AMD: R8xx Speculation

[Ailuros]

Look, I'm not sure about anything, of course. But, history says that normally is more difficult to get higher clock speeds on a new process, and that in the past new GPUs on a new process were not so much higher clocked with respect to slightly older GPU made with an older process. RV790 was a "singlularity" as it was specifically targetted for high clocks but this was realized on a process proved for more than one year, after some refinement. If initial rumored 40 nm problems were not a fairy tale, most of the effort should have been spent on the yield issue instead of getting the absolutely better clock speed obtainable. Moreover, power consumption usually goes higher with clock speed (as higher clock often needs higher voltage), so a bigger chip with lower clock speed sometimes could be more easily obtainable. If the hints about the die size of Cypress are right, DX11, RBE and texturing improvements are unlikely to be responsible for such a big increase in size, considering the scaling.
Of course there are so many "if" but this is the "AMD: R8xx Speculation" thread and we shoud reason about what is rumored to be true.
If we want to reason about the "2 Teraflop" wording (but again, the sentence said "more than 2 teraflops") we can see that 2 teraflops could be reached by a 800SP @ 1.25 GHz, 1200 SP @ 850 MHz, 1600 SP @625 MHz, 2000 SP @ 500 MHz or 2400 SP @ 417 MHz. This if the architecture is quite similar to what RV770 is, of course.

I'd expect something like 5850 to be on the 2TFLOP level and the 5870 to distance itself from it by a reasonable persentage. The question still remains how they reach it. I'm in no way claiming that it has to be say 1280 or 1600SPs at all, I'm merely wondering where the safe clues are that point into one or the other direction. Frankly each case scenario works for me, as there doesn't seem to be any significant performance difference after all at least in theory.

As seahawk already pointed out, there are no indications that TSMC's 40nm yield problems have anything to do with frequencies directly. Of course there is a chance that some time ago when yields weren't at the stage they are today, that any IHV could try to lower frequencies and rise yields by a margin X. But I haven't seen anything this far that points in the direction that AMD has missed its original target with Evergreen.

And yes power consumption rises with higher frequencies but it's in no way irrelevant to chip complexity either. A@900MHz consumes more than A@700MHz, but if you try to compare it to B then B would have to have the same complexity as A.

4770 is 750 Mhz (same as Rv770) and does not seem to overclock to 900+ Mhz so easily.

A chip that went into production way earlier than any Evergreen SKU, with a lot more problematic yields than today, with merely 640SPs and was targetting the mainstream space of the RV7x0 family of GPUs. With Evergreen the entire picture of performance levels changes a lot and if anything you'd better compare a 4770 with Juniper than with Cypress.

Look at it that way: I wouldn't want to read soon that if Cypress turns out to be a 16 cluster GPU any signs of possible "disappointment" or "is that all" kind of comments. What's important to me is what shoots out at the other end and as I said it won't change much.

 

[no-X]

Every radeon since RV100 had TWO DACs

No. R200 has single internal DAC:

You can see the external DAC here (near the fan, marked by red and blue):



Some boards doesn't have it, e.g. this Hercules R8500 LE:



This model isn't capable of independent analog output (it was also cheaper because of that and quite popular among gamers)

Here is datasheet of the external DAC:

http://www.analog.com/en/digital-to-analog-converters/video-encoders/adv7123/products/product.html

 

[psolord]

Some pricing rumours.

Well if the euro/usd rates are as the www.xe.com suggests (1euro=1.43USD), then the 5870X2 could fall to about 420 euros, which would be at the same price level of the 4870X2 when it launched (450 euros), which is accepted by me!

I wouldn't want to pay more than 450 euros for a 5870X2. No matter how you see it thouth, the 5850 at 209 euros is quite expensive, even if it is faster than the 4890. Price/performance ratio should be better compared to older products, not the same!

 

[leoneazzurro]

I'd expect something like 5850 to be on the 2TFLOP level and the 5870 to distance itself from it by a reasonable persentage. The question still remains how they reach it. I'm in no way claiming that it has to be say 1280 or 1600SPs at all, I'm merely wondering where the safe clues are that point into one or the other direction. Frankly each case scenario works for me, as there doesn't seem to be any significant performance difference after all at least in theory.

