CHIP DESIGNER ARM believes that CMOS transistors and Moore's Law will not provide the performance gains required to meet future processing demand.
ARM CEO Warren East delivered the Mildner Lecture at University College London (UCL) yesterday, and said what most smartphone users already know, that processing demand is outstripping battery capacity growth. However while Intel likes to talk about process nodes, East said gains from transistor scaling are diminishing and will not provide the "useful performance gains" needed in processors.
East mentioned a problem that Intel and every other chipmaker has been facing, the issue of current leakage as transistor scales shrink. Intel talked in depth about leakage when it was briefing journalists on its 22nm Trigate process node and East went even further, saying that the CMOS transistor has effectively had its day. East said, "We've had transistor scaling, we've been getting effectively 'free' increased levels of performance and improved power consumption. As we look forward, these CMOS transistors - you can't make them much smaller - so you need to look at new kinds of materials to solve these problems, or challenges, of performance and flexible hardware."
East continued to say that Moore's law is no longer the industry's saviour that many like to claim because while it predicts smaller transistors, the useful effect of smaller transistors becomes questionable.
"Once upon a time Moore's Law was really great for both performance and power. But beyond a certain stage, power starts going the other way [up], and when you look at what that means historically, you're looking at performance increase, but you're looking at practical performance increase for a given level of power consumption. In the good old days of not very long ago, then most of the gain you got was effectively for free by making the transistors smaller - they went faster and you got more power efficiency. But as time goes on, unfortunately, the amount of gain you get just from scaling gets less and less, so if we are going to make these smaller transistor designs actually deliver better performance, better power consumption, then you have to get clever, you have to get clever with materials and with the device architecture."
Although it was not surprising to hear East talk down the importance of Moore's Law in the future, even Intel has said it needs to go beyond Moore's Law to meet its goals for exascale computing. The cost of shrinking the process node is in the region of $10bn and all indications seem to be that the costs will only go up, and even for cash-rich firms such as Intel it is a very significant amount of money.
Judging by East's comments, we could be hearing less about shrinking process nodes and more about exotic materials in a processor's electrical design. µ
Sign up for INQbot – a weekly roundup of the best from the INQ