***3D 1997 - A Look Forward
by John Latta
3D - Points to Ponder
Is 1997 a "3D Year?" In spite of our enthusiasm for the technology we
increasingly expect that 1997 will be another year where the supporting
3D technology is too little too late to have a major impact on Q4. Why?
Making 3D happen is all about creating a 3D infrastructure (see the
illustration at http://www.fourthwave.com/wave/infrastruct.html) and
creating infrastructures is a slow process. Here are the factors which we
see gating the 3D marketplace in 1997 and beyond.
In spite 3D's heritage from flight simulators going back 20+ years and
more recently SGI workstations, the only real market of import is the
mass market. It is here, where the desktops and home PC media appliances
incorporate 3D, that the technology will have the greatest impact. With
impact comes the prospect for being a growth company who supplies 3D
hardware or software in the next 5 years. PCs will compete with next
generation game platforms and the "war for the eyeballs" as Andy Grove of
Intel described it in his COMDEX keynote. Just as HP found it could not
be a major PC company without selling consumer PCs 3D chip companies will
find that they cannot have a major role in the market without gaining the
economies of scale and economic access to state-of-the-art fab that large
unit volumes enable. Being successful in 3D is about delivering the best
technology, the best experience and new applications in 10's of millions
of units. 3D has close parallels with the development of the
microprocessor based PC industry in the early 1980's. Delivering high
quality real time 3D is a price performance game. Being a niche player is
the same as being a small player - just ask Apple.
With software rendering now having achieved at least 20f/s, hardware
acceleration is only of value if it brings a significant improvement in
image quality. There is no consumer market pull for 3D which is a "little
better." This means at least bi-linear perspective correct texture mapped
images at 30f/s in a 640 X 480 window - minimum. "Free D" does not cut
it. Turning filtering features off to pump up the pixel fill rate is a
losing game. Although frame rate has been an important parameter to judge
performance this will rapidly fade as anything less than real time
(30f/s) will be judged as a loser. Today the best measure of image
quality is correlated with pixel fill rate and this will remain for some
time to come.
The gating performance metrics in 3D will evolve in three steps. First,
the leading chips will be those which excel in pixel fill rate. 3Dfx has
already set the market threshold at 50+Mp/sec and we expect this will to
continue to rise. This parameter is basically a rasteriser performance
metric based on the bandwidth to the texture memory. Next, the market
will shift to polygon rate. There are a number of reasons for this and
the primary one being that the number of polygons per frame determines
the number of objects in the scene. (We are implicitly assuming that the
fill rate continues to rise so that the depth complexity is not
compromised as the polygon rate also rises.) Another related reason is
that content designers will put an emphasis on humans in their polygon
budgets. A credible human takes 5,000 polygons. Thus, if you mingle with
4 of your closest friends, assuming the scene takes 5,000 polygons, this
results in a requirement to process 25,000 polygons at the frame rate or
1.5M polygons/sec at 60 Hz. The third transition will occur at the level
of 1 - 2 M polygons per second when there are "excess" polygons (most
will say there are never enough polygons). At this level, explicit
quality features will enable those with the best and most to become
market discriminators. To some extent we saw this in the E&S Harmony
video tape at COMDEX. The SGI Infinite Reality also fits into this class.
Quality features include anti-alaising, lighting, bump mapping and
filtering beyond tri-linear MIP Mapping. The realization of this
performance is not one of understanding the core technology but of design
and process technology - getting the best performance and the most
features in the least gates with the best process technology available to
the 3D chip company.
Geometry Engines will play an increasingly important role in 3D chip
design. The reasons are three fold: flexibility, features and
performance. As the emphasis shifts to polygon performance and lighting
along with flexible API compliance watch for programmable geometry
engines to play an increasingly important role in future chips. In
addition, the more sophisticated chips will provide feedback on pipeline
loading so that developers can optimize the scene content for maximum
performance.
