rather silent about my support of SoC’s like BRCM’s. What gives?

Many point out to me that SoC’s and NPU’s are technically the same. NPU’s and SoC’s are both highly integrated “programmable” semiconductor devices. The subtle difference is that an SoC vendor is deploying programmable elements to solve specific problems in a narrower application space. The vendor of an SoC is providing a finished ( or close-to-finished ) solution. The SoC vendor could have easily “hard-coded” the device, but the best way to provide the solution was to deploy programmable elements.

In summary, the subtle difference is that an “SoC Solution/Ecosystem” is provided by a “system focussed vendor leveraging technology” and the “NPU device” is provided by a “technology focussed vendor”. The SoC vendor knows the system problems that they want to solve and thus can prove that the solution delivers. The NPU vendor does not fully know the system problems that their device will be required to solve and thus have to work very closely with their customers to prove that it works.

Advertisements

2 Responses to System-on-a-Chip (SoC) vs Network Processor (NPU) – What’s the Difference?

  1. ozman says:

    OK, it’s taken me a while to be able to formalize my thoughts enough to articulate a response.
    I had a visceral reaction to the assertion that NPUs are a technology sell and SOCs are a system sell. My contention is that they are pretty much the same sell, and both are doomed to failure if poorly conceived.

    Simply from first principles, the cost of developing a large scale SOC is of the same order of magnitude as developing a large scale NPU. How the money is spent might be different – NPU vendors focus on RISC core and programming tool development, SOC vendors costs are around IO, integration and verification. However, at the end of the day, both cost around the same. Which kinda means that they need roughly the same size business oppty to meet the same ROI targets. And that’s where the difference is. NPU vendors hace gone after line card sockets in networking equipment. SOC vendors have gone after consumer oppty’s – handsets, printers etc. There just maybe enough money in those markets to justify the intial investment.

    However, what comes next? After the silicon Gods have given you 2-4x more gates in next process node, most vendors go and stick more application specific functionality into the SOC device – increasing development and verification costs, and simultaneously reducing the market oppty. This sounds a lot like the curse of the ASSP.

    A second more logistical point – I think you vastly under-estimate the cost and challenge in creating an broad based mult-vendor eco-system. Unless your SOC has an embedded x86 processor, you’re on your own, or it’s a very, very expensive proposal. You end up paying for the same work to be done at Windriver, MontaVista, QNX, and so on. The NPU vendor didn’t have the requirement to create an eco-system, and was simply developing a point solution (Si + s/w tools).

    At the end of the day, whether it’s an SOC or NPU (so what is a Cavium Octeon BTW?), it gets down to how much it costs to design, verify, make and sell, verses how much money there is out there. It remains to be seen whether the SOC market goes the same way as the NPU market – where the Si vendor has to win the 800 pound Gorilla just to break even on the investment.

  2. Iain Verigin says:

    I’ve been stuck on how to reply.
    * On one hand we agree in many places 🙂 and on the other hand we disagree a lot 👿
    * In your 3rd paragraph you seem to agree with my post that NPUs focus on developing a technology platform and that SoCs focus on building focussed solutions.
    * In your 4th paragraph you introduce a new point. I agree with this point. Technology platforms like NPUs ( and PLDs) scale much better than ASSP’s and SoCs.
    – this seems to be helping PLDs win more business than NPUs.
    * wrt/ ROI … it depends on how you calculate it for the NPU industry.
    – historical return on sales — sucks!
    – historical return on investment — awesome! ( startup investors cash in while the acquirers of NPU technology all look at “the emperors new clothes” )

    * I get the feeling that you’re just not liking that I believe more in SoCs than NPUs
    – Maybe I’m stuck in the past where I can’t remember many NPUs that have made any return on sales. I can’t remember many NPU companies that have done very well for their investors.
    – SoCs have done very well measured on sales, but few have done well for startups. Because as you point out. They cost a ton to develop.

%d bloggers like this: