A Decade At The Ceiling

Processor performance cannot increase without making energy efficiency a primary objective.


This month marks the tenth anniversary of the introduction of the Intel Pentium 4 HT 570J, which had an advertised operating frequency of 3.8 GHz. It was manufactured in a 90nm process, had a VID voltage range of 1.2V-1.425V and was rated at 115W TDP. In a previous article, Power to Fly, we looked at the graph that I’m including again here below for reference.

Screen Shot 2014-11-05 at 8.12.42 AM

The microprocessor industry, after having gone through nearly two decades of practically doubling the clock frequency every year, suddenly came to a grinding halt. Predictions were made back in 2002 of 30GHz processors by the year 2010. It’s interesting to wonder what people would have thought if they knew that the clock frequency gravy train was about to stop. Contrary to what many maybe would’ve believed, the industry didn’t come to a stop. There were some bumps for sure as engineering teams had to come to grips with the new realities of designing in smaller technology nodes that weren’t going to scale as well for power as they were once accustomed, but through architectural innovations we are definitely seeing new products that deliver more performance per watt.

The plot below shows the past 10 years on the same log scale as the first plot above. The Core i7 4790K, announced in June 2014, is (to the best of my knowledge) the first Intel x86-based processor announced since the P4 570J to have a base core clock frequency at 4.0 GHz. It is listed as having a “turbo” clock frequency of 4.4GHz. It’s a 4-core 8-threaded machine and also includes built-in graphics processing capability and is fabricated in a 22nm process. While the clock frequency is only marginally faster than that 2004 P4, it has substantially more compute capability and it fits within an 88W TDP versus 115W. The 570J is listed at ~125Million transistors and the 4790K at ~1.4Billion transistors.

Screen Shot 2014-11-05 at 8.13.10 AM

There are other processors that run at core clock frequencies faster than 4GHz in the marketplace (some of IBM’s Power processors run at ~5+GHz, for example), but none that I know of that run at anything near the predicted for 2010 30GHz range for desktop/server class machines. This hasn’t stopped the assault though towards exascale computing with a target of 10^18 floating-point operations per second. Given the limitations on clock frequencies, architects must now look to squeeze out more efficiency in terms of getting the highest possible number of operations for every Joule of energy put into the system. In other words, there is no performance without first making energy efficiency a primary objective.