Rumors with numbers ATI vs Nvidia

[metro.cl]

well i've been searching for info about r580 and g71 for a while now, i have a clear picture that is realistic.

Here is what i wrote:

http://www.chilehardware.cl/index.ph...=143595#143595

google translation:
http://translate.google.com/translat...language_tools


i based everything on a couple of chineese guys that seem to know a lot, and some info i get from insiders at ATI and Nvidia (nothing really special but i know when someone is not telling something realistic).

regards

feel free to post comments here or there i'll ad info if you have a good source

im goin out to dinner then i'll do a small translation

[Shadowmage]

700mhz+ on G71 yet only 650mhz on R580?

Looks like NV and ATI are switching positions again after ATI's brief stint as clock speed leader...

[DilTech]

Interested....G71 will also include HDR + AA....

Unfortunately, I don't see it being fast enough on either ATi or NVidia's side to actually be useful for us... If it can be done higher than 1024x768 at GOOD framerates I'd be shocked...

[Tim]

Interested....G71 will also include HDR + AA....

Unfortunately, I don't see it being fast enough on either ATi or NVidia's side to actually be useful for us... If it can be done higher than 1024x768 at GOOD framerates I'd be shocked...

yea, I wonder how this will work out...anyway, this is from the penstarsys article

While nothing concrete is known about the 90 nm refresh of the G70, we can speculate that the product replacing the 7800 GTX will at least reach 550 MHz and that the high end could possibly hit 700 MHz at maximum. NVIDIA will not stop at just clock speed. We can expect that it will have a full 8 quads of pixel shaders, 10 vertex shaders, and 16 “Super” ROPS that will be able to handle the output of those shaders. I have heard rumors that the AA unit will be getting a makeover and it will be able to handle HDR anti-aliasing. I have also heard rumors that texture filtering will also be getting a boost and we can expect texture quality to match that of the older FX series. This product could easily hit 380 million transistors, and with the addition of 90 nm Low-K (remember, the regular G70 is 110 nm FSG- it does not get a transistor performance increase by using Low-K) this product will hit some impressive clockspeeds. One thing that does not look to change will be the memory controller. NVIDIA does not feel the need for a programmable memory controller like ATI has as of yet, and will instead rely on faster GDDR-3 memory to make up the difference.

So little is know about these cards, but we ARE talking about next month for the R580 and Feb(?) for the G71.....direct opposite of R520 (which was rumour galore )

It's amazing that we are talking about infact new cards here.....it's all a bit crazy.... 4 months for R520 and a vapor 512mb GTX.....so far for 12month refresh, seems like they improved the original 6 month refresh here

One very nice thing is angle independent AF also for nVidia, it's amazing that they had it on the 59XX series...but it wasn't a big deal back then...dunno why, wasn't really into hardware back then

[biohead]

heh Tim, development seems to speed up exponentially

[perkam]

Metro.Cl, could you post a version translated into English so that the link can be more handy

Perkam

[softpain]

Yes please Metro in case I missed something....?

[MaxxxRacer]

no offense to anyone here, but speculation is just that.. speculation.. remember the 32pipe rumors of the R520...

[ahmad]

Wasn't the G71 already announced to be 750MHz?

ATi already does HDR+AA at 30FPS. And if this is only slightly faster, I am sure 50-60FPs shouldn't be much of a problem.

[nn_step]

Until I can buy the cards.. They aren't real to me..
I have no need for Rumors.

[Frisch]

I have no need for Rumors.

I saw your girlfriend hanging out with an ati guy the other day.


Sorry couldn't resist

[nn_step]

I saw your girlfriend hanging out with an ati guy the other day.


Sorry couldn't resist

and I Saw you making out

Sorry I couldn't Resist.

I understand the reasoning, but pls keep such comments within bounds - Perkam

[Cybercat]

I never understood the "16 pipes processing 3 pixels each" implementation. Hopefully they'll explain it in detail when these cards come out.

[perkam]

Let's see if we can't use what we know now to make some estimates...

The X1300PRO is 4 pipes processing 1 pixel each. Gets 3k in 05
The X1600PRO (which I now loathe ) is 4 pipes processing 3 pixels each. Gets 4k in 05 (PROvsPRO because same mem) (33.33% increase in perf)

The R520 is 16 pipes processing 1 pixels each. Gets 8.7k in 05
If the R580 is 16 pipes processing 3 pixels each meaning going from 16x1 (r520) to 16x3 (r580?)....taking into account the increase in 05 score from 4x1 to 4x3, if they both end up having the same memory, the R580 would get 11.6k stock in 05. Thats assuming same clocks and same memory on an average system. (8700 x 1.333333333333)

In a much faster system and assuming higher core and mem clocks than the current XT (both of which are expected), we should see the R580 scoring 12.8k stock in 05. (9600 x 1.33333333333333)

Perkam

[Flak Monkey]

I never understood the "16 pipes processing 3 pixels each" implementation. Hopefully they'll explain it in detail when these cards come out.

It's the same way that the 5900 series has 4 pixel pipelines but 2 texture units per pipeline. Making it the theoretical equivalent of 8 pixel pipelines. That didn't work out to well though and from what I've seen of the X1600 so far it appears to be just as inefficient on the X1k series as on the FX series. Good thing they decided to give the r580 48pipes (16x3) instead of 32 because with the G71 being >= the r580 in clock speed it's going to need at least one step up in the number of shaders.

[Shadowmage]

The reason X1600 is slow is because it has 4 texture units, 4 ROPs, and 4 of some other stuff (can't remember off the top of my head).

