Every year in the fall, the Linley Group runs their processor conference. There are other conferences earlier in the year that are more specialized, addressing servers, mobile, and IoT. The conference, originally called the Microprocessor Forum, has been going since the late 1980s, originally under the umbrella of Michael Slater's Microprocessor Report , then by Cahners, and today by the Linley Group, who still also publish Microprocessor Report . As usual, the opening keynote was by Linley Gwenapp, who focused on the economic slowdown (halt?) of Moore's Law and the consequent scramble for more performance through specialization. Moore's Law In the past, Moore's Law was the main factor that drove processor performance. I was at the Microprocessor Forum in the mid-2000s and a famed processor architect pointed out that all the architectural innovation up to that point had produced a performance improvement of about 10X and all the other increases in performance had come about by improvements in semiconductor process technology, both just making the processor itself faster and enabling larger instruction and data caches. But that is not long, in Paul McCartney's grammatically incorrect words, the world in which we live in. Moore's Law has stalled economically, in the sense that you can't just get increased performance at the same cost by waiting a couple of years. Yes, you can still get more performance, but the cost per transistor is either flat or declining much more slowly than historically. You can get twice as many billion transistors on a chip as in the previous generation, but it will cost you twice as much. The foundries and IDMs (aka Intel) would dispute that it is as bad as that but we no longer in what An Steegen of imec calls "happy scaling." One side effect of Moore's Law was that there was little time for architectural innovation. All the effort went into moving designs to the next node and general-purpose designs were the only ones with enough volume to drive the investment. But specialized architectures can produce 10X to 100X the performance/watt. The two examples Linley quoted were Cadence's Tensilica procesors, and Microsoft's Catapult, which is an Altera (now Intel, of course) FPGA programmed to accelerate Bing searches. The two areas where this dynamic is most notable are vision processing and neural networks. Both of these can be done on general-purpose processors but specialized solutions are dramatically more efficient. Datacenter growth continues to be strong, now over $10B (for the processors) and growing at 11% CAGR, with public cloud (Amazon, Microsoft, Google, Alibaba, etc.) the fastest growing sub-segment. Intel's Xeon family has "literally 99% market share" but it is under attack. There are three primary attackers: AMD has its new Zen CPU, which has increased per-clock performance by 40% (plus faster clocks) IBM created OpenPower to promote its Power architecture in third-party servers ARM and its partners have been working hard on improving single-thread performance to get close to Intel, with Applied Micro Gene-x the most visible but with Qualcomm and Broadcom lurking in the wings All three of these architectures were presented in more detail during the conference, and I'll write about them in depth in a separate post. IoT Moving away from the high end, another potentially large growth area is the Internet of Things, IoT. This is not an independent market, though, since IoT devices almost all connect back to the cloud for back-end services. The big bifurcation in IoT is between industrial IoT (sometimes called IIoT) and consumer IoT. Consumer is driven mostly by convenience, status, and social engagement. Although the devices need to be cheap to drive volume, cost savings directly are not the driver. For example, smart light bulbs might save you some money on electricity in the long run, but the motivation to put them in would be more the remote control, perhaps the capability to change color. And they are cool. Industrial IoT is the other way around. It is all about cost saving. In fact, some business models are based totally on this. For example, Joe Costello (whose name should be familiar since he used to be CEO of Cadence) is CEO of a company called Enlightened that will replace all the lighting in an office complex, with no up-front cost, and take a cut of the energy savings. The big challenge in the IoT market is security. One thing everyone was talking about was the St Jude Medical story. St Jude makes implantable medical devices like pacemakers. A security company called MedSec analyzed them and found security issues. They teamed up with a private equity company called Muddy Waters to short the stock. St Jude has sued them that the warnings are false. But that is another unexpected business model: short the stock of companies and expose security weaknesses. It is the IoT devices that are the weak point, due to lack of resources. The "things" typically don't have powerful processors, and are battery powered. But encryption is inherently computationally expensive, otherwise it isn't secure. It is generally accepted now that security needs to combine software solutions but build upon a hardware "root of trust". Automotive Automotive (and mobile) are regarded as separate markets from IoT with their own issues. Not least is that cars have to operate 20 years in extreme cold and heat. Even today, when processors mostly control seats and mirrors, with a subsegment for things like engine control, the processor market is $10B. This is expected to double by 2025, driven by ADAS and autonomous driving. Linley's view of the current status of autonomous vehicles is: Tesla Autopilot, Volvo Pilot Assist, etc. are available now but require continuous driver supervision Initial autonomous vehicles likely in 2018—but self driving only in good conditions on highway Ford has promised fully autonomous vehicles in 2021 incremental hardware cost below $5000 by 2022. For personal cars, cost may be reduced by cheaper insurance The über conclusion: In post-Moore era, processor architectures are becoming specialized. Previous: Memories Are Made of This: Preview of MemCon![]()
![]()
↧