Last week was the Linley Microprocessor Conference. Not the mobile one, which I find the most interesting since smartphones are such a bit part of what drives process technology these days, this is the one focused on networking and servers. But increasingly both markets are being driven by the same thing, namely mobile data. In fact smartphones are growing so fast that they are already the primary way that the internet is accessed and that trend is only going to accelerate. Desktop and notebook computers are a comparative niche. Jag Bolaria (of Linley) gave the opening keynote giving an overview of this part of the industry. Reading their data is a little difficult since they have a, to me, slightly weird definition of embedded, namely processors that last for a long time. I think they really mean anything other than a standard PC, which also has the slightly weird effect of showing Intel losing market share since the more that is done with standard server chips then the more it doesn’t count in the embedded market, as opposed to selling chips that go into routers, for example. Another complication is that Avago purchase LSI and then sold the Axxia part of that business to Intel. For this year, Linley continued to track LSI as a separate company.
Overall market share is lead by Intel followed by Freescale. To my surprise, AMD increased market share against Intel in this market since suppliers are looking for lower costs and an alternative to Intel. Cavium are also doing well selling their Octeon SoCs into many segments. Of course ARM don’t appear on this list since they don’t sell processors. However, when we look at share by instruction-set-architecture (ISA) they show up, but with a suprisingly low share. x86 slightly trails the Power architecture (Freescale, Applied Micro, Axxia for now, all use this).
ARM has a tiny share. But as I reported last year, that is all set to change. The 64-bit ARM v8 instruction set has opened up new markets and almost all embedded vendors are moving their future investment to ARM. However, the time to design-in, ship and ramp equipment in a conservative market means that the crossover will take 5-10 years, but:
- AppliedMicro shipping X-Gene and sampling X-Gene2
- Cavium plans to sample Thunder in Q4 (their current products are MIPS based)
- Feescale sampling LS1 and plans to sample LS2 this quarter
- LSI/Avago/Intel shipping ARM version of Axxia (although presumably this will be short lived now Intel owns that business)
- AMD sampling Hierofalcon for embedded market
- Broadcom shippping StrataGX and developing Vulcan CPU
Linley believes that a dual architecture strategy (such as MIPS and ARM) will not endure since qualifying two architectures is too expensive. So, provided the ARM parts are not rejected by the market, then MIPS and Power will lose market share.
As I said above, the market is driven by cellular data. Next year there will be 2 billion connected devices shipped, around half of which will use high data rate LTE (a big part driven by Chinese deployment although based on Xilinx conference calls and other information this is somewhat pushed out). A new technology like LTE forces the operators to focus on coverage (so then they can sell handsets) and then on capacity expansion as the transition happens and more handsets use more and more data. Capacity expansion means small cells which is a big growth opportunity.
In the server market, no surprises, Intel has 96% market share. Their strategy is to leave no holes. No matter which price/performance point you are interested in, Intel has a processor for you: Xeon E7, E5, E3, Atom and even customized processors. But ARM has this market in their sights along with their partners. As I said above, AppliedMicro are in production, AMD, Cavium, Broadcom and more. Calxeda were a casualty, only having 32 bit ARMv7 and not enough money to do 64-bit. Jag also said “expect more entrants” which I suspect he knows some non-public stuff that I don’t.
But there are challenges for ARM in servers. Unlike with Intel there is no common platform and each vendor has its own uniqueness. OS vendors are sitting on their hands waiting for volume before they invest. Of course they are attacking a walled-city but Intel has very high pricing (good for them) which creates market opportunity. But it is hard for ARM processors to match Xeon E5/7 performance, which matters for some markets (where single thread performance matters) but not others (where cost, power, physical size and general TCO is important, such as, say, Facebook data centers).
Linley himself talked about the Internet of Things (IoT). Industrial applications are already broadly deployed with, for example, 300M smart meters installed. For that market a discrete solution that costs $20 is OK. Large consumer goods are next with an SoC desired with a cost under $10. But the really big market is small things with an SoC at under $3. I’m personally not convinced since the market is so fragmented. Also, I’m not convinced that IoT devices have minimal compute needs since they also have very low power needs which means thay cannot be uploading huge amounts of data to the cloud continuously for processing. But it seems some vendors are developing IoT ASSPs integrating CPU, memory, analog and wireless to drive the cost down.
Finally, a little bit of process roadmap. But one message is that cost-focused processors will stay on 28nm LP indefinitely. Jag pointed out that 16nm reduces power but increases cost and the density is the same as 20nm (it has the same metal fabric). For high end servers that is an acceptable combination but for IoT and other lower end applications, 28nm will be around for a long time.