What do you expect for R650

[nutball]

If "silent_guy" could not see how R600 is limited in a lot of situations vs. G80GTX I'm very surprise.

I don't expect for nobody to come up in percentage by how much R650 be faster. All I would like to know what is limitation is and how R650 will be based on this estimate change.

You have to understand that the question you're asking isn't one you're going to get an answer to, at least not a simple answer like you seem to want.

The people who know the answer can't tell you the answer (NDA). The people who are giving out reasons on Internet forums don't know the answer, they're just guessing.

That's presuming that there is one, single answer and not several complex factors playing together...

 

[Shtal]

Sorry, don't mean to hurt anybody feeling, all I was interested was if it is true that R650 will have more texture units + higher clock frequency. Based on that change how would it effect performance of the GPU.

 

[ANova]

I'm sure ATI is aware or currently in the process of figuring out the bottleneck in the design and plan to do their best to rectify and maximize performance for the refresh, the question is whether or not it will be possible to do so without a costly rehaul.

At this time it seems rather pointless to debate.

 

[Zvekan]

A bug in R600? Texture limited? Inefficient ring bus?

Pretty much everybody thought that R600 would outperform a GTX, given the MC width and shader specs. Yet all of us were seriously wrong. That being the case, how do you expect anyone to come up with a meaningful percentage? Would it make any of us wiser if I really meant 42%? (That's average of course: 39.5% for Doom3 and 44.7% for CoD.)

On paper if counting bus width and number of "stream" processors, one would think that R600 should be faster, but if we look at the other perspective - number of transistors we should see that R600 doesn't really have the edge if we suppose that engineers in ATI/AMD and NVIDIA are equally capable.

Also we have to take in the account different working frequencies - to me at least - it appears that NVIDIA pretty much shocked it's competition via two, quite radically different clock domains.

Also we have to take in account that G80 uses NVIO external chip, and that R600 has an audio codec, CRYPTOrom (does G80 have it also?) and other things integrated right into the die.

Also 80HS process that R600 is built upon, while offering, typicaly higher speeds has quite a big problem regarding leakage as far as I have seen, even on AMD presentation slides - that makes clocking, or if fact cooling R600 quite a challange.

Zvekan

 

[silent_guy]

Also 80HS process that R600 is built upon, while offering, typicaly higher speeds has quite a big problem regarding leakage as far as I have seen, even on AMD presentation slides - that makes clocking, or if fact cooling R600 quite a challange.

Actually, that's one generally accepted point that I really don't understand.

Have a look at the following power consumption graphs on TechReport (similar to many others).

R600 seems to do just fine compared to the competition wrt power in idle mode, and only jumps much higher when under load. How this can be explained by leakage? Leakage is supposed to be a function of die size, voltage, and process, not clock speed. So what's going on here? Are they switching off internal power planes during idle or reducing voltage? Both are fairly common technique for battery powered gizmo's, but I'm not sure if they are also used high power applications.

It'd be nice if some in-depth review would look at voltages, but that's probably a bit much to ask.

 

[Sound_Card]

Not sure if voltages are lowered in 2D, but I'm believe that portions of the GPU are shut down during 2D along with lower clocks.

 

[Bjorn]

R600 seems to do just fine compared to the competition wrt power in idle mode, and only jumps much higher when under load. How this can be explained by leakage?

I'm guessing that leakage problems sounds much better then "AMD's design isn't as power efficient as Nvidias" to some people.

AFAIK, the R600 contains more transistors then the G80, and is clocked higher. Disregarding the shaders of course. So i don't see why we should be so surprised that the R600 consumes more power then the G80.

On the other hand, it seems to be a good overclocker so who knows, perhaps AMD had a much higher clock in mind when they designed it but the process didn't agree with them.

 

[silent_guy]

Not sure if voltages are lowered in 2D, but I'm believe that portions of the GPU are shut down during 2D along with lower clocks.

Yes, but shutting down or lowering clocks doesn't reduce leakage.

Let's switch over to the graphs of Xbitlabs, which isolates the power of the GPU card itself: a GTX and 2900 are more or less identical at idle. Yet at full power, the difference between them is 30W (vs 36W for Tech report). A 20% difference for (let's be charitable ) similar performance.

