8800GTX and GTS specs in card data base

[MrWizard6600]

K, it's high time the stats changed on the 8800 series cards. I'm not going to pretend I understand all the jargon related to video cards, (because everytime I do, someone says something that contradicts with what I thought was right). But the numbers you have posted are wrong, and they need fixing.

--edit: please anyone feel free to comment if i get any of these terms wrong--

Seems to me, that all pixel, shader, and vertex (and geometry? is that mearly point-collisions (IE physics) calculations?) are handled by the 128, or 96, unified shaders on the card. As such, the traditional method of "multiply core clock with pipes" doesnt apply to any of the three.

So, what you should do, IMHO, is take the Stream Processers (SP)(IE unified shaders), and multiply them by the shader clock, which is 1.35GHz. Leave the resulting number either with a star, or an asterisk, or just the term "/3" after it. While it could produce these numbers, if it needed to, it would never have to, as some of the SPs would be diverted to other purposes. So while a 8800GTX could do 1.728e11 vertex calculations per sec, some of those would be pixel calculations, and some would be pixel calculations

If I've done this horribly wrong correct me, but I dont think I have, and the numbers need fixing.

[Mike]

That's actually a very viable solution.

My only issue is...what happens when R600 comes out? How will we state the specs then?

Before I go through the time and effort to change all the pages on the site to accommodate these new unified cards, I want to know that it will, at the very least, be able to compare G80 to R600 with some level of accuracy.

That, and I've been pretty busy with school. But yeah, something really needs to be done...when this semester ends, I'll hack something together, then do a major overhaul when R600 comes out.

[MrWizard6600]

yeah well the R600 is gonna just has 64 SPs right?

so, 64 X (w/e the clocks are) + your star, asterisk, "/3" or w/e.

Note: one problem, and this will always been a thorn in your side, is if GPUs start to develop something like Hyper threading, in that a second thread can start processing before the first is even finished, then we have a problem, you'll just have to accomdate your clock cycles accordingly

ex: if it exicutes 25% of the seccond thread, when the first thread is 100% exicuted, increase the clock speed you multiplied into the SPs by 25%.

[Track]

So ur saying that instead of doing this:

Stream Processors (128) x Core Clock Speed (575Mhz) = X

We do this:

Stream Processors (128) x Shader Clock Speed (1350Mhz) = X

And then we say /3 because the shaders are allocated and split apart to 3 tasks?


Well then, where does the Core Clock come into place when talking abt Shaders? Do the shaders now have their own 1.35Ghz clock, and the Textures their own personal one?
Im not sure abt this new architecture, and i dont know what this little shader chip is, but if its what is driving the shaders, then doesnt it perform the same action as a second core?

I agree with what u have said abt the HT technology, even though im pretty sure that is an Intel trademark and a similar technology wont be used in GPUs, and there is still a way to have more than one thread, wich is having more than one core.

Also, the R600 has 256 Stream Processors (as with current speculation, of course).

[MrWizard6600]

@track, i saw a lamens version of a diagram of the core, but i cant seem to find it.

the Sps are devided into groups, Firingsquads calling em banks. each bank consists of 16 unified shaders, an L1 cache, and some "texture adress units" which I dont know the exact funcion of. the 8800GTX has 8 of these banks, and the GTS has 2 disabled, making for 6.

The Vertex thread issue, Pixel thread issue, and Geometry thread issue, all run at the "core" speed. That is to say (i think) the shaders run at 1.35GHz, and the Rosterization runs at 575/500MHz.

before writing this post, I hadnt figured that out. That would explain the G80s tendancy to max out on one type of thread. IE if you increase the supposid "core" clock, you increase the cores abillity to rosterize things, and hence more threads of one type can be processed. Ok so, blagghh hard to put into non-jibberish. if the Core clock is 575MHz, you can only process 575 million pixel threads per second because thats the max ammount that can be given an adress (rosterized) per second. So, if you increase the core clock, you increase the cores flexibillity, as well as increasing rosterization times.

and where are you getting 256 sps from?

[Track]

Vertex, Pixel and Geometry. Vertex and Geometry are shader calculations and use the 1.35Ghz core, while the ROPs use the 575Mhz core, just like the textures.. Thats my guess.

So that means that the shaders are not only plentifull but are also clocked much higher - 128 (or 8 banks) x 1.35Ghz.

The R600 has 4 banks, each quad-pumped. Same as 16 banks, or 256 SPs.