G70 Vs X1800 Efficiency

[trinibwoy]

I would not be so sure on that..going to 90 will give them sum head room on yeilds and allow for faster clocks, but at faster clocks you get more heat and more power drain...

Well, the same can be said for going from 130nm to 110nm which netted them faster clocks and more pipes - you weren't seeing too many Ultras doing 490 (530) on stock voltage. But somehow they managed to actually not increase heat and power drain - they even managed to go back to a single-slot cooler.

Let me put it this way. Put a dual-slot cooler on a GTX and crank the voltage. What kind of clocks might we see?

It seems to me that ATi has not made good use of the 90nm process when it comes to savings on power consumption. Look at the XL vs GT numbers for example - http://www.techreport.com/reviews/2005q4/radeon-x1000/index.x?pg=16

 

[Pete]

I'm probably just grumpy, stupid, or ignorant, but I don't like the way DH uses the term "efficiency".

In my mind, a high level of efficiency suggests that the resources of the GPU are not idle very much. However, it seems to me that DH is actually measuring how much work each GPU performs each cycle. Which is fine, but it isn't "efficiency".

Hmmm, you're right. Technically, they're testing IPC. If they had power draw info and clocked the two to draw the same power, I suppose they could claim an efficiency per watt test.

Then again, every site has already done that with its 7800GT vs. X1800XL tests.

 

[trinibwoy]

Let me put it this way. Put a dual-slot cooler on a GTX and crank the voltage. What kind of clocks might we see?

Which makes me wonder - is there some hard ceiling on clocks at 110nm?

 

[jb]

Well, the same can be said for going from 130nm to 110nm which netted them faster clocks and more pipes - you weren't seeing too many Ultras doing 490 (530) on stock voltage. But somehow they managed to actually not increase heat and power drain - they even managed to go back to a single-slot cooler.

It depends if the tweak the design to run on the 90 nm tech. Meaning they really did a great job at making it run cool and with less power usage at their current tech. Shrink it down to a smaller process does not mean it will be running just as optimial. I was not throwing cuation to the wind as power usage and heat profiles dont really change when you shrink the process. In fact it could be "worse" for them as they have a smaller die which a simular amount of heat has to be spread over = "hotter" cores.

It seems to me that ATi has not made good use of the 90nm process when it comes to savings on power consumption. Look at the XL vs GT numbers for example - http://www.techreport.com/reviews/2005q4/radeon-x1000/index.x?pg=16

I think its "better" to compare them to a 850 xt as they have simular specs and its not bad compared to the 850 xt. Dont forget that TR power is over the whole card...so memory useage is also using power...faster memory = more power drain which has nothing to do with the R520 power usage.....

 

[FrameBuffer]

I think its "better" to compare them to a 850 xt as they have simular specs and its not bad compared to the 850 xt. Dont forget that TR power is over the whole card...so memory useage is also using power...faster memory = more power drain which has nothing to do with the R520 power usage.....

If this is the case, then wouldnt the x1800 XL (256MB) underclocked vs the 7800 GTX OC'd with only 16 Pixel Pipelines be a better comparison as opposed to DH's configuration ? Would such a test be a better litmus as to equate performance/watt ??

 

[Bouncing Zabaglione Bros.]

What's your point? The G70 can't hit 600Mhz+ just like the R520 doesn't have more than 16 pipes. You're conveniently ignoring what I pointed out earlier, that the only siginificant difference in theoretical numbers between DH's setup and retail configurations is the memory bandwidth advantage of the XT.

That the whole exercise is pointless when you have to change the basic states of the products. "Car A is as fast as Car B when we drive Car B more slowly".

 

[jb]

If this is the case, then wouldnt the x1800 XL (256MB) underclocked vs the 7800 GTX OC'd with only 16 Pixel Pipelines be a better comparison as opposed to DH's configuration ? Would such a test be a better litmus as to equate performance/watt ??

Supose soo... but who know what cards they had to pick and choose from at the time. Again trying to get everything the same is hard to do (same memory specs/speed, core speeds, ect)...

 

[Pete]

jb, the XL has the same speed and amount of RAM as the GT, and TR reports it draws 30W more "idle." That's a lot more, especially given the smaller process.

OTOH, they're much closer under load, with the XL only drawing ~5W more than the GT.

 

[trinibwoy]

That the whole exercise is pointless when you have to change the basic states of the products. "Car A is as fast as Car B when we drive Car B more slowly".

Horrible analogy! Anyway, you have a point but the exercise isn't completely "pointless". It highlighted for me once again the boost that extra 25% of memory bandwidth could be giving the XT.

But as other people have said, we're asking for something we already have. The perfect comparision is right here for us to see XL vs GT. Identical fillrate and memory bandwidth. And in games where the XT puts the smack down on the GTX - the XL just matches the GT.

http://www.xbitlabs.com/articles/video/display/radeon-x1800_9.html

 

[Tweaker]

Now I missed probably something but GTX pipeline has 2*ALU+miniALUs vs 520 1*ALU+miniALU i.e 100% more ALU power. They perform roughly same(except D3) yet GTX is more efficient? Of course nV ALU!= ATI ALU but still..

 

[jb]

jb, the XL has the same speed and amount of RAM as the GT, and TR reports it draws 30W more "idle." That's a lot more, especially given the smaller process.

