JANUARY 2007 MAXIMUMPC 31
PHYSICS PROCESSING
Nvidia and ATI have both been touting
the idea of using their GPUs to acceler-
ate physics computations. We’ve pooh-
poohed the concept in the past, because
it seemed as though the graphics compa-
nies were simply reaching for an excuse
to sell more parts. If Nvidia’s Quantum
Effects technology takes off, we might
have to change our tune.
David Kirk, Nvidia’s chief scientist,
justifi ed our initial skepticism during the
company’s GeForce 8800 press briefi ng
with his comment that “...previously, you
could do only part of the physics process-
ing on the GPU because the underlying
hardware wasn’t cut out for it.” But the
parallel-processor nature of Nvidia’s new
architecture renders it much more adapt-
able for physics processing; in fact, the
company maintains that it has baked
some features into the chip that are dedi-
cated to non-graphics processing.
As Kirk explained it, CPUs are not
well-suited to physics processing because
they can solve only one equation at a time
and they execute most instructions out
of their cache. This requires a large and
complex set of control logic to keep the
CPU busy. The unifi ed shader units in the
8800 series can share data via a parallel
data cache without even having to use
video memory. This is a major shift from
Nvidia’s previous position of “We can do
physics processing, too.” With the 8800,
they’re saying “We’ve architected this new
GPU to do a very good job of physics
processing.”
It will be diffi cult to measure Nvidia’s
claims in the absence of good bench-
marking tools—not to mention games that
make suffi cient use of physics to really
matter—but you’ll fi nd the results of our
hands-on testing of the 8800 GTX on the
next page. Turn to our reviews section for
more in-depth results.This block diagram illustrates the highly parallel nature of the
GeForce 8800’s architecture. Instructions from the host CPU are
fed to an input assembler, which issues a vertex thread to be pro-
cessed by the shader core. Once a vertex instruction has been
executed, the thread processor sends the result on either for geom-
etry processing (where entirely new vertices might be created—ordestroyed) or for pixel-shading.
The GPU can also temporarily store computational results in L
or L2 cache and then route them back to the top of the shader core,
where they’ll be repeatedly dispatched, processed, and re-looped
until all shader operations have been completed and the resulting
pixel fragment is sent on to the rasterizer.INSIDE THE GEFORCE 8800 GTX
SP SPLFB FB FB FB FB FBL1 L1 L1 L1 L1 LSP SPLSP SPLSP SPLSP SPLSP SPLSP SPLSP SPLThread ProcessorVtx Thread Issue Geom Thread Issue Pixel Thread IssueVtx Thread Issue Pixel Thread IssueVtx Thread IssueHOW WILL ATI AND
AGEIA RESPOND?We know that ATI’s next graphics processor will fea-
ture a unifi ed shader architecture because the com-
pany publicly said so way back in April 2006. ATI has
also made a lot of noise about its stream-processing
initiative, which harnesses the multiple multithreaded
cores in its GPUs to run distributed-computing appli-
cations such as Folding@Home.
Unfortunately, the only other information we’ve
been able to pry from ATI is that we can expect to
hear more about its new GPU—code-named R600—
around the time that Windows Vista is released. Our
hope is that AMD’s acquisition of ATI hasn’t been too
much of a distraction for the product-management
and engineering teams, and that they have tremen-
dously ambitious plans to outdo their arch rivals at
Nvidia. Anything less would undoubtedly slow the
advance of the state of the art, because no company
sees value in competing with itself.
It’s more diffi cult to predict how the 8800 series
will impact Ageia. Physics remains a wide-open
fi eld that game developers have yet to exploit. We
remain believers in the concept of building dedicated
hardware for such a resource-intensive task, but it’s
been more than six months since Ageia introduced
its PhysX PPU and we still don’t have the killer app to
justify buying a card equipped with one.