G71/G73: Summary & Pre-Launch Speculation Frenzy

Yup, pretty much. Nothing has changed from the 7800 to the 7900 in terms of shader strength, they both have 24 pixel shader processors with 2 ALUs per, only now nvidia has decided to throw in the 48 number (since 24 x 2 = 48) to make it seem like it's that much more powerful, when in fact it's just a G70 running at 655 MHz.

Welcome to nvidia's marketing department.

 

Where did you get that, please? I'm aware of exactly one credible source for that, and frankly that one struck me as an ATI rep having a "bad speculation day" post NV40 transistor count head-scratching.

 

Well, 24-pipelines was fine for marketing when they were up against 16 pipelines from ATi. But now that ATi has jumped to 48-pipes/alu's, marketing jumped all over G70's second ALU improvements.

There was a discussion between myself and andypski on this subject a couple days back. It is no more fair to compare G70's 24-pipes to R580's 16-pipes than it is to compare G70's 24-pipes to R580's 48-ALU's or G70's 48-ALU's to R580's 48-ALU's or G70's 48-ALUs to R580's 96. There's no perfect match there due to architectural differences and from what I've seen, what you consider more "right" depends on which side you're batting for.

Good thing none of that matters and it's the performance that counts in the end.

 

Phew.. I was a bit worried nvidia really did have 48 alus.
Nvidia marketing never ceases to amaze me.
First the NV30 had 8 pipes and now the G71 has 48 alus.

 

I'd like to know more about these magic drivers! Did they find a way to make better use of the mini ALU??? 20-30% is a huge jump.

 

All I know is that I would not want to count my balls using either ATI or Nvidia method.
With the latest buzzword being efficiency per mm^2, I would not want not to measure my johnson with Nvidia ruler. However, I benchmark my performance using the Nvidia method.

 

LMAO! Quote of the day!
________
Sexy girls

 

From our "More fun with stupid benchmark graphing tricks" file: http://www.theinquirer.net/?article=30089

 

Just a question from a noob on this forum (though not a noob in chip design): I've been following this thread for a long time now and every now and then this discussion about the amount of transistors comes up.

Why does it matter?

Because, honestly, in a 4 semiconductor company I've worked for, this has never been even mentionned... The only thing that counts is the size of the die and the percentage of memory of that die. Number of transistors is only used by the evil side (that's the marketing department for us, engineers) who use it to impress potential customers.

There are a number of ways in which you can exagerate the amount of transitors:
- lie about it knowingly. Hey, who's going to prove you wrong?
- lie about it unknowingly: no engineer is actually going to spend time counting them. Really. So all the numbers you read are based on extrapolations. "Let's say, x mm2 at a typical average of so many transitors per mm2, hmmm, that should be this amount of transistors." The fact that transistor densities for RAMs are way higher than those of random logic make it much easier to fudge numbers. Who's going to count the exact amount of mm2 of RAMs of the total die? Or the IO pads of a chip take a huge amount of area, even though they usually only contain only a few (huge) transistors. etc etc.
- optimistic counting. The last one is very easy to do: in current designs, there's a lot of redunancy, especially when it comes to RAMs. Say you put a number ram of 64 KB rams on your chip to implement your transmogrifier. In a 90nm process, there's no way you're *not* going to add redundant row and/or columns to increase the yields (dramatically.) If this adds an overhead of 10% for the RAM, but increases the yield by more than that, nobody is going to stop you from counting those extra 10% as real transistors, even if they're never going to do anything useful in addition.

In other words: nobody on the engineering side ever talks about the amount of transitors. So if Nvidia and ATI are using different numbers of same sizes of dies: it's all marketing. Just ignore it.

The parameters that *do* count are:
- die area
- % of area used for RAM. (On one side more is better, since it allows more redundancy. On the other side less is better, since densitity is higher which makes it hard to produce error free when there's no redudancy, which is the case when you're dealing with a lot of small RAMs.)
- Amount of logic redundany (typically, this is rather small.)
- cell density. This is mostly determined by the amount of wiring that interconnects that cells. An elegant design uses less wiring to to the same kind of stuff.

 

Well, by some measure, transistor count is more tightly-connected to the underlying architecture than die size, so it's somewhat nice to have the information for speculation on how architectures will scale over different numbers of pipelines, and how well they'll do at different processes.

But you're absolutely right: for the most part, transistor count is just a proxy for die size in these speculations. It is the die size that is the important bottom line. I think we mostly make use of transistor count because it's largely independent of process technology, and we're used to thinking in that way on these forums.

 

http://www.driverheaven.net/articles/efficiency/index.htm

Actually here it is. I could not recall it at first.

Disregarding opengl because of drivers, which isn't fair to Nvidia but anyway, in half life 2 the G70 is only 14% faster in the second case (less in the lower res).

And the 3Dmarks tests show decent parity as well if you ask me.

Sure Nvidia IS more powerful pipe for pipe clock for clock, but I wouldn't say by a lot. Which is sort of amazing.

 

Well, there are two things that can really throw a wrench into that logic.

First of all, you're making the assumption that those extra ALU units will be used. With only 16 texture units, it's quite possible that most games will be limited by those instead of the ALU's, though future games may change this dramatically.

Secondly, once games do obtain a higher ALU to TEX ratio, nVidia theoretically gains too: it can make better use of its first ALU, since nVidia makes use of the first ALU unit when a TEX op is called. So we're going to see how it all pans out.

 

Maybe ATI should come out with a slide "ALU for ALU, we are 150% more powerful!" or some such nonsense

Because by Nvidia's ALU definition, you would have to chop G70 to EIGHT pipes to make 16 ALU's vs R520's 16 "ALU's". I wonder who'd handily win that bench..

 

True but I'm just showing you an ATI "pipe" is probably 85% as good as an Nvidia pipe or something like that.

Sure that wont show in R580 currently though, tis true. But with the right game that's what's in store.

I think people need to stop the ALU nonsense and get back to pixel pipelines. ATI has two ALU's per pipeline just like Nvidia. To be true, ATI pretty much started it. The ALU's are SECONDARY to understanding the architectures, That ATI has a weaker second ALU can tell us why their pipes are weaker. But to call both together one ALU is silly to my understanding (which isn't much)

 

As Chris Goffey used to say on Top Gear whilst smalling the bonnet down :-

"Right, that's what's under the bonnet, lets see what she will do on the road ... "

 

This is getting silly.

Edit: I also wish these charts with axis that don't start from Zero died a horrible death.

 

Gowd, what's going on here? We started a new topic and it starts over with the same crap we threw out on the first pages of the old one? Hello?

I thought we were over with that 48/96 crap...

If anyone mentions full/half/almost-ALU's again, I'll take that post and beat him to death with it!

So any new benchmarks or anything substantial yet?

 

Exactly, we should be at the stage of the first few results leaking out now followed in a couple of days time with the normal mutterings from assembled web writers how XXXXXX always jumps the gun

Anyone kept an eye on the Far Eastern Web sites ? Galaxy have a promo banner on Gzeasy for 7900 and it mentions 3dmark but I cannot read the rest.