Moore's Law will continue to be a viable business proposition for many years to come, because when it is no longer possible to scale in two dimensions, chip manufacturers will simply grow in the third dimension, according to Steve Pawlowski, chief technology officer of Intel's Digital Enterprise Group.
Moore's Law is a rule of thumb in the history of computing hardware whereby the number of transistors that can be placed on an integrated circuit doubles every 18 to 24 months. The law is named after Intel co-founder Gordon E Moore, who described the trend in his 1965 paper, “Cramming more components onto integrated circuits”.
His prediction has proved to be uncannily accurate, in part because the law is now used in the semiconductor industry to guide long-term planning and to set targets for research and development. However, the law cannot continue indefinitely, because transistors will eventually reach the physical limits of miniaturisation at atomic levels.
Last month, scientists at the University of New South Wales, Australia claimed to have created the first transistor from a single phosphorous atom using near-atomic precision, which could keep development of processors on track with Moore's Law until at least 2020. However, most industry commentators expect the law to reach its limit between 2013 and 2018.
Speaking to Techworld at the Xeon E5 launch event in London last week, Pawlowski asserted that while the number of transistors cannot continue to grow forever in two dimensions, the law can be extended by integrating multiple wafer-thin layers of electronic-grade silicon – known as dies – in a stack.
“I'm always asked, is Moore's Law going to end? Well maybe some day, but I plan on working for the company for another 10 years, and I don't see us not continuing on the Moore's Law trend in that time frame,” said Pawlowski. However, a key challenge will be to remove the heat generated as chip volumes become smaller and smaller.
“My guess is there'll be some potential trade-off between the cost of cooling the chip and performance optimisation,” he said. “We always run financial analysis to work out whether we're keeping up with the two-year cadence, and we will run financial analysis to determine when it makes sense to stack multiple dies, versus trying to grow in one dimension.”
Focus on efficiency
Intel operates what it describes as a tick-tock model. Every two years, the company aims to increase transistor density in line with Moore's Law, known as a tick, resulting in higher performance levels and greater energy efficiency. In alternate years, known as tocks, Intel uses the previous year’s manufacturing process to introduce a new microarchitecture, which enables new capabilities.
In 2004, Intel and the other microprocessor manufacturers found that they could no longer produce faster processors, because the chips produced too much heat, so instead they began producing chips with multiple cores (processors) to increase total performance.
Intel's recently-released Xeon E5 chip falls into a tick cycle. It uses the same Sandy Bridge architecture that was released last year, but supports up to eight cores, providing an 80 percent improvement in performance and 50 percent improvement in energy efficiency compared to the previous Xeon 5600 series chips, which had six cores.
“Performance is still a key metric, but so is energy efficiency,” said Pawlowski. “My very first customer meeting after coming back from the labs was with a supercomputer company. I sat down, and before I could even give them my business card and introduce myself they said, 'We hate your multi-core strategy.' I asked why. They said, 'Because we want you to improve memory bandwidth and I/O bandwidth as long as you're putting more cores in the processor.'”
Pawlowski explained that, for customers running high performance computing (HPC) environments, such as the European Nuclear Research Organisation (CERN) near Geneva, power can often be limited, and this has a knock-on effect on performance. By integrating the memory controller onto the chip, customers can make more efficient use of the memory bandwidth and reduce latency.
“Integrating the memory controller onto the chip gives us greater influence over the entire system footprint, not just the CPU,” he added. “Working with network vendors, we can do a better job of power managing but over time it may make sense to have more network functionality closer to the processor to allow us to do finer grain power control.”
Pawlowski said that Intel's E5 processors manage power more intelligently than chips based on the Nehalem architecture, due to a new version of the Turbo Boost overclocking mechanism.
“In a very low utilisation scenario, we'll slow the processor down, turn down the memory unit – we'll save power across the whole platform and we'll take credits for it,” he explained. “As those credits build up, and we come to a high utilisation environment, the power control unit will kick in and push the machine way above the thermal design point for a short period of time, until we can guarantee that it's reliable, and start dropping down again.”
This results in a 12 to 14 percent improvement in performance, compared to just a two to three percent improvement with the previous version of Turbo Boost on Nehalem, said Pawlowski.