: News
: Reviews
: Editorials

: Archives
: About
: Contact
: Advertising
: Privacy
: Links
: Forums

The Rising Cost of Graphics

Yet Another Glass Ceiling?

by Josh Walrath

The People

            At the beginning of the 3D revolution, we had two notable companies that were started in the early 90’s.  3Dfx was founded by Gary Tarolli, Scott Sellers, and Ross Smith.  NVIDIA was founded by Jen-Hsun Huang, Chris Malachowsky, and Curtis Priem.  The 3Dfx guys had a small office space in San Jose, while the folks at NVIDIA basically started out of a garage.  While each company had a slightly different vision how things should be, they all started out as a handful of guys designing basic 3D hardware.

            In the case of 3Dfx the SST-1 (Sellers, Smith, Tarolli-1) was a set of two chips, the raster chip and the texturing chip, each of which was comprised of around 1 million transistors.  These two chips made up the Voodoo Graphics.  It took these three guys (plus a few others) about two years to develop this product, and initially it didn’t look like it was going to fly because it needed so much expensive memory to work correctly.  Lucky for 3Dfx memory prices took a huge drop right before the product was to be released.  3Dfx was funded for around $10 million for the two years of development and the first production run, and later on received a few more monetary injections.  The combination of speed, luck, OpenGL and Direct3D support*, as well as a robust and strong Glide API made the Voodoo Graphics the top 3D card in the world.  By the time 3Dfx released this part, they had approximately 20 full time engineers working for the company.  Before the launch of the Voodoo 2, the engineering staff had grown to 40+ people.  During this time 3Dfx announced that they had a record quarter with $20 million in revenue (and an operating loss of several million).

            With NVIDIA things were a bit different.  The three engineers knew the pricing issues of memory at the time, and they had to find a way to make an affordable product that could actually compete in terms of performance to the other 3D parts coming down the line, yet undercut the competition in price.  To do this NVIDIA took a unique way to render a scene.  Instead of using polygons for rendering, they used quadratics.  This gave them the ability to create complex scenes using less memory than the competition, as well as give surfaces a more curved appearance (vs. the standard, blocky polygon models of the time).  This would then allow the NVIDIA product to use about one half the memory as the other upcoming parts.  The NV-1 also had the added feature of being a sound card.  Diamond was the only manufacturer that picked up the NV-1, and the expense of developing this as well as its unpopularity due to rendering issues with quadratics nearly killed off NVIDIA.  NVIDIA’s second product was a console part that was cancelled midway through (NV-2), but they were paid for some of the development they did on it.  Barely holding onto life, the engineers at NVIDIA developed the NV-3, or what would eventually become the Riva 128.  This product saved NVIDIA’s bacon, and allowed Jen-Hsun and the gang to expand the company due to the great success that the Riva 128 generated.

            Now that we know what the first generations of 3D graphics looked like from a transistor count and engineering perspective, let us see what today’s products and companies look like.  3Dfx turned into 3dfx, and then they died and were bought up by NVIDIA.  NVIDIA right now is probably the strongest graphics company in the world, though ATI is definitely not far behind.

            When NVIDIA bought 3dfx out, there were around 400 engineers working at the company.  They then made job offers to around 100 3dfx engineers.  Now in late 2005 there are approximately 700 full time engineers working at NVIDIA, not to mention the other support staff to keep the company running (secretaries, IT people, accounting, investor relations, management, etc.).  Gone are the days of three guys working out of a garage.  It is estimated that there are over 15,000 NVIDIA employees worldwide at this time.  Total estimated revenues for FY2006 for NVIDIA appear to be around $2.2 billion.  Compare that to 3Dfx’s 1996 revenues of approximately 40 million.

            This is not just about how many extra people NVIDIA hired, but also about the overall complexity of the latest products to hit the market.  The original Voodoo Graphics had a total of 2 million transistors in two chips running at 50 MHz.  The NVIDIA 7800 GTX 512 MB features the G70 GPU which is comprised of 302 million transistors running at 550 MHz.  151x the transistor count and 10x the clockspeed in only 9 years.  This is an exponential growth that is simply unmatched in the computer industry.

            This increase in complexity requires more engineers and better tools to work with.  End-users demand more features, speed, and quality from every generation of 3D product, and to meet that demand ATI and NVIDIA spend millions to hire and retain engineers, upgrade workstations and server farms to handle the complex designs, and buy new software tools that can handle the massive job of routing and placing millions of transistors and make a working product at the end of the process.  More money is then spent to utilize the latest fabrication processes, as well as create the complex masks associated with these huge designs.  One single design can easily eat up $100 million in basic costs for design and production, not to mention all the surrounding costs of running a large and complex business.  As Brian mentioned above, NVIDIA spent over $1 billion US in just R&D for their graphics products in the past three years.  ATI spends a comparable amount of money for their products as well.

            Developing such products costs money.  Lots of money.  To be able to stay in business and continue to make products such as these, the company needs to get money from somewhere.  These people don’t work for free, and certainly all the other physical products a company uses are not donated to them from somewhere else.  Sun does not give NVIDIA and ATI free servers, nor does Staples provide free copy paper and toner cartridges.  All of these things cost money, and that money has to come from somewhere.

            The other big question we have to ask is what are we actually getting for the extra money spent?  When the GeForce 2 Ultra was released it was selling for around $450.  At the time the competition was the GeForce 2, Voodoo 5 5500, and the original ATI Radeon.  Those other products were selling for around $250, but what did the Ultra offer over the other standard cards?  It certainly didn’t give anti-aliasing quality to match that of the Voodoo 5, and it was not significantly faster than the regular GeForce 2 (less than 15% in most applications).  The price difference and performance did not match up that well.  Sure, the user got 64 MB of memory with the Ultra, but did it really help out the card at that time?  Not really.

            We can look at our current situation and see a much larger difference.  The GeForce 6800 GT is priced around $289 average, but the overall difference in performance between this card and the new 7800 GTX 512 MB is actually huge.  Two 6800 GT’s can almost perform on par with the 7800 GTX 512.  When comparing each card singly there is nearly an 80% difference in performance, and there are extra features in the 7800 over the 6800 (such as Transparency AA).  Going up to the $330 range we find the 7800 GT’s, and it still takes two of those to overcome the performance of the 7800 GTX 512.  Even just taking the $50 jump from the 6800 GT to the 7800 GT results in some very tangible performance and feature benefits.  We are not talking a small jump in performance for a lot of money.  What we are now seeing are obvious performance benefits that more closely reflect price changes.

*OpenGL support for the Voodoo Graphics was achieved through a mini-GL driver that translated the OpenGL calls to Glide, 3Dfx’s native API for the Voodoo series.  Direct3D support was added after Voodoo Graphic’s introduction, and the first iteration of Microsoft’s 3D API was simply atrocious.

Next: Economies of Scale and Milestones

Copyright 1999-2005 PenStar Systems, LLC.