State-of-the-art 3D chip technology will continue to enter the market at
the high end - at end user price points => $1,000. However, in spite of
being a technology entry point this is still a limited and narrow market
(in terms of unit volume). For companies to grow and gain market share
they must move down market in what is already a brutal battlefield.
However, just as SGI has found that high priced systems are the means to
define the 3D workstation market with state-of-the-art capabilities it is
also the markets which have the gross margins sufficient to cover the
high development costs. In the PC space companies will participate in the
high end for many of the same reasons - 3DLabs is a good case in point.
The PC cost metrics also give these same companies an excellent lever by
which to compete with SGI. The PC will offer performance comparable to an
SGI system but at 1/10 the cost. Again 3Dfx is playing to this strategy,
at least in the game space. A critical factor in the development of
performance driven 3D, including the high end, is that segmentation by
APIs will disappear. Only the pristine will have a single API solution,
be it Direct 3D or OpenGL.
Integrating state-of-the-art 2D with 3D onto an accelerator chip is a
given. Better think of it as a display processor. Pass thru cables have
gone the way of the dodo bird, NSP and UNIX.
Texture compression and data management will have an increasing role in
next generation chips. The effective use of texture is the most efficient
means to increase scene content complexity and realism. Texture memory
sizes will rapidly reach 10MB+. As a result of the continuing demand on
performance, as outlined above, only local texture memory will be an
adequate solution - even with AGP. Thus, texture compression and
management of texture storage will be critical to keep system costs down
and continue the upward spiral of performance at mass market prices.
An issue largely unaddressed by the emerging 3D hardware industry is
platform performance optimization. Intel is very sensitive of the need to
have an integrated design (which they call optimized) which is a lowest
cost solution for a mass market platform. Yet, there are a number of
factors which Intel does not control including the add-in card market. In
many respects platform optimization is similar to L1 and L2 cache
integration into the microprocessor, the core logic chip set and the
platform itself. Yet, what 3D brings is the need to optimize the platform
in much the same way that multimedia does - nearly all system components
from the OS to the I/O subsystem to the memory design to the driver to
the 3D chip and the microprocessor influence 3D performance. Although AGP
is an attempt to raise the bandwidth to main memory for texture storage
at 500MB/sec even this rate is inadequate for many of today's 3D chip
designs. Tuning the PC platform for optimum 3D performance could make
Plug and Play look like a cake walk. But being able to continue to push
the 3D performance envelope with a PC system design is critical if the PC
is to compete against SGI at the high end and the home video console at
the low end.
Benchmarking is one of the most difficult issues which the 3D industry
faces, aside from the chip design challenges and creating compelling
software applications. Benchmarking will split into two segments:
application level benchmarks and performance characterization. Although
application level benchmarks will focus on system level performance
issues they will not address the image quality issue. Two systems could
have virtually the same performance but different output quality levels
which could result in significantly different user experiences. In
performance characterization there are a number of techniques to measure
isolated parameters such as pixel fill rate, polygon rate and frame rate
and early benchmarking tests will quantify these. However, as the
industry transitions to the 20,000+ polygon/frame performance level,
image quality metrics will become increasingly important. The fundamental
issue is - how is image quality measured in video and especially where
multi-frame quality impacts are to be assessed? We believe that image
quality will only be evaluated subjectively. However, we expect that
techniques will evolve which quantify typical 3D image generation quality
artifacts. These will then be categorized, ranked and rated in terms of
their impact on the perceive video image. With such tools being used in
side-by-side comparisons by experienced observers relative image quality
measures can be assigned to 3D output. Comprehensive and fair benchmarks
are critical to the industry. If 2D accelerators are any indicator we
expect that winners in 3D benchmarks will be correlated with 3D market
leaders. Sales will rise and fall based on benchmarking results.
Benchmarking is a high stakes game and if done well could cause the bar
of 3D performance to rise continually as vendors release new chips to
better the competition.
