Graphics Reference
In-Depth Information
Like the Extreme CPU, the GTX GPU is a single component—a packaged
silicon
chip
—that is mounted to a circuit board by hundreds of soldered circuit
interconnections. The CPU, chipset, and CPU memory are mounted to the pri-
mary system circuit board, called the
motherboard,
along with most of the PC's
remaining electronics. The GPU and GPU memory are mounted to a separate
PCIe circuit board that is connected to the motherboard through a multiconnector
socket (see Figure 38.2).
Perhaps the most notable characteristic of the GeForce 9800 GTX is its
extremely high performance. The GTX can render 340 million small triangles per
second and, when rendering much larger triangles, a peak of almost 11 billion pix-
els per second. Executing application-specified code to compute the shaded col-
ors of rendered pixels, the GTX can perform 576 billion floating-point operations
per second (GFLOPS). This is more GFLOPS than the most powerful super-
computer available just over a decade ago, and is more than five times the
102.4 GFLOPS that the Intel Core 2 Extreme QX9770 CPU can sustain. Mem-
ory bandwidth—the rate data is moved between a processor chip and its exter-
nal random access memory—is an equally important metric of performance. The
Core 2 Extreme QX9770 CPU accesses external memory through the X48 Express
chipset. While the chipset supports memory transfers at up to 25.6 GB/s, the Front
Side Bus that connects the CPU to the chipset limits CPU-memory transfers to a
maximum of 12.8 GB/s. The GPU accesses its memory directly and achieves a
peak transfer rate of 70.4 GB/s, more than five times that available to the CPU.
In the same way that interest compounds on invested money, CPU and GPU
performance have reached their current state through steady exponential increase.
Underlying this exponential increase is Moore's Law, Gordon Moore's 1965 pre-
diction that the number of transistors on an economically optimum integrated cir-
cuit would continue to increase exponentially, as it had since Intel had begun
producing integrated circuits a few years earlier [Moo65]. Over the succeeding
four decades the actual increase held steady at about 50% per year, resulting in
the near-iconic status of Moore's prediction. Compounding at high rates quickly
results in huge gains. A decade of 50% annual compounding, for example, yields
an increase of 1.5
10
=
57.67 times. And a decade of 100% annual compounding
yields an increase of 2
10
=
1,024 times, so quantities with the same initial value
that grow at these two rates differ by a factor of almost 20 after a decade. Driven
by Moore's Law, the storage capacity of integrated circuit memory, which is pro-
portional to transistor count, has increased by a factor greater than ten million
since its first commercial availability in the early 1960s.
Increased transistor count allows increased circuit complexity, which engi-
neers parlay into increased performance through techniques such as
parallelism
(performing multiple operations simultaneously) and
caching
(keeping frequently
used data elements in small, high-speed memory near computational units).
Increases in transistor count are driven primarily by the steady reduction in the
dimensions of transistors and interconnections on silicon integrated circuits. (The
interconnections on the Intel Core 2 Extreme QX9770 CPU have a
drawn width
of just 45 nm, 1
Figure 38.2:
NVIDIA
GeForce
9800 GTX graphics card. (Cour-
tesy of NVIDIA.)
2,000th the width of a human hair, and about one-tenth the
wavelength of blue light.) Smaller transistors change state more quickly, and
shorter interconnections introduce less delay, so circuits can be run at higher
speeds. Increases in transistor count and circuit speed compound, allowing the
/
