AMD: R7xx Speculation

[Squilliam]

Fudzilla : Supposed to use GDDR5

"Samples are being tested"

http://www.fudzilla.com/index.php?option=com_content&task=view&id=5456&Itemid=1

 

[Mat3]

In Crossfire setups, how do the two GPUs divide the work? Alternate frames? If the same frame, how do they divide the vertex work?

For the R700, is it known if it's just going to be another Crossfire type setup, or might we see some real integration between multiple GPUs (i.e. - sharing memory so textures don't have to be duplicated)?

 

[AlexV]

In Crossfire setups, how do the two GPUs divide the work? Alternate frames? If the same frame, how do they divide the vertex work?

For the R700, is it known if it's just going to be another Crossfire type setup, or might we see some real integration between multiple GPUs (i.e. - sharing memory so textures don't have to be duplicated)?

AFR(alternate frame rendering) is the current preferred multi-GPU rendering scheme. In it one GPU gets frame n and the other frame n+1. As to the second part of your question, if the rendering load is shared per-frame(rendering modes like Split Frame Rendering-SFR, Scissor or SuperTiling) vertex work isn't divided, it's duplicated-that's one of the drawbacks, no geometry oomph scaling with those rendering modes.

 

[Jawed]

For the R700, is it known if it's just going to be another Crossfire type setup, or might we see some real integration between multiple GPUs (i.e. - sharing memory so textures don't have to be duplicated)?

These questions have been standing for quite a while now with no sign of an answer I'm afraid.

I expect CrossFire already shares memory twixt multiple GPUs for textures.

Logically there's no reason why a render target can't be shared jointly by two or more GPUs - the memory attached to all these GPUs comprises a single address space, latency hiding within each chip should be able to cope. But, erm, could it work this way in practice?

This style of render target handling may only become practical when blending/z-testing/stencil-testing become shader programs rather than using fixed function blocks.

Jawed

 

[Squilliam]

Gt200, 240-50w TDP, a door or a window for the r700 to claw through and snatch a lead?

AMD seems to have the most energy efficient chip designs right now. What are the expectations for the r700?

I think it's becoming more and more important. My pc is in a closet practically and it gets hot enough with a 7900gtx.

If the r700 enthusiast is under 200w TDP, it will make it my number one contender.

Also I can go for the Intel chipset which is cooler as well.

Edit: Fuddy duddy linky winky:http://www.fudzilla.com/index.php?option=com_content&task=view&id=5519&Itemid=1

 

[turtle]

Looking at it logistically, ATi's part will probably consume more power.

Rv670 is 105W on 55nm.
R680 is 190W, and is Rv670x2, 55nm.

Now, while "R700" (or whatever the single die is named) may be on 45nm, it seems likely (at this point in time) that it will be on 55nm, be it that samples are supposedly out and such. I take this to mean we can't expect much power savings per spec using the same architecture.

I would assume each R700 chip will be an increase of the spec of RV670, with a little pixie dust thrown in such as changing of how AA is done (ROPs as well as shaders, or whatever). Let's say 50% increase in spec, which is realistic. While not an exact science, it would put it at around 150W for the single core part, and (if R680 is any indication) a dual core part around 280ish. Granted R680 has higher clocks than a single 3870 and thus power savings may be more than what we see now if the R700x2 has similar clocks to the single core part.

While whatever the dual core part of R700 is may not be similar to R680, there is only so much wiggle room of how much juice two chips will need.

I believe the single core part will aim for <150W, or one power connector on a pci-e v1 board (75W from slot, 75 from 6-pin), none on a pci-e 2.0 board, as it can provide 150w of juice. This just seems logical to me on a mainstream scale, and doable with a 50% increase in spec, while it might be tight. 140-150W?