Now according to some articles, the 80nm process is very unpredictable wrt leakage: low in some cases, high in others. Dynamic power is more predictable and constant within the same process. Spec'ed at 210W worst case, 160W suggests that both Techreport and Xbitlabs had low leakage samples!

I have no trouble believing that leakage is, indeed, a problem for some cards. But I also think that it's not a factor in these particular examples, and that their overall architectural power efficiency is also part of the problem.
Looking back, this is in line with previous generations: wrt power efficiency, R580 and R520 weren't exactly stellar examples either.

 

[Geeforcer]

On the other hand, it seems to be a good overclocker so who knows, perhaps AMD had a much higher clock in mind when they designed it but the process didn't agree with them.

How good of an overclocker is it really, relatively speaking? I mean, 825-850MHz with a stock cooler looks impressive, but it's very much in the same percentage range as the boringly common 630-640MHz GTX overclock.

 

[Geeforcer]

[OT]

I am more interested in how it would do against G80 with 1955MHz shader core, 1.21 gigahertz RAM and 88 TMUs.

Note to self: no more veiled Back to the Future references.
[/OT]

 

[Shtal]

Yes, but shutting down or lowering clocks doesn't reduce leakage.

Let's switch over to the graphs of Xbitlabs, which isolates the power of the GPU card itself: a GTX and 2900 are more or less identical at idle. Yet at full power, the difference between them is 30W (vs 36W for Tech report). A 20% difference for (let's be charitable ) similar performance.

Now according to some articles, the 80nm process is very unpredictable wrt leakage: low in some cases, high in others. Dynamic power is more predictable and constant within the same process. Spec'ed at 210W worst case, 160W suggests that both Techreport and Xbitlabs had low leakage samples!

I have no trouble believing that leakage is, indeed, a problem for some cards. But I also think that it's not a factor in these particular examples, and that their overall architectural power efficiency is also part of the problem.
Looking back, this is in line with previous generations: wrt power efficiency, R580 and R520 weren't exactly stellar examples either.

Hi! Mr. silent_guy!

No offense in anyway.... Do you believe this problem based on architecture/design or the chip is too complicated based on 80nm-tech; if so, then it reminds me NV30 based on 130nm. "similar problem"

Also when ATI will switch to 65nm, in your opinion will it solve a problem by reducing enough - that you might even use single-slot solution for GPU-heatsink+fan.

 

[INKster]

I'm guessing that leakage problems sounds much better then "AMD's design isn't as power efficient as Nvidias" to some people.

AFAIK, the R600 contains more transistors then the G80, and is clocked higher. Disregarding the shaders of course. So i don't see why we should be so surprised that the R600 consumes more power then the G80.

On the other hand, it seems to be a good overclocker so who knows, perhaps AMD had a much higher clock in mind when they designed it but the process didn't agree with them.

Actually, if you count the 66 million transistors of the NVIO -an essential part of the design, since there would be no video output without it-, then the G80 products actually have more (681+66= 747 M) of them than R600 (720 M).
The NVIO is made at the 110nm half node, while G80 is 90nm. Both probably have higher main vcore values than R600 (80nm).

It is my opinion that R600 was never meant to be running at these high clockspeeds so soon.
However, they had no choice but to release it as fast as they could, and to cut back on the (very rare) HD2900 XTX'es core clocks (due to thermal issues), leaving the higher GDDR4 frequency as the primary differential factor against the regular HD2900 XT.
Let's not forget that a larger core also has a larger surface to dissipate heat, which could explain why the G80 makes due with a relatively simple copper/aluminum heatsink, while the R600's go for an all-copper solution. The similar exterior shrouds and fans are somewhat misleading at first sight.

R650 may well be what R580 was to R520, perhaps even more if the 65nm node is confirmed. The question remaining is what will Nvidia counter it with, since the G80 is far from having the same structural deficiencies as the NV4x/G7x lines had against R5xx (namely, image quality).
A simple shrink seems out of the question, unless FP64 support is somehow already embedded (but still deactivated at chip level, like Hyperthreading in the early "Willamette" Pentium 4's) in the G80 primary design.