The difference in performance should justify the price premium, so it should not be less than 20%. If both variants have GDRR5 (and it's very likely to be so), the difference will be almost completely given by core clock speeds. So, in the case of 1200SP, Cypress should be 1+ GHz at launch and have volume of the 5870. This speed was reached only in selected OC RV790 parts. TSMC is rumored to have fixed the yield problems 40 only after the RV740 release. This smells not like a "mature" process node.
So, my guess is that a 1200SP, 1+ GHz part is probably more unlikely to appear than a 1600SP, 750 MHz part. Then maybe we could have a 1520 or a 1760 SP part, who knows

As seahawk already pointed out, there are no indications that TSMC's 40nm yield problems have anything to do with frequencies directly. Of course there is a chance that some time ago when yields weren't at the stage they are today, that any IHV could try to lower frequencies and rise yields by a margin X. But I haven't seen anything this far that points in the direction that AMD has missed its original target with Evergreen.

What I was pointing out is that RV740 launched at 750 Mhz, not 850. So, at that point the process seemed not to be mature enough to allow good yields (which are also affected by the desired clock)
Also, RV770->Rv790 took at least 8 months. RV740->Cypress four and half. With all the process problem we heard on the 40nm and that seems to be much higher than those probably encountered on the 55nm node. So, very high clocks are possible? Yes. Are they likely? Given the history, I would not say so.

And yes power consumption rises with higher frequencies but it's in no way irrelevant to chip complexity either. A@900MHz consumes more than A@700MHz, but if you try to compare it to B then B would have to have the same complexity as A.

What I was saying was merely that power consumption normally rises more than linearly with the frequency, because to achieve higher clocks it's common practice to increase voltage, too. So a slightly bigger chip with lower clocks and voltage could have a lower TDP and it could be better cooled than a slighly smaller chip with higher clocks and voltage. And designers could also to decide that way, especially for the higher margins parts.

A chip that went into production way earlier than any Evergreen SKU, with a lot more problematic yields than today, with merely 640SPs and was targetting the mainstream space of the RV7x0 family of GPUs. With Evergreen the entire picture of performance levels changes a lot and if anything you'd better compare a 4770 with Juniper than with Cypress.

Yes, they will be different chips, but they have to comply to what design and process can do. Up to date, 1+ Ghz with architectures similar to Rv7x0 were not so easily attainable. Rumors hints about Cypress being bigger than Rv770. So I'm giving my guess to a wider but lower clocked variant than a "slim-but-fast" part.

Look at it that way: I wouldn't want to read soon that if Cypress turns out to be a 16 cluster GPU any signs of possible "disappointment" or "is that all" kind of comments. What's important to me is what shoots out at the other end and as I said it won't change much.

I'm not speaking about "disappointment" (I gived up with desktops, so my personali "interest" in these boards' performance is zero) , I was merely looking at the rumors and making a guess that, IMO, could cope tha most with them from a technical point of view. :smile: Of course then ATI will surprise us all with 800+ scalar shaders @ 1.5 Ghz (double pumped)

 

[Jawed]

In the time of RV770 deploy 55nm was an 9 month old process,

RV770 worked well enough on its initial version that that's what was released. That's reasonably rare.

I can't understand the rest of your post. I think you're disagreeing with me, but I can't work out the reasons for your disagreement.

Jawed

 

[Ailuros]

The difference in performance should justify the price premium, so it should not be less than 20%. If both variants have GDRR5 (and it's very likely to be so), the difference will be almost completely given by core clock speeds. So, in the case of 1200SP, Cypress should be 1+ GHz at launch and have volume of the 5870. This speed was reached only in selected OC RV790 parts. TSMC is rumored to have fixed the yield problems 40 only after the RV740 release. This smells not like a "mature" process node.
So, my guess is that a 1200SP, 1+ GHz part is probably more unlikely to appear than a 1600SP, 750 MHz part. Then maybe we could have a 1520 or a 1760 SP part, who knows

Why should it be at over 1GHz in the first place? Assuming no change in SIMDs could get you easily to 2.3 TFLOPs with 16 clusters and 900MHz.

And no there's still no connection between TSMCs manufacturing process problems and the real targets of each IHV. Either way whether 1280 or 1600 they didn't make that decision yesterday or within this year either. If they'd gone hypothetically for any "plan B" they wouldn't release "on time" soon.

What has a 55nm chip to do with it again? If anything it was a valuable experience learned for ATI, which could have worked both ways.