I also believe the X2 part will shoot for < 300W, and will achieve wiggle room for higher clocks from combining the two chips on one pcb, like we see in the 3870x2. The 300W number is derived from the pci-e 2.0 spec, as well as the logical amount of connectors (2 six pins on a pci-e 2.0 board or 1 six pin and 1 eight pin on a pci-e v1 board). At best I imagine we see similar results to R680, and about a 20% efficiency (250W) and at worst similar to what we see on R680 through pure wattage differance (20W or so...260-280W?). It's impossible to guess considering the X2 product of the R7xx generation will be completely different than R680, and we don't know how the 2 chips, 1 board scales through the use of bigger chips. I think nearing the cusp of 150w for a single part, and 250-280W for a dually is a reasonable guess though.

Personally, I don't think we can ask for much more (less) than that unless the actual chip technology incorporates a more efficient way to do the same tasks or the architecture is dramatically different than R600.

 

[Shtal]

GDDR5 will help some for R700 to reduce power consumption.

 

[silent_guy]

GDDR5 will help some for R700 to reduce power consumption.

I suppose that's because high speed DRAM has a long history of ever decreasing power consumption?

 

[Shtal]

I suppose that's because high speed DRAM has a long history of ever decreasing power consumption?

What I mean is overall total consumption will be lower for Radeon card with newer generation memory.

 

[Arun]

Actually I've been thinking about that lately. Fundamentally, you have three criterias to look at: power/chip, power/bitrate, power/bit. What GDDR5 certainly improves is power/bitrate, but I've not seen anything that'd confirm whether it actually improves power/bit and power/chip; my understanding is that it would, in fact, makes them worse. I'd be very interested if anyone else had some more serious data on that question.

 

[aca]

Actually I've been thinking about that lately. Fundamentally, you have three criterias to look at: power/chip, power/bitrate, power/bit. What GDDR5 certainly improves is power/bitrate, but I've not seen anything that'd confirm whether it actually improves power/bit and power/chip; my understanding is that it would, in fact, makes them worse. I'd be very interested if anyone else had some more serious data on that question.

It seems to me that the power/chip ratio is not so easy to determine as the power consumption will not be constant. Power/bitrate seems a fair manner to compare, but what do you actually mean by power/bit? Shouldn't that be energy/bit? And then again, a bit produced at higher frequencies will cost more energy...

 

[NocturnDragon]

It seems to me that the power/chip ratio is not so easy to determine as the power consumption will not be constant. Power/bitrate seems a fair manner to compare, but what do you actually mean by power/bit? Shouldn't that be energy/bit? And then again, a bit produced at higher frequencies will cost more energy...

I guess he means (power) / (amount of storage).
So W/bit, W/MiB or W/GiB.

 

[silent_guy]

Actually I've been thinking about that lately. Fundamentally, you have three criterias to look at: power/chip, power/bitrate, power/bit. What GDDR5 certainly improves is power/bitrate, but I've not seen anything that'd confirm whether it actually improves power/bit and power/chip; my understanding is that it would, in fact, makes them worse. I'd be very interested if anyone else had some more serious data on that question.

It's very likely that power/bitrate and maybe power/bit will be better. But if history has shown anything it's that, in the long term, those improvements are seldom used to make a product less power hungry.

It seems to me that GDDR4 was on a downward section of a sawtooth that's sloping up. As far as I can tell from the white paper, GDDR5 has been designed performance. A lower voltage will definitely help a bit to keep things in check, but with transfer rates more than double those of GDDR3, it's a safe bet that overall power consumption will be higher.

 

[Sound_Card]

My hunch on R700 is that it will be using GDDR4 with the return of the 512bit bus consisting of two .055 230mm dice each boasting with a mere 20% performance impovement(no aa/af) over a single RV670. Followed by a spring refresh in 2009 shrunk down to 45nm. Power consumption at around <250w with two 8 pin power connectors. 1gb sounds right for the frame buffer with 512mb offerings for lower segments. Of course, praying that AFR is not needed and the GPU's work as one.