 

[silent_guy]

Do you believe this problem based on architecture/design or the chip is too complicated based on 80nm-tech;

I don't believe there is such a thing a 'being too complicated' for a certain technology, if only because the word complicated is extremely vague...

Power is still one of the major things that's hard to predict accurately unless you're too close to tape-out. It's very hard to pin point why one architecture is intrinsically more efficient than the other. The ring bus definitely doesn't help, but is it a major factor? Probably not. A separate (5-way ported?) cache in front of the register file sounds weird and power and area expensive. Is it really necessary to run everything at 760 MHz instead of just the performance critical parts? etc etc. It's often also the sum of a lot of small parts that seem insignificant by themselves.

Also when ATI will switch to 65nm, in your opinion will it solve a problem by reducing enough - that you might even use single-slot solution for GPU-heatsink+fan.

For their high-end part? No way: that would mean they're leaving a lot of potential extra performance on the table. Too risky.

 

[Shtal]

You did made a valid point in both questions----> Mr.silent_guy.....

Thanks

 

[leoneazzurro]

I'm guessing that leakage problems sounds much better then "AMD's design isn't as power efficient as Nvidias" to some people.

AFAIK, the R600 contains more transistors then the G80, and is clocked higher. Disregarding the shaders of course. So i don't see why we should be so surprised that the R600 consumes more power then the G80.

On the other hand, it seems to be a good overclocker so who knows, perhaps AMD had a much higher clock in mind when they designed it but the process didn't agree with them.

R600 uses also the Fast14 technology, which can help to speed up the chip, but increases power consumption.

 

[satein]

I don't believe there is such a thing a 'being too complicated' for a certain technology, if only because the word complicated is extremely vague...

Power is still one of the major things that's hard to predict accurately unless you're too close to tape-out. It's very hard to pin point why one architecture is intrinsically more efficient than the other. The ring bus definitely doesn't help, but is it a major factor? Probably not. A separate (5-way ported?) cache in front of the register file sounds weird and power and area expensive. Is it really necessary to run everything at 760 MHz instead of just the performance critical parts? etc etc. It's often also the sum of a lot of small parts that seem insignificant by themselves.


For their high-end part? No way: that would mean they're leaving a lot of potential extra performance on the table. Too risky.

I could not answer or response any of your post and it is always a great read on your post :smile:

However, I did find an article on Japanese page at pc.watch.impress.co.jp that they tried to get an answer why R600 isn't this fast as expected...
The reason where AMD with Radeon HD 2000 does not take time fast シェーダ (This is a translation, I cann't read out Japanese anyway )
English translation by babelfish over here...

The translation bit may be not that good, but for you I think it would be ok.

Regards,

 

[Bjorn]

Actually, if you count the 66 million transistors of the NVIO -an essential part of the design, since there would be no video output without it-, then the G80 products actually have more (681+66= 747 M) of them than R600 (720 M).
The NVIO is made at the 110nm half node, while G80 is 90nm. Both probably have higher main vcore values than R600 (80nm).

Oops, forgot the NVIO.

It is my opinion that R600 was never meant to be running at these high clockspeeds so soon....

Probably true.

R650 may well be what R580 was to R520, perhaps even more if the 65nm node is confirmed. The question remaining is what will Nvidia counter it with, since the G80 is far from having the same structural deficiencies as the NV4x/G7x lines had against R5xx (namely, image quality).

It'll sure be an interesting battle. The G80 vs R600 battle turned out to be rather lackluster.

 

[seahawk]

Considering G80 ist still 90nm and believing both will move to 65nm with their next offering and considering that the transistor count is roughly equal a,d NV keeps NVIO, I would believe that NV has way more headroom to use, then does ATI.

 

[nicolasb]

Does anyone think R650 might include UVD?

If it were HDMI 1.3 compliant that would also be cool, but I'm certainly not holding my breath for that.

 

[Jawed]

Leakage is supposed to be a function of die size, voltage, and process, not clock speed. So what's going on here? Are they switching off internal power planes during idle or reducing voltage? Both are fairly common technique for battery powered gizmo's, but I'm not sure if they are also used high power applications.

Supposedly all ATI GPUs are a common design, mobile and desktop, so the power saving techniques are common to all.

Jawed