***Benchmarking Fires
by John Latta
Our reporting on the Gemini 3D benchmarks generated a response, to say
the least.
Evans and Sutherland Responds
Rick Maule
Vice President and General Manager Desktop Graphics
Evans & Sutherland
The graphics market has been traditionally divided into two arenas:
design automation and like applications (for which OpenGL has emerged as
the current standard API) and simulation and game applications (for which
OpenGL has been seldom used). The reason is that the former typically
uses large numbers of small polygons and the latter, small numbers of
large polygons (an oversimplification to address this issue).
To address the design automation and OpenGL market, the E&S Freedom
Series hardware renders large numbers of small polygons at a very high
rate by utilizing the parallelism of the graphics processors. The gvf and
gvr are simulation benchmarks performed by Gemini that use a small number
of larger polygons, which does not take advantage of the Freedom Series
parallelism. The Freedom Series is better suited for OpenGL applications
which are more geometry-intensive.
When benchmark results are made public later this quarter, products based
on the REALimage technology (demonstrated by Evans & Sutherland at
SIGGRAPH) will combine the best of both these world's with performance
metrics comparable to the traditional professional UNIX workstation
products (such as SGI's High Impact) and the newest gaming boards (such
as 3Dfx).
www.gemtech.com
Real 3D Responds
The Real 3D Pro-1000 performance shown in the benchmarks is not fully
representative of its performance. Since the design is based on a double
frame buffer there is no need to have a frame rate above 60Hz. Thus, when
a benchmark is based on updating at a faster rate than what the eye can
see this has little meaning as a measure of performance, especially when
Image Quality plays such an important role and is not discussed in these
tests.
The impact of using features which dominate high quality image generators
was not fully reflected in the benchmarks. For example, it has been Real
3D's experience that features such as: anti-aliasing, large databases,
high depth complexity, special effects such as fog, microtexture, shading
and translucency when combined will place a significant load on a system.
The benchmarking puts little emphasis here. A more discernible test
would use a dense area of database with a large model count and high fill
rate. i.e. Any car can drive 60 mph down a straight level road, but
throw in some curves, hills and bad weather and most of those cars will
be forced to slow down.
Another important factor in evaluating the performance of systems is the
variation in update rate as a function of the scene. The Pro-1000
maintains a 60Hz update rate in the hardware by using advanced load
management algorithms. This is possible through the use of a proprietary
parallel architecture which can handle very complex database culling
structures. The load balancing of the geometry processing is also
handled automatically in the hardware. This design limits the load on
the host CPU to merely updating model positions where other architectures
often load the CPU down with translations, culling and clipping. A
further performance test would be to increase the number of models in the
scene until the hardware being tested drops to a frame rate below 60Hz.
The Pro-1000 uses a patent pending S-Buffer method of anti-aliasing which
provides the same image quality level as the "full scene" method
associated with SGI. Although image quality is not something that can be
readily quantified, by applying our patented tri-linear MIP-mapped
texture, anti-aliasing, shading, fogging, illumination, and translucency
to all the polygons, all the time, the Pro-1000 is able to produce an
image of the highest quality with no impact on performance.
www.real3d.com
Gemini Responds
John Archdeacon stood by the benchmarks as reasonable tests of 3D
performance. One area which could be made more evident - the 3Dfx cards
used 16 bit color depth while the others were full 24 bit. In addition,
there needs to be more education on the role and value of anti-aliasing.
In spite of the pressures for an Image Quality measure his response is
"If we knew how to measure it we would do it."
Gemini recently received the Power Storm 3D accelerator from DEC,
Permedia and 500TX cards from 3DLabs and an O2 (new Indy replacement)
from SGI. It is expected that benchmarking results will be reported by
the end of November.
Reflecting the need to measure CPU utilization accurately as part of the
benchmarks John sees a hole in Windows NT. In UNIX they can call system
level function to get the CPU utilization but see no equivalent function
in NT. If WAVE readers have suggestions you are encouraged to communicate
with John directly at jarch@gemtech.com.
www.gemtech.com
Wave Issue 9611 10/25/96 Article 6-01