The dominant technology in computer manufacture from personal computers to supercomputers for more than a decade has been VLSI semiconductor based on silicon. Continued growth in this technology over that period has led to orders of magnitude improvement in performance, device density, and cost. Other semiconductor technologies, in particular Gallium Arsenide (GaAs), provide alternate operating points than silicon with different trade-offs. The Semiconductor Industry Association (SIA) has developed a detailed projection of the most likely path for semiconductor technology evolution up through the year 2007. Recent examination of the dominant issues has resulted in tentative extensions to the year 2014. These results have proven key to determining viability of PetaFLOPS computing systems within the next two decades and have provided the basis for the PetaFLOPS computer architectures proposed at this workshop.
The state of the art in silicon semiconductor technology employed currently in delivered commercial computers is dominated by feature size which is now approaching 0.5 microns. Current manufacturing yields enable processor chips of half a million gates. DRAMs which provide the bulk of main memory are being delivered with 16 Mbits and SRAMs used for very high-speed memory and caches have up to 4 Mbits. On-chip clock speeds have now reached 200 MHz, although DRAM access times continue to be substantially slower with typical values in the 65 nanosecond range.
The SIA projection shows a steady rate of improvement in most key parameters out to and beyond 2007. It is estimated that at that time, feature size will have reached 0.1 microns. This will enable more than 20 million logic gates to be integrated on a single chip. At that time, DRAM will have a capacity of 16 billion bits and SRAM should be capable of storing 4 billion bits. Advances in clock speed and logic performance will not be so dramatic and is not expected to go much beyond 1 GHz or a clock cycle time of 1 nanosecond. During this time, logic voltages will shrink down to 1.5 volts. Even so, power consumption will be an important issue. It is projected that high-performance dies may experience power requirements of between 40 and 200 Watts, a significant increase over the 10 to 30 Watt demands of today's high-end processors.
There had been a serious concern that beyond this point in feature size, quantum effects would begin to dominate and new models would be required to estimate progress at finer resolution. However, recent studies have indicated that current trends will continue to feature sizes as small as 0.025 microns by the year 2014. However, it is at this level that semiconductor technology begins to make PetaFLOPS computer architecture viable.
The potential of GaAs is less understood. Significant advances have been made in this technology in the last five years with commercial high-end computers being delivered incorporating GaAs integrated circuits. And new systems employing this technology are being developed. Clocks speeds between a factor of 4 and 10 times that practical with silicon are achievable using GaAs. However device density is measureably less while costs and manufacturing difficulties are significantly more. It is clear that this technology is having an impact on high-performance computer architecture. It is premature to assert that it will become the device technology of choice in the future.