G70 Vs X1800 Efficiency

[j^aws]

Its easy to forget that these tests aren't actually showing that.

Sorry, maybe my post wasn't clear. I'm showing how '32bit' theoretical peaks, taking into consideration issue rates (something I usually use between different pipes etc.) and how they correlate to the actual test numbers here, i.e. there's a 33% difference per cycle between R520 and G70 PS units, therefore the G70 has more raw PS power per cycle which this test is showing.

However, the numbers are not meant to suggest dynamic branching performance etc... but 'raw' shading horsepower per pipe, per cycle...

 

[Dave Baumann]

The point I was making is that the tests aren't actually testing FP32 processing capacity, since probably all of the tests will be using partial precision hints, ergo these "like for like" tests are showing one processing FP32 always and the other a mixture of FP16 and FP32.

 

[Geo]

The question IMHO should be not if that review is made perfectly right, but rather, is this view INTERESTING.
Is it?
It is for me. And it seems its interesting for you too.

Me too. Tho I think Wavey put his finger on what it should have been labeled --"IPC" rather than "efficiency". Sure, you can say "Looking at the paper specs I would have predicted these results", but it is always nice to test these things and put quantifiable numbers to them.

 

[neliz]

Well I don't think ATi's plan orginally was to clock this thing to sky is the limit. I would like to see the comparision between the r520 and the nv40, I think efficiency wise in sm 2.0/3.0 will be very similiar per clock other then dynamic branching.

Philosophies change depending on competitors products. Its almost as if ATi had no choice but to clock it up this high.

if you can hide an instruction for say, 4 cycles, would you rather have those four cycles take 10 seconds in total or 6 seconds in total?

 

[ChrisRay]

The point I was making is that the tests aren't actually testing FP32 processing capacity, since probably all of the tests will be using partial precision hints, ergo these "like for like" tests are showing one processing FP32 always and the other a mixture of FP16 and FP32.

I'm a bit confused. Shadermark uses FP32 unless you request a PP hint. And Half Life 2 uses FP32. At least half the tests performed here at least are relevant to FP32.


*Edit* Just checked. There is no difference between FP32 and default mode for shadermark on my 7800GTX SLI setup.

 

[Geo]

Well I don't think ATi's plan orginally was to clock this thing to sky is the limit. I would like to see the comparision between the r520 and the nv40, I think efficiency wise in sm 2.0/3.0 will be very similiar per clock other then dynamic branching.

Philosophies change depending on competitors products. Its almost as if ATi had no choice but to clock it up this high.

You never know about scuttlebutt, and how much is "liar's dice" being passed on. Having said that, going back over the rumor mill on 16 pipes (and you'll all recall that was in a minority on its own right to the end), it seems to me the prevalent belief was 650mhz and north for the tippy-top part. I had PMs from some typically sensible types over the course of that period that suggested maybe quite a bit north. Which is just to say that I'm not convinced that ATI saw 625mhz as "sky is the limit" for this part.

It'll be very interesting to see what R580 is clocked at. Any hints there yet?

 

[neliz]

It'll be very interesting to see what R580 is clocked at. Any hints there yet?

50Mhz over the XT-PE?

 

[Bouncing Zabaglione Bros.]

Philosophies change depending on competitors products. Its almost as if ATi had no choice but to clock it up this high.

I think R520 is actually clocked very conservatively, looking at how easy it is for the latest cores to go even higher. I would have been very surprised for ATI to have decided to go to .09 and not expect to get higher clockspeeds out of it.

 

[CMAN]

So basically people are arguing whether what we see/learned from the article is useful. One side says it is useful because it exposes IPC. While others see it is not useful because we already knew it and doesn't expose anything. Is this correct?

 

[trinibwoy]

So basically people are arguing whether what we see/learned from the article is useful. One side says it is useful because it exposes IPC. While others see it is not useful because we already knew it and doesn't expose anything. Is this correct?

Well one side says it is useful because it exposes IPC and rebukes ATi's claims of higher efficiency (although such claims were obviously in comparison to the R420). The other side of the argument seems to say it's worthless and misleading because it has no bearing on the performance of retail configurations.

It was a purely academic exercise, but people seem to be getting their panties in a twist because they think it casts the R520 in a bad light.

 

[soylent]

So basically people are arguing whether what we see/learned from the article is useful. One side says it is useful because it exposes IPC. While others see it is not useful because we already knew it and doesn't expose anything. Is this correct?

The argument would have been less confused and less heated if the tread title was changed to IPC instead of efficiency.

Most people would probably agree that "efficiency" = IPC/theoretical IPC or "effeciency" = actual/theoretical max of some property. And then there is also the question of how we take a weighted average over efficiencies for a bunch of different things and come up with an overall efficiency rating and so on.

 

[Joe DeFuria]

The argument would have been less confused and less heated if the tread title was changed to IPC instead of efficiency.

The problem is, the thread title reflects the "conclusions" drawn by the article in question. The article talks about "efficiency" (indeed, crowns the G70 the most "efficient design"), which is a big problem.

Many people (myself one of them) are disappointed more with the conclusions drawn and presented from the article, than with the exercise itself.

 

[saf1]

It's a totally artificial comparison that is never going to happen in the real world.

The problem is that it did happen in the real world. And the results are there to make of it what you will.

While I agree it is not apples to apples and in many ways very different - yet it many ways it is the same. Just as it is the same in the pain ATI went through to bring the card to market much like the NV30 with Nvidia. The question I will ask is this - is the CEO of ATI able to recover as the CEO of Nvidia did?