[Cybercat]

It's the same way that the 5900 series has 4 pixel pipelines but 2 texture units per pipeline. Making it the theoretical equivalent of 8 pixel pipelines. That didn't work out to well though and from what I've seen of the X1600 so far it appears to be just as inefficient on the X1k series as on the FX series. Good thing they decided to give the r580 48pipes (16x3) instead of 32 because with the G71 being >= the r580 in clock speed it's going to need at least one step up in the number of shaders.

The FX's texturing and Z-op performance was theoretically the same as 8 pipes, but what made it suck was shader performance.

The X1600 is the complete opposite. Shader equivalent of twelve pipes, but the texturing performance of four. This isn't so much inefficient as it is a pretty major bottleneck. That along with memory bandwidth.

The thing I don't understand about this "16 pipes processing 3 pixels each" is, this is implying that the X1600 doesn't actually have 12 shader pipes, and the R580 doesn't actually have 48. There is a lot of confusion in this area, which I hope future articles clear up. There's already plenty of documentation that disproves this assessment though, as they all say the X1600 physically has twelve pipes. It would be a grave miscommunication on ATI's part if it actually had much less than that.

But what is a "pixel pipeline"? The shader pipes do the math calculations, the ROPs do the pixel drawing. With so many asymetric designs going around, the term "pixel pipeline" isn't as straightforward anymore. It used to mean the shader units, rasterizers, and texture units, all in one structure - in the end refering to how many pixels could be drawn. As far as "pixel pipelines" go, the G70 only has 16. However it's refered to as having 24, to emphasize the shader and texture units. With ATI, it's a bit different. The RV530 is said to have 12 pipes, but with 4 ROPs, and the R580 with 48 pipes, and 16 ROPs. So yes, as far as "pixel pipelines" is concerned, the R580 has 16. On that same note, you'd say the RV530 has 4.

So, how does the R580 have 16 pipes that can draw three pixels at once? This must mean the ROPs are revamped. The shader units, which there are 48 of, don't actually draw the pixels themselves, AFAIK. They do the math calculations behind them. I'm going to say that ATI probably thinks of pipes as the number of ROPs and texture units, which is convenient because with ATI they're always the same number. That would be more accurate when talking about "pixel pipelines" than to talk about the math capacity of your architecture. Of course, ignoring the shader units as the main point of the architecture takes a lot away from their marketing appeal. Obviously a RV530 with 4 pipes isn't as appealing as a 6600GT with 8.

[trakslacker]

^ damn fine post. Quite informative.

[Hicks]

Talking about rumours is all good fun, and we can all speculate what we want, but most of the time it gets out of hand, with people saying it's 100% definate because some idiot on another forum said so.

[turtle]

@ cybercat's post:

We're all confused, but here's my assumption based on what i've gathered from the pros at B3D, although don't misunderstand me, I understand about 1/3 of what they're saying.

It's 16ROPs processing 3 ALUs per rop instead of 1 like R520, kind of like a C1 Xenos (x360), but without unified architecture (and EDRAM etc). This does seem to make the major change in the architecture between R520 and R580 in the ROPs. 16 texture units like R520, 16 ROPs like R520. The 48 fragment shader pipes most likely pertain to ALU's. To my understanding that means while only 16pixels can be displayed at a time, 48 shader processes can be calculated waiting to be displayed, and somehow ATi made this process feasible. So while it seems to be future-proofing for games in the future that are more shader intensive (3:1 shader/texture or higher) some things coded more texture intensive may show little improvement over r520 except for the higher clocks. But again, if the future goes that way, more shader intensive, ATi should be set up until unified shaders, just adjusting it's product to run faster (80nm) and perhaps use GDDR4 in an update (R590?).

One things for sure, it'll probably whoop-ass in 3dmark05, if nothing else. Like perk said, probably up 30-35%, because parts of 3dmark05 are very shader intensive (hence how a x1600 does relatively well regardless of it's 128bit bus and 4rops). So if 3dmark is representive, it may be well prepared for the future, but I expect it to do little better in games/benches not geared toward these tasks. You'll probably see results all over the board dependant on the game and how it's coded (Far Cry will probably go through the roof again...), where-as Nvidia's may be more consistant because 2:1, as G71 seems to be, is more in gear with current games. This could turn in the near future though, who knows?

My big question on the matter is which compensates better in the areas it performs worse than the other for now, and in the future, as it's hard to guess. If you say R520/G70 are relatively equal (430/1200 256mb vs 625/1500 512mb with fairly similar vs/tmus/alus), you have to weigh what each company is bringing to the table for the refresh (NV:200+ mhz, slightly better ALU's [2], more ram, faster ram, 8 shader pipes, more tmu's, and possibley more vertex shaders; ATi: slightly faster clocks, faster ram, 3x ALU's, "32 more pixel shaders") it becomes very difficult to judge the two, as they are no longer comparible with now completely different architectures, and will perform much different in different scenarios depending on how things are coded.

[metro.cl]

nice explanations, for what i've read the new ati pipes have a drawback, they are slower than normal pixes processing, also think about all of this as maxium theorical values that are dependant in many other things.

regards.

[softpain]

Yes Turtle that is about what I understand....nice summury btw!

[Shadowmage]

They definately have a drawback if you consider them to be 48 pipes. They have no drawback if you consider them to be fattened 16 pipes

[Cybercat]

Ah, turtle, I haven't had the pleasure of reading your posts for a while now. What you say does seem to be the case, although I don't know if I'll rule out 12 true quads totally just yet, until it's out and everyone has had the last word, so to speak.

[perkam]

I hope rumours that ATI will be stopping production of x1800 arent true

Was hoping to see a few failed cores bumped down to 12 pipes in the form of the X1800PRO...though with the lack of the 7800GS...doesnt seem practical

Perkam