NV30 FP performance statistics

I saw this post over at Beyond3D and thought you might interested.

Look at the posts by thepkrl in this thread: http://www.beyond3d.com/forum/viewtopic.php?t=5124

He’s done some very extensive performance and dependency testing on the various different fragment instructions.

It might be nice to cut-and-paste some of the numbers onto this board in case something happens to that thread, but I’m not sure of the netiquette of doing that

– Zeno

OT:

Just wanted to know if anyone here actually have this card(5800 ultra)? I still cant find anything except pre-order.

I disagree as far as that beeing OT.
Performance as far as fp (and perhaps the DX version of the same code) is important.

It should be put on a website, not a forum.

The guy called ‘reverend’ tells a interesting thing… How exactly those diagrams and speculations come from?

I would take this with a grain of salt.

My speculation is that NV40 is a single-pipe chip. It has a special quantum-device that links its bus with a dimension in which the framebuffer is already full with the rendering you asked for. It can connect to 31000 universes/sec so, this is the maximum framerate but expect lower performance. This new device gives the best with at least 300 rendering passes in which the performance is 300000x.

Sorry for the joke but really, those may just be numbers posted on a forum - unless someone other confirms them (and puts them on a more organized and readable form), I would take them as a very irrelevant information. BTW, having a faster or slower FP unit does not mean so much on overall performance, and it is what counts.
Really, excuse me for the joke.

Obli,

If you start looking at the different 3D benchmarks, and actually run the numbers, and read publicly available specs, as well as publicly available hints or semi-publicly-but-likely rumors, you can piece together something which might be a very good model of the truth. Who knows, the chip may internally work totally differently, but if the model models performance accurately (and it seems to do, according to the data he posted) then the model is valid.

Yes, you’re definetly right.
The model is wrong, but it works. Well, the idea is good enough. You remembered me a very important thing, thanks!

However, I’ll continue to take the posted thread as a not really-reliable source. The information provided by “thepkrl” looks useful even if I don’t know how he get those numbers. Looks like however, he’s knowing about what he’s speaking about. BTW, zephyr himself tells us the thread needs verification so, information may or may not be reliable, depending from your discretion.

The fact is that most of the thread is, from my point of view, just a discussion of how nv30 MAY work. The only doubt I have in this is that, if backengineering a graphic chip were so easy, sure other companies would already did it.

However, internet is a free place, and averyone is free to give its opinion. Here’s mine.

[This message has been edited by Obli (edited 04-11-2003).]

The fact is that most of the thread is, from my point of view, just a discussion of how nv30 MAY work. The only doubt I have in this is that, if backengineering a graphic chip were so easy, sure other companies would already did it.

There’s a difference between saying that Opcode X probably takes Y amount of time on NV30, and getting sufficient information to build an NV30. The information they are gathering through benchmarks, while not useful to nVidia’s compeditors, is useful and potentially valid to us as users of the hardware.