What I was pointing out is that RV740 launched at 750 Mhz, not 850. So, at that point the process seemed not to be mature enough to allow good yields (which are also affected by the desired clock)

Compare it against Juniper; add the necessary "ifs" and you're at the same dead end.

Also, RV770->Rv790 took at least 8 months. RV740->Cypress four and half. With all the process problem we heard on the 40nm and that seems to be much higher than those probably encountered on the 55nm node. So, very high clocks are possible? Yes. Are they likely? Given the history, I would not say so.

What has the distance between 740 and Cypress to do with it anyway? Again RV790 is a 55nm chip and RV740 is a mainstream chip. I still don't see why one is "likelier" as the other. If I'd have a blond moment I could eventually say that 1600 matches my purse better....

Yes, they will be different chips, but they have to comply to what design and process can do. Up to date, 1+ Ghz with architectures similar to Rv7x0 were not so easily attainable. Rumors hints about Cypress being bigger than Rv770. So I'm giving my guess to a wider but lower clocked variant than a "slim-but-fast" part.

No one said anything about over 1GHz frequencies; normally the units that are less tolerant to higher frequencies are TMUs, ROPs and way less ALUs afaik. On the other side I can see Arun mumbling about possible >1TMU clock domains on the other side of the river bank and yes I am wondering what is and what isn't possible.

900MHz under 40nm doesn't sound like a problem to me but eventually I might be proven wrong soon.

I'm not speaking about "disappointment" (I gived up with desktops, so my personali "interest" in these boards' performance is zero) , I was merely looking at the rumors and making a guess that, IMO, could cope tha most with them from a technical point of view. :smile: Of course then ATI will surprise us all with 800+ scalar shaders @ 1.5 Ghz (double pumped)

Neither the term "scalar" nor the clock domain theory matches my blond moments so let's leave it as is

 

[Sampsa]

$299 confirmed? I doubt PowerColor would throw anything but top-of-the-line card for big top prize

I couldn't find the specific image from PowerColors site though, so it might have been removed as unintentional leak
http://www.powercolor.com/Global/NewsInfo.asp?id=701

Here is a full PR from PowerColor:

POWERCOLOR DESIGN CONTEST KICKS OFF NEXT GEN VIDEO CARDS
PowerColor

“Design and Choose the Best” contest will give away more than USD$2,000 in prizes

Taipei, Taiwan –September 1, 2009– TUL Corporation, a leading manufacturer of AMD graphics cards, announces the “Design and Choose the Best” online contest. Gamers and designers from around the globe can download design elements from the contest website and incorporate them into wallpaper, screensavers and IM chat icons. Artwork can be uploaded via the contest detail page and each category will be voted on by the public. Top vote getters from each category (3) will be some of the first to own a high end DirectX 11 next generation PowerColor graphics card!

Designers won’t be the only ones with a chance to win great prizes. Voters will have a chance to win 1 of 20 Go! Green HD4670’s. Voting will begin Sep. 21st and participates will automatically be entered into a random drawing for the video cards at the conclusion of the contest.

Ready to design and vote for the best? Visit our website for more details:
http://www.powercolor.com/event/wallpaper/index.htm

About TUL Corporation
TUL is a leading supplier of AMD graphics cards under the PowerColor brand. We offer award-winning products based on our technology leading components, quality design and superb engineering. TUL employs more than 150 talents worldwide, with offices in China, Europe, Russia and USA to support over 300 channel partners and distributors in 50 plus countries. For further information, visit: http://www.tul.com.tw/global/

For more product information, visit: www.powercolor.com.
Or contact: pr@powercolor.com

 

[leoneazzurro]

Why should it be at over 1GHz in the first place? Assuming no change in SIMDs could get you easily to 2.3 TFLOPs with 16 clusters and 900MHz.

And no there's still no connection between TSMCs manufacturing process problems and the real targets of each IHV. Either way whether 1280 or 1600 they didn't make that decision yesterday or within this year either. If they'd gone hypothetically for any "plan B" they wouldn't release "on time" soon.

What has a 55nm chip to do with it again? If anything it was a valuable experience learned for ATI, which could have worked both ways.

Compare it against Juniper; add the necessary "ifs" and you're at the same dead end.

What has the distance between 740 and Cypress to do with it anyway? Again RV790 is a 55nm chip and RV740 is a mainstream chip. I still don't see why one is "likelier" as the other. If I'd have a blond moment I could eventually say that 1600 matches my purse better....