 

[nexus_alpha]

My hunch on R700 is that it will be using GDDR4 with the return of the 512bit bus consisting of two .055 230mm dice each boasting with a mere 20% performance impovement(no aa/af) over a single RV670. Followed by a spring refresh in 2009 shrunk down to 45nm. Power consumption at around <250w with two 8 pin power connectors. 1gb sounds right for the frame buffer with 512mb offerings for lower segments. Of course, praying that AFR is not needed and the GPU's work as one.

I think you are right about the 512 bit interface, reason being is..... in a recent interview when asked why in some titles the min fps is lower in a multigpu setting verses a single gpu setting. It was stated that that when loading textures for a new level it has to be copied x times where x is the number of gpus involved. So to alleviate this issue in the short term for a dual gpu or x gpu card aka r680 and r700 more bandwidth. The rest I think you are wrong..... 45 nm is already sampling not the high end 45nm stuff, but perfect for rv770 but the actual r700 who knows.


Here is a article from late 2006

Edward Chou, marketing director of ATI Technologies Asia-Pacific division, stated that yield rates on 80nm technology production at Taiwan Semiconductor Manufacturing Company (TSMC) remain stable. ATI has been working closely with the foundry, according to Chou, adding that the company is looking to adopt the 45nm node by 2008. ATI began 80nm production at TSMC at the end of 2005 and started production of its 65nm-made chips from the middle of this year.

ATI on October 19 held a press conference to officially announce its Radeon X1950 PRO graphics chip (codenamed RV570), manufactured using 80nm process technology. By the first half of 2007, ATI aims for its complete lineup of its graphics chips to adopt a 80nm node, whose production costs are 10-15% less than the cost for 90nm process technology, Chou said.

 

[nexus_alpha]

Also AMD is being extremly smart right now in that the have reliased mobile is the future. Look at all the sales figures mobile is growing faster then desktop in almost all segments except maybe storage. AMD is competing with intel in this segment and ATI has already beaten nVIDIA in q208 if intial data is to believed (remeber laptop designs are usually ready more than a quarter before.

AMD's most ambitious plans are all mobile ala SWIFT GRIFFIN PUMA and rv770.

 

[Blazkowicz]

Probably ATi is on a new dogma where single chip solutions will/should all fall under RV* class of products and R* would be saved for multi-chip flavours.

yes, but that still is consistent with their dogma of reserving the R for the highest end product.

 

[LordEC911]

My hunch on R700 is that it will be using GDDR4 with the return of the 512bit bus consisting of two .055 230mm dice each boasting with a mere 20% performance impovement(no aa/af) over a single RV670. Followed by a spring refresh in 2009 shrunk down to 45nm. Power consumption at around <250w with two 8 pin power connectors. 1gb sounds right for the frame buffer with 512mb offerings for lower segments. Of course, praying that AFR is not needed and the GPU's work as one.

With recent rumors of RV770, aka single core R700, being roughly 50% faster than RV670, I doubt most of your speculation.

As for the GDDR5 discussion, then how did these articles come up with 30% power savings vs mainstream GDDR3?

Link
Hmmm... says 20% in link but I swore I read 30% somewhere.

 

[Sound_Card]

With recent rumors of RV770, aka single core R700, being roughly 50% faster than RV670, I doubt most of your speculation.

As for the GDDR5 discussion, then how did these articles come up with 30% power savings vs mainstream GDDR3?

Link
Hmmm... says 20% in link but I swore I read 30% somewhere.

My performance guess was just that... a guess. I admit, I'm being a little humble, but also trying to be realistic. 20 to 50% sounds right.

I don't think we well see GDDR5 in use untill early 2009 to the least.

 

[DegustatoR]

Gt200, 240-50w TDP

Am i the only one here thinking that at this stage of "retail availability" of G100/GT200 all we can talk about WRT TDP is the projected envelope? If the chip is designed for 240-250W TDP _max_ it doesn't mean that retail boards on that chip will consume 250W. How much they will consume in the end depends on how good the competition (7x0) will be.