What do you do with a billion transistors on a chip?
05/01/2006
Some time ago, a group of chip industry veterans pointed to a quandary facing the industry as we moved toward a billion transistors on a chip and beyond. They did not question the continuing march along the Moore’s Law semi-log track for the next few years, which looked difficult but feasible. Their concern was: What could you do with all those transistors? Memory is never big enough, of course, but what about processors, logic, and auxiliary functions?
Monster microprocessors with steadily faster performance would run so hot they would incinerate. But design costs appeared to be too high for short runs of more specialized system-on-chips (SOCs), especially in the price-sensitive consumer market. Transistors could be used to add functions, but too many functions might make end equipment overly complex, and putting different capabilities on the same chip is difficult using a uniform CMOS process flow.
Not to worry. As always, the industry is finding ways over, around, and under seemingly impenetrable barriers. System-in-a-package (SIP) technology is now developing as an alternative to SOCs, so that optimized processes can be used for different thinned chips that are stacked in low-profile packages. Meanwhile, the march toward ever-faster microprocessors has shifted to multicore designs, with two, four, nine, or even more moderate-speed processors on the same chip. Individual programs can use multiprocessing, multithreading, and a growing array of parallelization techniques, so that they can run faster and with much less energy per instruction (EPI) than in the past, even without a lightning-quick processor.
Some years ago, Gordon Moore explained how, in the early days, a lot of progress came from design techniques rather than simply shrinking circuitry. When most of the circuit tricks were used up, the industry came to depend almost completely on the shrink for progress. For decades, all the pressure has been on toolmakers and developers of materials and fab processes to hew to Moore’s Law targets.
That is changing as shrink rules are reaching some physical limits. Now chip designers, system architects, and software developers are being challenged as never before. Running hundreds of millions of transistors with energy efficiency without sacrificing performance is proving very difficult, requiring new ideas for processing architectures along with clever software. Parallelism used to mean splitting a sequential program into sections that could run separately, but frequent loops made concurrent processing difficult because pieces of code had to keep waiting until other loops were finished before they could execute.
One advance was the Connection Machine designed by Danny Hillis at Intelligent Machines in Cambridge. The architecture was structured around a parallel instruction in LISP, an artificial intelligence language, allowing the machine to run the same program concurrently on large data arrays. It was ideal for jobs like pattern recognition and ranked information searches that are now done by Google-type software on much faster processors.
No one worries about what to do with so many transistors any more. Cell phones with built-in CMOS camera chips, video screens, and Internet links are going with minimal chip counts to push performance and with stacked-chip packages to keep down size and weight. Ultralight notebooks are running movies and high-resolution games along with e-mail access and Internet interactivity. Used well, the extra transistors can offer multiple functions while simplifying user interfaces. Sleep modes, multiple supply voltages, and variable processor speed are some of the tricks used to minimize power dissipation while optimizing performance. The world of applications will open ever wider as all the developers up and down the chain find better solutions, often by collaborative efforts.
The developers of tools, materials, and wafer processes will still have to struggle to keep pushing the Moore’s Law envelope. But as the industry moves forward, they should have lots of company.
 |
Robert Haavind
Editorial Director