No one said anything about over 1GHz frequencies; normally the units that are less tolerant to higher frequencies are TMUs, ROPs and way less ALUs afaik. On the other side I can see Arun mumbling about possible >1TMU clock domains on the other side of the river bank and yes I am wondering what is and what isn't possible.

900MHz under 40nm doesn't sound like a problem to me but eventually I might be proven wrong soon.

Neither the term "scalar" nor the clock domain theory matches my blond moments so let's leave it as is

*Sigh*
1+ Ghz are needed with 1200 SP to achieve 2 teraflop (5850) + 20+% (5870). If there will be 1280, it will be different, but you'll anyway need more than 900 Mhz.
55 nm chip are representative of a manufacturing process that had not so big problems. Making a big volume of chips on a 40 nm process with good yields (failures also include parts that are not capable to reach a given clock) seemed until now a "bit" harder for both Nvidia and ATI.
Of course 900 Mhz will be reachable. Now? Maybe, maybe not.
The distance between RV770 and RV790 could be representative of the improvement in we can expect to reach on a given process (in this case 55nm) in a certain time range. 55 nm seemed to give less problems than what 40 nm is rumored to do.
And yes, RV740 should be compared to Juniper, but, even if we suppose an improvement like the one had with the Rv770->RV790 upgrade, and we consider that chips with the same basic architecture do not normally differ so much in term of clock speed (i.e. RV730-RV770) a 1280 SP chip will be a little off the mark.

And, anyway, this is called the speculation thread, so if we cannot speculate about something because we don't know exactly what it is has no reason to exist, can we?

 

[seahawk]

I would dare to say that 5870 needs to be 20% faster then 5850, but I doubt it does need to clock 20% higher or have 20% more TFlops for that.

The combination of GPU and Memory just needs to be 20% slower. Now if you think about a 256 Bit part, there is one obvious solution to make the part slower and cheaper for the partners.

 

[Tim]

*Sigh*
1+ Ghz are needed with 1200 SP to achieve 2 teraflop (5850) + 20+% (5870). If there will be 1280, it will be different, but you'll anyway need more than 900 Mhz.

You math is off.

With 1200 you need 833Mhz to achieve 2 teraflop, with 1280 SPs only 781Mhz is needed 1200 SP at 1 Ghz equals 2.4 teraflops.

 

[leoneazzurro]

I would dare to say that 5870 needs to be 20% faster then 5850, but I doubt it does need to clock 20% higher or have 20% more TFlops for that.

The combination of GPU and Memory just needs to be 20% slower. Now if you think about a 256 Bit part, there is one obvious solution to make the part slower and cheaper for the partners.

I said "At least" 20%. 4870 was 20%-30% more with 20% more clock AND GDDR5, for instance.
Yes, a crippled memory bus + GDDR5 can do the job - but only in certain cases (bandwidth limited situations or, if RBE are linked to MC, fill rate limited situations), whereas in most of the cases the difference will be given primarily by clock core, if the 5850 is not a crippled part in the shader department.
(and if it will be, this means that 5870 and hence the chip will have more SP )
In the given example (4870 vs 4850) most of the difference at "human" resolutions was due to core clock increase alone, there was a review on hardware.fr where the two parts were tested at the same core clock but with default memory clock and the difference was not very high (5%)

EDIT: of course 192 bit bus + lower RBE count is a little different, however I don't think it will be a problem until raching high resolutions...

 

[leoneazzurro]

You math is off.

With 1200 you need 833Mhz to achieve 2 teraflop, with 1280 SPs only 781Mhz is needed 1200 SP at 1 Ghz equals 2.4 teraflops.

Yes, it's the same I said: 2 teraflops(R5850 performance ) + 20% of that (5870 performance increase with respect to 5850) = 2.4 Teraflops. The clocks are referred to the 5870, which should be higher clocked than 5850 - if they have the same 1200SP number.

 

[Alexko]

http://www.semiaccurate.com/2009/09/03/and-cypress-shader-count/

In a nutshell, Charlie claims Cypress has 1600 shaders and Juniper 800.

 

[mczak]

Yes, it's the same I said: 2 teraflops(R5850 performance ) + 20% of that (5870 performance increase with respect to 5850) = 2.4 Teraflops. The clocks are referred to the 5870, which should be higher clocked than 5850 - if they have the same 1200SP number.

