AMD and NVIDIA document ping-pong
Author: Vedran Dakic
Date: 14 May 2007

So, here's the real deal - I came across some pretty unnecessary documents that AMD and NVIDIA have been sending to the press over the weekend and decided to write something a bit funny and poetic about this. I mean, I appericiate and respect both companies' product but sometimes they just indulge themselves in this PRware that's completely unnecessary. It's perhaps a bit funny and this is how it's ment so take your cup of coffee and check this one out...

Quote, NVIDIA document:
"Radeon HD 2900 has 320 shaders but GeForce 8800 GTX has only 128 shaders. What?s the deal?
AMD counts their standard ALU and special-function ALUs to reach 320. GeForce 8800 GTX has 128 standard ALUs and 128 special function units. By this method of counting, GeForce 8800 GTX has 256. So Radeon HD 2900 has 25% more shader ALUs, but GeForce 8800 GTX ALUs are clocked 82% faster and 8800 GTS ALUs are clocked 62% faster. The GeForce 8800 GTX is faster overall. The 8800 GTS and Radeon HD 2900 are comparable."

At this particular time, we're really reminded of Mark Twain's (1835-1910) words we all liked so much when we were little kids. Check these out:
?You?re a coward and a pup. I?ll tell my big brother on you, and he can thrash you with his little finger, and I?ll make him do it, too.?
?What do I care for your big brother? I?ve got a brother that?s bigger than he is ? and what?s more, he can throw him over that fence, too.?
(c) by Mark Twain, The Adventures of Tom Sawyer, 1876.

Quote, AMD document:
'?By this method of counting, GeForce 8800 GTX has 256.?
Interesting that the characterization of GeForce 8800 GTX has all of a sudden changed after 6 months of telling people that its 128 units ? why change it now based on an incorrect understanding of a competing architecture?
While we won?t try and teach others about an architecture we didn?t build, we can certainly point to performance testing that shows that we can achieve instruction rates incredibly close to the peak rate expected for both general purpose operations and special function operations. Testing on G80 certainly doesn?t give us any indicator that their special function performance is close to how they have decided to characterize it here':(snip....)

In the end, maybe it's the immortal F.D.R. words we're really remembering, so nicely quoted in one of the Dawson's Creek episodes (Season 3, Episode 17, Cinderella Story):
"The only thing we have to fear is fear itself."

Quote, first NVIDIA and then AMD document.
"2900 uses a superscalar architecture with VLIW instructions. Which shader architecture is better?
First it should be noted that NVIDIA introduced a superscalar architecture many years ago with the GeForce 6. And the generation before, we designed a VLIW (Very Long Instruction Word) architecture with the GeForce FX. Both of these architectures have their respective advantages and disadvantages, but it was through this experience that we designed the GeForce 8 as a scalar architecture.
(snip)
In essence, for the Radeon HD 2900 to sustain peak performance, the driver needs to read in a stream of shader code and output five independent instructions at all times. If the driver could only extract two instructions per clock, performance will be reduced by three-fifths.
The GeForce 8800?s scalar architecture does not face this problem. The driver needs to only extract two instructions per clock (one for the ALU, one of the special function unit) in order to the keep the shader unit fully utilized. As a result, the GeForce 8800 provides much better real-world performance."(End of NVIDIA doc quote)

'?In essence, for the Radeon HD 2900 to sustain peak performance, the driver needs to read in a stream of shader code and output five independent instructions at all times. If the driver could only extract two instructions per clock, performance will be reduced by three-fifths.?
Radeon HD 2900 XT?s shader compiler is already incredibly effective at its job and it and the hardware is capable of scheduling even vector instructions across the Stream Processors either in parallel or serial, allowing for a very flexible instruction issuing in order to extract the maximum performance out of the compute core of the ASIC.'(End of AMD doc quote)

Nevertheless, this ping-pong-mumbo-jumbo reminded us of something else, as well. We did read quite a lot of novels when we were younger, you know? How about Oscar Wilde's (1854-1900) Picture of Dorian Gray?
?Have you really a very bad influence, Lord Henry? As bad as Basil says??
?There is no such thing as a good influence, Mr. Gray. All influence is immoral,?immoral from the scientific point of view.?
?Why??
?Because to influence a person is to give him one?s own soul. He does not think his natural thoughts, or burn with his natural passions. His virtues are not real to him. His sins, if there are such things as sins, are borrowed. He becomes an echo of some one else?s music, an actor of a part that has not been written for him. The aim of life is self-development. To realize one?s nature perfectly,?that is what each of us is here for. People are afraid of themselves, nowadays. They have forgotten the highest of all duties, the duty that one owes to one?s self.'
(c) by Oscar Wilde, The Picture of Dorian Gray, 1891.

Here's one thing that we actually agree upon (the bold part):
AMD states that they are 50x faster in their geometry shader test.
Different geometry tests will yield different results. We have a geometry shader test called ?metaballs.? You can download it from:
here. To run the test under Vista / DX10, execute the Metaballs application in the /bin directory. Move the Grid Resolution slider all the way to the right. Turn off vsync and AA through the "Change Device" button, and ensure maximum number of metaballs (100) is selected with "Add metaball" button. Set to "GS single-pass" mode. Note the frame rate in upper left corner. You should see the GeForce 8800 is much faster than Radeon HD 2900. In the end, only real world games are a valid benchmark for comparing such features.

The conclusion? It's very simple. Please, take care of your own product, develop drivers, additional features, better picture quality, better game performance and don't waste your time with these documents. They are completely unnecessary and, in the end, there are only two things that mainly decide what's right and wrong in this crazy hardware world - testing results and market share/acceptance of new products.

But if we do some deducing on the Oscar Wilde quote, we might have a hipocrisy-phase, as well. If influence is immoral, then what's press doing? Work of the devil himself? :-) Well, we're pretty sure we don't. But that's up to you to decide...