AMD: R8xx Speculation

[MarkoIt]

Errrmmmm...why is there still a discussion of the tech specs of a launched product? Didn't the review sites already said it has 256bit bus? Do you believe there are some kind of hidden surprises? I mean ok speculating before the launch is one thing, but speculating after 100 reviews is another!

Could all this be indicative that there is too much wishful thinking, because people think that ATI underdelivered and they are waiting for some kind of miracle? And no, I am not suggesting that they underdelivered, I am just wondering if people think so (I do have an opinion on the subject, but this is not the point of this post).

I do, gaming performance wise. Only 40% over GTX285, it's not enough for a card with those specs. It out of question that G300 won't top it, even if it is a doubled GT200.

 

[flopper]

I do, gaming performance wise. Only 40% over GTX285, it's not enough for a card with those specs. It out of question that G300 won't top it, even if it is a doubled GT200.

isnt the card for dx11?
If you see 40% with dx11 path, its 80% faster than previous gen.

 

[I.S.T.]

Myself, I was expecting a G80esce jump, given they decided to go nuts with it. As it stands, we got less than that. It's still a good card, mind you, a very good one. However, it's not as good as I was hoping.

I hope GT300 lives up to that, or is at least fast enough over the 5870 to make it a better product. I have my doubts, though...

 

[XMAN26]

I'm just bookmarking this post for the G300 launch. If the 5870X2 tops the G300 then its a terrible product from XMAN26's perspective.

Like twisting what people say?

Other than the BF patch, what other DX11 games are there? Also when did MS release DX11 for download? The games I play are DX10 or 9. They perform just fine for me with my setup, to gain 20-25% performance for 380 bucks for DX11 and the hopes that there will be a patch(And the Turbine DDO DX11 patch is very amusing in their slides, DX10 still doesn't work right and they want to patch it to DX11?!) or game(s) released I'll buy to play in DX11 is simply silly. And I'm not the only person who says this. Maybe not here, but many other sites.

Now as I have said and will capitolize it this time so as to help you read it better, IF I OWNED A SINGLE GTX260 OR LOWER OR 4870 OR LOWER CARD OR EVEN SLI/XFIRE ANY MID RANGE(NOW) G92/R740/G80 CARD(S), IT IS AND WOULD BE A VERY WORTH WHILE UPGRADE AS IT IS A HUGE IMPROVEMENT OVER THOSE CARDS. Does that help, can you understand what was typed this time? The 5870 compared to certain cards/setups just isn't/doesn't have that "Gotta get it now" feeling to it for alot of folks so they are waiting for price drops and to see what else is coming.

I'm sure the card line will sell well, Just not as good if it did to the GTX285 what it does to the 4870 for an increase.

 

[no-X]

Anyway, has anybody any idea, why ATi didn't publish die-shot of the GPU?

We haven't seen die-shot, we don't know, how many SIMDs does RV830 and RV840 have (knowing parameters of two differently sized GPU would help us to make an estimation of functional unit size etc.)...

 

[Tchock]

Anyway, has anybody any idea, why ATi didn't publish die-shot of the GPU?

We haven't seen die-shot, we don't know, how many SIMDs does RV830 and RV840 have (knowing parameters of two differently sized GPU would help us to make an estimation of functional unit size etc.)...

1920/384! 1120/192!

Would mean TMUs have to scale up too?

Anyway a slight OTish paragraph:
Am taking literature this sem with my lecturer kinda arguing that form dominates context for natural human responses. A certain post above seems to indicate that- it doesn't matter by most if you're vindicated, you're still using full caps against a single witty sentence.

(Again, just my observations. All in good fun. )

 

[Wirmish]

I believe the difference is more. Someone on this forum made a more accurate calculation and found out that there are about 500M transistors more than what they should be in RV870. But it's also true that AMD/ATi said that the scheduler of RV870 would have been twice as complex as RV770 one. It maybe be the transistor cost of DirectX11.
Though, RV730 has about 500M transistor and it has 320 ALUs, 32TMUs, 2 RBEs and 2 MC + other things.
It could be that RV870 has a 384bit MC and also 4 more SIMD. It would be great for the HD5890, but it would get way too close to the performance area of the HD58x0X2s.