In the end, the general consumer will buy whatever brand they are loyal to. This test will not do anything for them. It was just a test for the geeks out there who had some time, a couple of cards, and good music to test to. No more, no less.

 

[erick]

Nice discussion going on.

Wanted to bring your attention to the 3DMark 05 geometry (triangle) test table shown here :

http://www.driverheaven.net/articles/efficiency/testing2.htm

R520 is showing A HUGE lead over 7800 GTX here:

128 Triangles (batch size)
G70 27.4 MTriangles/s
R520 119.5 MTriangles/s

R520 is a WHOPPING 4.3 times faster 'clock-per-clock' (although I know that games do not use so small batches anymore).

Looking further, into the batch sizes that I am told today's games actually use (512-2048), we see such numbers:

512 Triangles (batch size)
G70 67.9 MTriangles/s
R520 133.5 MTriangles/s (2x faster than G70)

2048 Triangles (batch size)
G70 100.3 MTriangles/s
R520 137.4 MTriangles/s (about 1.35x faster than G70)

These numbers are pretty interesting just like that, but for the fun of it let's factor in the ACTUAL clockspeeds as well (G70 430Mhz and R520 625MHz):

512 Triangles (batch size)
G70 64.9 MTriangles/s
R520 183.3 MTriangles/s (2.8x faster than G70)

2048 Triangles (batch size)
G70 95.8 MTriangles/s
R520 188.7 MTriangles/s (1.969 or nearly 2x faster than G70)

The most interesting question for me is: Why does this HUGE difference not present itself in the games? And will it ever?

Anyhow, one thing is for sure, and this is that R520 wastes G70 in geometrical prowess (and efficiency) as long as triangles are involved

 

[Hubert]

The problem is, the thread title reflects the "conclusions" drawn by the article in question. The article talks about "efficiency" (indeed, crowns the G70 the most "efficient design"), which is a big problem.

Many people (myself one of them) are disappointed more with the conclusions drawn and presented from the article, than with the exercise itself.

Exactly. lemme' just quote:

Bolded out conclusion: "the G70 is the most efficient design and therefore performance leader when both architectures are configured similarly"

and then

"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."

It seems obvious that the usual gamer will choose G70 based on this conclusion. It is more efficient, isn't it. So while it is agreedly artificial and purely academic, still, it says that the G70 is better, buy it. Heck, I wont buy it 'cos does not have HDR+AA, and it has much performance loss with AA enabled. And it's green. And Ruby is a sexy chick.

 

[John Reynolds]

Just as it is the same in the pain ATI went through to bring the card to market much like the NV30 with Nvidia. The question I will ask is this - is the CEO of ATI able to recover as the CEO of Nvidia did?

This assumes the market will react to X1xxx series the same as it did to NV30.

 

[Bouncing Zabaglione Bros.]

The problem is that it did happen in the real world. And the results are there to make of it what you will.

What people are going to downclock their 625 mhz cards down to 400 mhz? I don't think so.

While I agree it is not apples to apples and in many ways very different - yet it many ways it is the same. Just as it is the same in the pain ATI went through to bring the card to market much like the NV30 with Nvidia. The question I will ask is this - is the CEO of ATI able to recover as the CEO of Nvidia did?

Totally different. NV30 was a year late and very poor - to the extent that Nvidia cancelled it straight away. R520 was six months late and is very good.

In the end, the general consumer will buy whatever brand they are loyal to. This test will not do anything for them. It was just a test for the geeks out there who had some time, a couple of cards, and good music to test to. No more, no less.

And pointless, with flawed methodology and conclusions that are not supported by the "facts".

 

[Geo]

50Mhz over the XT-PE?

Heh. And XT-PE clocks at?

I tend to think they'd be doing pretty good to bring R580 in at X1800XT clocks. If it is clocked much higher than that, I suppose it might be explained by maturation of 90nm, but I'd tend to lean towards just learning some lessons with R520 and applying them.

 

[saf1]

The real equation here of interest to me is that while you can drastically underclock an XL to match slightly overclocked G70 speeds, there's no way in hell to slightly overclock the XL and then bring up the G70 clock to match it, is there? Gaw lee... Whadd'ay'a think about dem apples?...

I think you answered your own question - there has been no need for Nvidia to clock their GPU to the speeds ATI does.

 

[saf1]

What people are going to downclock their 625 mhz cards down to 400 mhz? I don't think so.


Totally different. NV30 was a year late and very poor - to the extent that Nvidia cancelled it straight away. R520 was six months late and is very good.


And pointless, with flawed methodology and conclusions that are not supported by the "facts".

I didn't say people let alone the general consumer. I said it happened, you said it wouldn't happen. Totally different to use your words.

Regarding NV30 vs the 5xx - is it different? How so? Late is late. If you don't believe me, look at the stock and tell me what the investors say, ok? It is late, they made mistakes. My point is that both companies made mistakes in bringing a much hyped card to the market. It does not matter if the card was good or bad, they "BOTH" had problems in bringing a card to market. You want to know what else is funny, both companies delt with Microsoft and the xbox project...

While the 5xx may be good - the question remains how long it will be on the market. We can let the fan boys and girls determine what card they will buy.

Nothing is pointless - unless the folks who wrote the article work for you and you had another project for them to work on that didn't get done. That article will do nothing for the general consumer who is going to buy a product that is available today on the market by either company.

It was nothing more than a "what-if".