Stimulating developer attention (commitment?) with 3D hardware spec
sheets is a vacuous proposition. To an application or game developer the
only thing which counts is having running 3D accelerated hardware. Spec
sheets are a dime a dozen. We also see where chips get "binned" within
quality bands (today's bands are based on pixel fill rate as discussed
above) and those who break out of those bands with significant price
performance advantages get developer attention. 3Dfx has shown that its
superior fill rate allowed it to draw developers and, as a result, games
were ported to its Voodoo chip.
3D is a chicken and egg market. Without an installed base of 3D
accelerators developers will not invest to create content or applications
which uses them. (There will be a few pioneers with early development
titles, however, most of them will be paid for their efforts.) But more
importantly, developers want only one or two APIs or accelerators to
write to and not a proliferation of 3D interfaces, be they hardware or
software. The key operatives are installed base, stability and minimum
performance level. In this respect 3D closely parallels the development
of the early CD-ROM and multimedia title market.
There is more to 3D software than games, CAD, 3D GIS and visualization.
The problem is that no one has a clue as to what. It may be a new user
interface, personal shared environments or interactive movies. Just as
Visicalc and 123 created market pull and Doom did the same as a game
title, the same can happen with a 3D title or application which exploits
acceleration. We see the opportunities for 3D software as being a wide
open playing field. We believe the challenge is clear - create compelling
3D and they will come.
With better 3D acceleration we expect that much of the content will begin
to look more like traditional flight simulator data bases which are
optimized for visualization. In the case of games, data bases would
contain scene and object descriptions, animation scripts, audio content
and events which key the audio, environment physics and object dynamics.
Just one problem - it is too difficult and time consuming to create these
data bases. The barriers to making compelling 3D content are too high.
Presently the best tool for real time data base creation is Multigen but
it is too complex for a large market or one which is production oriented.
RealiMation by Datapath is a significant step forward in database editing
and viewing. In the future 3D databases can come from other 3D
information sources such as CAD, but sophisticated simplification
technology is required as SGI showed at the recent Autofact show. The
types of information and level of detail required for object
visualization is substantially different than that required to design,
manufacture or build the item be it an alternator or building. Think of
3D data base and display technology as the portal to the "real"
information about that same item through the use of a visual metaphor.
Multiple types of 3D content exist which describes objects but these,
from the standpoint of visualization and access to the object
information, are, in most cases, incoherent and incompatible. Thus, the
potential for real time 3D technology to enable the seamless exploration
of 3D content, where the content comes from many sources based on its
purpose and application, is unrealized. To visualize the problem one only
has to think of the all digital design of the Boeing 777.
Closely related to data base creation is object behaviors and
intelligence. Today the emphasis is on having libraries of objects which
are either in polygonal form or can be reduced to polygons. What is
needed is a modular method of adding object attributes and behaviors
based on the use for the model, object or character. In fact, the natural
progression of the technology is that object attributes will change with
time and the conditions under which the object is within the scene.
Deformation characteristics are only required when needed, for example.
We see object attributes as an area largely untouched. An elegant
solution will make it easier to create environments, populate them with
objects, have useful interactions between objects in an environment and
implement participant and equipment interfaces from the real to the
virtual. A key problem is that a proprietary solution is not what the
industry needs to build a 3D infrastructure.
The union of networking, 3D, and a social experience is a virtual
community. We find today's implementation lacking for a number of
reasons. First, given the reliance on software rendering at the client
the images of avatars look silly. Not only must there be a better image
representations of humans but of human attributes as well. Communities
have social structures and a culture, however, in order to bridge the gap
between real (as we experience in our normal lives) and virtual (as
experienced in the virtual community) the participants must use fantasy
to imagine what the virtual environment represents. For most practical
applications adults find fantasy unappealing in the conduct of work and
many of life's functions. Finally, our review of virtual communities
finds that unique applications - short of games - are hard to isolate.
Thus, virtual communities are another combination of technologies in
search of a market.
Animation, or past time 3D, is for movies and television productions.
Only real time 3D counts on the PC. In many respects 3D is just another
media form which allows the participant to see, experience and interact
in scenes in real time. 3D is a form of multimedia and its potential
comes when there is a seamless blend between 3D and high quality video.