Yes, the difference is more than 242M transistors:
· RV870 = RV770 x 2 ... but with only 4 MC, 1 UVD2, 1 PCIe interface, 1 CFX compositor, 1 tesselator, ...
· Scheduler = twice as complex ... but RV770 x 2 = dual scheduler = twice as complex.

 

[neliz]

192/384 bit busses lead to awkward memory configs. everything up to Q2'10 shows only 128/256 configs.

 

[no-X]

Thanks. It seems, that the RV870 has ~300 more pins (over RV770) for better grounding and eyefinity support...

 

[CarstenS]

fellix: 420 pins total for GDDR5 or extra over GDDR3?


Radeon X800 was introduced as a GPU with 12 pipelines. XT part with 16 pipelines wasn't officially mentioned until introduction of 16-pipelined NV40. There isn't any single die-shot of RV870. I wouldn't be surprised by anything.

IIRC there was an NDA'ed launch of both X800 Pro and XT, so both configurations at the same time.

 

[Arty]

Like twisting what people say?

Other than the BF patch, what other DX11 games are there? Also when did MS release DX11 for download? The games I play are DX10 or 9. They perform just fine for me with my setup, to gain 20-25% performance for 380 bucks for DX11 and the hopes that there will be a patch(And the Turbine DDO DX11 patch is very amusing in their slides, DX10 still doesn't work right and they want to patch it to DX11?!) or game(s) released I'll buy to play in DX11 is simply silly. And I'm not the only person who says this. Maybe not here, but many other sites.

Now as I have said and will capitolize it this time so as to help you read it better, IF I OWNED A SINGLE GTX260 OR LOWER OR 4870 OR LOWER CARD OR EVEN SLI/XFIRE ANY MID RANGE(NOW) G92/R740/G80 CARD(S), IT IS AND WOULD BE A VERY WORTH WHILE UPGRADE AS IT IS A HUGE IMPROVEMENT OVER THOSE CARDS. Does that help, can you understand what was typed this time? The 5870 compared to certain cards/setups just isn't/doesn't have that "Gotta get it now" feeling to it for alot of folks so they are waiting for price drops and to see what else is coming.

I'm sure the card line will sell well, Just not as good if it did to the GTX285 what it does to the 4870 for an increase.

I'm not twisting anything you said ... "why should anyone upgrade from y multi GPU config to x single chip GPU". Objectively this should apply to all cases and we'll see if the G300 can trump a Cypress X2. Or else it would automatically be a FAIL as per your above statement.

 

[Lightman]

Well, AnandTech reported in their review that Juniper is simulated to be 12SIMD design. If Juniper is 181mm2 then why RV870 is 2x bigger and yet still only 20SIMD?
Also Charlie was saying in one of his forum posts that he know true number of shaders in RV870 and it's not what most people think ....

Anyway, the sooner nVidia releases their new card the sooner we find out if AMD is holding anything from us ATM. :smile:

 

[Wirmish]

192/384 bit busses lead to awkward memory configs. everything up to Q2'10 shows only 128/256 configs.

"Initially, NVIDIA has planned for an enthusiast GPU codenamed GT212 but was canned as it was pointless against the more formidable EG Cypress. Instead, NVIDIA concentrate their effort on the next generation DX11 GT300 GPU and is determined to get it out this coming December. Interestingly, sources told us that AIC partners will get to design the GT300 cards themselves, which means, no more boring reference cards. Information we have gathered so far implies that GT300 has more than double the shader processors of the previous generation. Which means, more than 480 shader processors since GT200 has 240. GT300 card will be fitted with GDDR5 memories featuring a 384-bit memory interface."

Link

 

[Wirmish]

Well, AnandTech reported in their review that Juniper is simulated to be 12SIMD design. If Juniper is 181mm2 then why RV870 is 2x bigger and yet still only 20SIMD?

AnandTech reported that Juniper = 14 SIMD.

Cypress = 20 SIMD
Juniper = 14 SIMD ? (-6 SIMD : 1120 ALU)
Redwood = 8 SIMD ? (-6 SIMD : 640 ALU = 2xRV730 = RV740)
Cedar = 2 SIMD ? (-6 SIMD : 160 ALU = 2xRV710)

 

[JasonCross]

I do, gaming performance wise. Only 40% over GTX285, it's not enough for a card with those specs. It out of question that G300 won't top it, even if it is a doubled GT200.

First, it's a fairly new architecture and GT200 isn't. I think it's safe to say, if history is any indication, that you'll see another 10-20% speedup on the 58xx series cards over the next 4-6 months of driver releases.