Who came up with 1200 anyway? I'd consider 1280 far more likely.
And yes 785Mhz for the 5850 part and 950Mhz for the 5870 part (which gives you that 20% difference) sounds doable imho (you can buy stock overclocked rv790 with that clock, so even assuming 40nm doesn't gain you anything there we know it's possible). Obviously, these parts would also have differently clocked memory, something like 900Mhz and 1.2Ghz GDDR5 sounds possible (though I've got no idea if that even makes a price difference). And since those parts are a lot faster than HD4870 bandwidth difference could definitely make a difference in performance.
Though a low-clocked 1600SP part could indeed have lower power draw. Not sure that would be really necessary (twice the power draw of rv740 doesn't sound too bad), though I'd assume it would have large OC potential in this case at least .

 

[digitalwanderer]

The number of shaders is irrelevant, these are extreme shaders....

 

[Pressure]

http://www.semiaccurate.com/2009/09/03/and-cypress-shader-count/

In a nutshell, Charlie claims Cypress has 1600 shaders and Juniper 800.

So they could be aiming at 2.4+ TFLOPS for Cypress (which happens to be what the Radeon HD 4870X2 is capable of with 1600 shaders) but with more bandwidth and other improvements.

 

[keritto]

Now you also have to consider their vagueness when they put down 1.6x faster. They put down that it would be 1.6 times faster than the 4870 but said compared to the rv870. This might be true, but when it is released it could be 2x faster at the top and 1.7x faster at the 5850. Wow they are within 10% now since they claimed it was 1.6 times faster.

It's just an estimation based on older releases and presentations x1900xt was 60% faster than x800xt x2900xt was 60% than x1950xt so this could look a like safe bet with 60% that would be eagerly expected. And nothing really in terms of 1.6x SPs.

But shouldn't dx11 be more efficient than even 10.1 so they even dont need 1280 SPs. But in past they need at least 2x of shader power to gain 60% jump in games and now that CS is needed for some fancy stuff maybeeven 2x is not enough anymore or they rally maxime efficiency with new threading levels in dx1.

 

[keritto]

It could be Terry Makedon's idea too, since he has Greek roots and all. I imagine him while writing drivers for the new babies and after seeing what they can do, mumbling to himseld "that's what i am goind to feed you Nvidia, hemlock"!

And MakeItDone is rooted with Greeks just like some DonaHue is rooted with Bavarians. It really sounds like picked up name or maybe an alias. So yep it can be enchanted with ancient Greece. And Envidia is greek oracle or what? Lol what kind of Xena Pop scenario.

 

[keritto]

No. R200 has single internal DAC:

You can see the external DAC here (near the fan, marked by red and blue):

I must say i'm pretty stunned with that. So RV100 (8month pre R200) RV150 (same as R200) and RV250 thats based on R200 after 12month came out with two internal DACs but R200 didn't have it. Pretty weird politics i must say. So did they done that cause they want to differentiate as manyproducts theycould _full_, LE, LELE, SE, VIVO. Something else doesnt come to my mind know. Tnx

add. Is this rigged picture i see and HDCP on this diagram just above DAC :scratch:

 

[keritto]

keritto: It was a bit different:

R100: Single display controller, built-in TMDS/DAC, external TV-out (Rage Theater) required. Independent outputs not possible.

I was talking about RV100 it had dual DACs as RV150 and RV250 had on RV100 was just one TMDS an RV150 they came out with two. RageTheater is neede only for VIVO functions not for TV-out since RV100

RV200/RV250 and SM2.0 Radeons: Dual DC, built-in dual DAC, single TMDS, external TMDS required for 2 independent digital outputs. RV100 was likely similar.

Dual (built-in) TMDS was present on R5xx and above.

What is DC i thought you missed push A for DAC? Since R5xx series we had dual DL (dual-link) TMDS all before we had TMDS but just up to 1920x1080 or SL. First 7800GTX came out with that DL support afair.

Sorry i refer RV200 as RV150 but it's based on DX7+ chips thats simply hardcoded in me So by that diagram you send after R200 also didnt have dual DAC but RV100 and RV150 (RV200 ) had cause you could connect two monitors via D-SUB. And they have built in TV-out. And GF4 was first that has TV-out on chip.