The genre of interactive movies will come on its own when real time
objects and characters participate within a DVD video playback and are,
for all practical purposes, unrecognizable as coming from a 3D image
generator. Thus, the interactive movie can be different each time and the
characters can change based on the plot and unfolding events in the
movie. Consider an interactive movie with Myst quality, mysterious
characters that change with every game play, characters who respond to
player actions and where exploration of the environment is done, in part,
with user controlled flashlights. The PC has the potential to reach a new
plateau of interactive entertainment. Part of this plateau will be
reached when the boundary between real (video) and virtual (real time 3D)
is difficult to discern. Unfortunately, this makes Wing Commander IV look
like a simple exercise.
This is all about creating a 3D infrastructure. Unfortunately, in spite
of many attempts at parallel evolution of components in the
infrastructure it seems to go one step at a time.
What does 1997 hold for 3D? Come share our crystal ball.
Threshold for acceptable 3D will be 55Mp/sec - minimum
Just as the 3Dfx Voodoo, with its 50+Mpixel/sec fill rate, set the
standard as a target for game content that level of performance will
remain as the minimum in 1997, only more companies will achieve it. The
infamous "free 3D" decelerator chips will fade into oblivion - nice try
but largely ineffectual in impacting the market. Market demand was absent
and content happened when developers were paid.
AGP will have only a minimal market impact
AGP hardware, Microsoft's essential software, both the OS and DirectX
5.0, and applications, including games, will be too late for AGP to have
much of an impact on the Christmas buying season. Increasingly AGP is
looking like MMX all over again. Look to 1998.
3D API competition will continue
The increasing role of OpenGL as a cross platform API will make it all
the more difficult for Direct3D to establish the dominant position on the
Windows platform. The expected shipment of Direct X 5.0 in June only
limits the ability of Direct3D to impact Christmas of 1997 because it
comes too late in the calendar year.
Software will continue to significantly lag chips and hardware
Without a large base of installed 3D chips or an universally accepted API
software developers will hold back from the market. This only retards the
introduction of cutting edge high quality 3D applications which leverage
acceleration.
VRML as an industry effort will implode
Without a strong consortium headed by a recognized industry leader VRML
will be just another file format with a diminishing contribution. There
are too many factions with strong vested interests to see VRML become the
force it should and could be.
Emergence of one significant accelerated 3D application
We have no idea what the application will be but we know its attributes:
priced under $50, usable by any consumer, not necessarily a game, fun,
and new. Today the best selling 3D applications are home design programs
for consumers. This may or may not be the seed from which the 3D
application will sprout. A 3D killer app which presses acceleration
technology is what the industry desperately needs.
Technology will be announced which will threaten SGI's dominance of 3D on
the Workstation
At the high end of the market winners will be made by attacking SGI with
a significant price-performance advantage. In 1997 it will be clear that
SGI's 3D performance advantage over other workstations and PCs will begin
to erode. The technology may not ship but the threat will be more evident
than ever.
3D on Portables will not make it but few will care
Already there is industry buzz about AGP on portables and desktop like 3D
performance on mobile computers. Portables are about e-mail and
productivity, not games.
3D Games will remain the software category which shows cutting-edge 3D
performance
Just as in 1996, the number of 3D game titles was significantly less than
promised. We expect that 1997 will be repeated - high initial
expectations with less delivery. Although we predict that no more than 10
titles will effectively push the envelope of accelerated 3D these titles
will show the potential of 3D as an entertainment medium on the PC which
directly competes with home video consoles.
3D will become an intellectual property minefield
Little has been said about the on-going legal battle between Lockheed
Martin and Silicon Graphics over 3D intellectual property (IP). Given
that much of the underlying technology for 3D image generation is 20+
years old it is not surprising that companies will lay stake in what they
feel is their own and where there is a market potential from which
royalties can be extracted. We expect that this will be continuing mine
field which will blow up on the unsuspecting. In 1997 many IP issues will
surface. Let the 3D developer beware.
Wave Issue 9617 1/3/97 Article 2-01