Second, saying GT300 will be faster is in itself a meaningless metric. Without knowing die size, power draw, cost, etc. it really doesn't matter if it's faster or not.

The most remarkable thing about RV870 to me is that it's roughly 330mm2. It's essentially the same size as G92 was when it was introduced. G92 was introduced, as I recall, in the GeForce 8800 GT, which went for about $220. Yes, it had less RAM and other board cost differences compared to the 58XX series. But it also ended up quickly moving to $179 and then down from there, and G92 was maybe Nvidia's most successful GPU - they used that sucker in tons of cards.

I'm willing to bet AMD has a ton of wiggle room in the 58xx price. As yields improve and stock builds up, and the GDDR5 costs come down, there's no reason the 5870 couldn't be a $200 card.

So the question isn't whether or not GT300 will be faster. I'm willing to bet it will. The question is whether or not it'll be over 450mm2, expensive, hot, low yields, etc etc. It's not a matter of faster/slower, but of what sort of products it'll go into.

Simply making something faster, costs be damned, is not terribly difficult or impressive.

 

[rpg.314]

Did the G92 use latest process (55nm, IIRC) at launch?

By the time, GT300 is ready to launch, 58xx will definitely have improved drivers.

And the real competition fro 5870 would not be with GTX380. I will be against GTX 360.

 

[AlexV]

Simply making something faster, costs be damned, is not terribly difficult or impressive.

The NV30 and R600 (or the V5 6000, or the Parhelia) want to disagree with that. Strongly. It's quite difficult to make fast GPUs, even when you have cost leeway. IMHO.

 

[MarkoIt]

First, it's a fairly new architecture and GT200 isn't. I think it's safe to say, if history is any indication, that you'll see another 10-20% speedup on the 58xx series cards over the next 4-6 months of driver releases.

Second, saying GT300 will be faster is in itself a meaningless metric. Without knowing die size, power draw, cost, etc. it really doesn't matter if it's faster or not.

The most remarkable thing about RV870 to me is that it's roughly 330mm2. It's essentially the same size as G92 was when it was introduced. G92 was introduced, as I recall, in the GeForce 8800 GT, which went for about $220. Yes, it had less RAM and other board cost differences compared to the 58XX series. But it also ended up quickly moving to $179 and then down from there, and G92 was maybe Nvidia's most successful GPU - they used that sucker in tons of cards.

I'm willing to bet AMD has a ton of wiggle room in the 58xx price. As yields improve and stock builds up, and the GDDR5 costs come down, there's no reason the 5870 couldn't be a $200 card.

So the question isn't whether or not GT300 will be faster. I'm willing to bet it will. The question is whether or not it'll be over 450mm2, expensive, hot, low yields, etc etc. It's not a matter of faster/slower, but of what sort of products it'll go into.

Simply making something faster, costs be damned, is not terribly difficult or impressive.

AnandTech reported that Juniper = 14 SIMD.

Cypress = 20 SIMD
Juniper = 14 SIMD ? (-6 SIMD : 1120 ALU)
Redwood = 8 SIMD ? (-6 SIMD : 640 ALU = 2xRV730 = RV740)
Cedar = 2 SIMD ? (-6 SIMD : 160 ALU = 2xRV710)

If Juniper is 14 SIMD, 24 Rops and has 3 MC (192bit), it SHOULD be about 30-40% smaller than RV870, but it's about 80% smaller.. (330 vs 181)
And also performance wise, there is something wrong. Juniper is born to replace the HD4800 series, and it will probably perform slightly better (10-20%). How can be Juniper able to top/match RV770 in about 181 mm^2 and RV870 be only 50% faster, when it's actually over 80% bigger then Juniper?

 

[Kaotik]

Did the G92 use latest process (55nm, IIRC) at launch?

G92 started on 65nm process, G92b was 55nm (this was after the first renaming round, 9800-series cards use both G92 and G92b, while 8800's used G92 (or G80 in case of GTX/GTS320/640)

 

[Wirmish]

How can be Juniper able to top/match RV770 in about 181 mm^2 and RV870 be only 50% faster, when it's actually over 80% bigger then Juniper?

Maybe the RV870 is CPU Limited.


3DMark06, single HD5870, Phenom II 965BE @ 6 GHz -> 30423 3DMarks