This section discusses device technology and its significance with respect to the possibility of building a PetaFLOPS machine in about 20 years-assuming that device technology continues the same trend in the future that is has taken in the past.
Since memory costs are predominant in two of the three design studies, consider the Semiconductor Industry of America (SIA) data for future technology. Figure 6.4 shows projected chip density following a straight line on a semilog scale for the next several years. The SIA projections stop at the year 2007 when feature sizes are sufficiently small to cause tunneling effects, although newer data from SIA cited earlier suggests that technology advances will continue beyond 2007, although then it is not clear what direction technology will take. Nevertheless, if lithography and feature sizes continue to diminish for whatever new technology comes along, the projections out to the year 2014 produce densities high enough for the construction of a PetaFLOPS machine at reasonable cost.
These projections account for a doubling in density in each linear
dimension (quadrupling of density overall) every three years. A
rough rule of thumb is a 10 times increase in density every five years, so
that 20 years should produce a net of increase in density and
a corresponding decrease in parts count. The parts numbers tabulated
in the previous section are quite manageable, and indicate that in 20
years the petacomputer will cost about what a supercomputer costs
today.
These findings are based on the previously noted crucial assumption that the technology road map can be extended beyond 2007, and therefore depend on a device technology that involves a tunneling phenomenon or some other physical phenomenon to produce a family of devices that do not yet exist.
Figure 6.5 shows parts counts as a function of memory density. These parts counts are valid for Category I and II designs if the memory capacity of a PetaFLOPS computer follows the scaling law of one gigabyte of main memory for each GigaFLOPS of performance for applications in the PetaFLOPS region; otherwise, a design can make use of less main memory. Since memory costs are a major factor, a reduction in the required size of main memory would have a significant impact on the design. This impact is evident in the Category III design.
Figure 6.6 shows the total bandwidth available as a function
of chip count. Note the dotted vertical line that shows the bandwidth
available at the pins of memory chip in 1994. The figure also
indicates low and high values for the internal bandwidth of a memory
chip, which measures the number of bits activated during each memory
cycle. In typical memory chips, from 1 to 8 bits are transmitted
off-chip for each memory access. Memory itself is laid out in two
dimensions, with rows and columns of approximately equal size. An
access selects one entire row of the chip, and reads this row into
buffers. From the buffers, the access mechanism selects the 1 to 8
bits to be sent off-chip. Since rows and columns are of equal size, a
typical -bit chip is organized into an array that is
, and thus
bits are accessed for each group
of up to 8 bits sent off chip. In 1994, memory sizes have reached 16
Mbits per chip, so that 4096 bits are accessed for each set of up to 8
bits sent off chip. Because for some chips only a single bit is sent
off chip per access, from 500 to 4000 times the internal bandwidth is
discarded. Category III machines attempt to gain performance by using
the internal bandwidth to drive processors colocated with the memory
chips.
Cooling and external bandwidth per chip are both important factors in devices, and have a major impact on device counts for the PetaFLOPS machine. Figures 6.7 and 6.8 show the SIA projections through 2007 and the panel's projections beyond that.
The SIA projections and the extrapolations through 2014 indicate that computer architecture as we know it today can support the design and construction of a PetaFLOPS machine in 2014. Technology in the next 20 years will bring a substantial decrease in cost and risk in the design and construction of such a machine. It is possible to build a PetaFLOPS machine today, but with extraordinary cost-perhaps in the hundreds of billions of dollars-and at great risk of failure because of the enormous number of parts required. As projected costs drop (by a factor of 10 every five years), the cost of a petacomputer drops to tens of millions in 20 years, well within the limits of supercomputers today. Also, the parts count in 2014 reduces to parts counts achieved today in large-scale machines.