Pete,
I ment to look at the two ATI cards to see how much it helped ATI going from the 850 to the 1800xl. And compare those numbers to see how much ATI's 90 process is doing as it was question how much that was helping ATI. Then realize that the 1800xl has almost x2 the number of transisters as the 850xt and still stays close to the power envelope its not bad...And I wanted to point out that TR numbers cover the whole card so things like memory and everything on the board (fan, ect) will effect the over all power drain.

I realize I did not state those and my post was confusing...I was trying to multi-tasking and guess maybe I should not do that when posting

 

[Maintank]

jb, the XL has the same speed and amount of RAM as the GT, and TR reports it draws 30W more "idle." That's a lot more, especially given the smaller process.

OTOH, they're much closer under load, with the XL only drawing ~5W more than the GT.

I dont believe ATI down clocks their core when in 2d do they?

 

[Moloch]

I dont believe ATI down clocks their core when in 2d do they?

Nope and in some review I read ati will start doing that in a future driver.

 

[Dave Baumann]

Which makes me wonder - is there some hard ceiling on clocks at 110nm?

Ceiling is probably more on the voltages. Check waht TSMC are saying about the rates for 110nm and what NVIDIA are already running at.

jb, the XL has the same speed and amount of RAM as the GT, and TR reports it draws 30W more "idle." That's a lot more, especially given the smaller process.

Most likely because NVIDIA are dropping the clocks a fair bit in 2D mode which ATI don't do yet. Terry mentioned this as a feature coming, and I presume that they will also take advantage of also lowering the voltages (like Intel's SpeedStep).

 

[Tokelil]

there's like a 25% difference in (3dmark) single texture fillrate between r520->g70?

(edit: in the DH tests.. 3432.2 MTexels/s vs 4288.1 MTexels/s)

This was what made me go as well. As far as I have read no one else has noted on this.
What is the explenation for such a big difference between the two?

Most likely because NVIDIA are dropping the clocks a fair bit in 2D mode which ATI don't do yet. Terry mentioned this as a feature coming, and I presume that they will also take advantage of also lowering the voltages (like Intel's SpeedStep).

Isn't a higher leak current expected when going to 90nm?

Edit:

Of course this is all theoretical as cards in the configurations tested don’t exist in retail channels however it’s an interesting aspect to take into consideration when choosing which card you wish to purchase.

Why they claim this info will mean anything for ones buying decision is beyond me...

 

[banksie]

The other thing that makes this test somewhat premature is that the ATi drivers are an unknown quantity.

We know for the R420 series of chips that they were very sensitive to clock frequencies and the memory caching patterns took a while to be tweaked to suit the frequencies the cards run at. Assuming even a mild amount of this has occured for the R520 then clocking down so drasticly may well be throwing these tweaks well off.

There also is the question as to how much optimisation time ATi has had with the R520 at full clock speeds - the implication thus far is not a lot.

This is where a good theoretical understanding of the chip and it's design is far more useful and why I like Beyond3D's reviews. Knowing the design well enough to see areas where drivers have room to improve is far more useful than a 'clock for clock' comparison between two quite different architectures.

 

[Sxotty]

Perhaps the option in RivaTuner doesn't work so it's actually still running with all pipes?

But I agree with OpenGL guy, this is not a particularly useful test. If we'd clock down a P4 to Athlon level, and find they run extremely slow in comparison, does that mean the P4 architecture is vastly inferior to the Athlon?

The first point should of course be checked, but beyond that...
Youguys seem very worried by this. No one is saying it is inferior. If it brings better performance at stock speeds then groovy, but it is in no way evil to test this, quite the contrary this is a very interesting thing that they have tested.

 

[AlphaWolf]

If it brings better performance at stock speeds then groovy, but it is in no way evil to test this, quite the contrary this is a very interesting thing that they have tested.

I see a few people saying they think it's interesting, perhaps you could help me understand why it's interesting to you?

 

[Geo]

Isn't a higher leak current expected when going to 90nm?

Well, at least apparently NV won't have to worry about a "soft ground" problem. It sounds like it was fab IP, and ATI already helpfully set that landmine off for them.

 

[Pete]

jb, OK, makes sense to compare to previous ATI gen. But the X850XT has just as much RAM as the X1800XL but clocked higher, so you'd want to compare an X800XT to the X1800XL to look at the same clocks. Even so, I suppose you'd also want to ensure that both cards are feeding their RAM the same voltage. I doubt the fan makes much of a difference, but I suppose it's also worth investigating.

Maintank and Dave, I know they don't downclock their cores, but didn't they tout features such as bypassing parts of the core not in use? Or is that just for their mobile parts?

hugh and Tokelil, I think nV's had very high single texture fillrate figures in 3DM since NV40. I remember how shocked we all were when the first ST fillrate figures were leaked, seeing as close they were to the theoretical ones compared to previous cards.

Judging from the comments on the article in DH's forum, not many know how to put this in perspective. Quite a few say they're "disappointed" in R520, but did they all forget that ATI ultimately clocks their cards higher?

I can't believe neither Veridian nor Zardon have heard of G70's clock steps. They don't even seem aware that G70 clocks its vertex shaders higher than its pixel shaders! In that sense, the test is unfair, because if you're going to let G70 clock parts higher, then why not run the vertex shader tests with R520's core at the same frequency as G70's VSs? If IPC is our goal, that would make sense. Kombatant, set 'em straight!