AMD: R8xx Speculation

[DegustatoR]

You plan for what you know you can plan for - costs are known, or at least can be relatively well predicted, so you can plan a strategy around that. Thats what the "sweet spot" strategy is about; the $200-$300 bracket at launch (with the knowledge that it will scale down afterward).

I'll try to explain again (maybe it's my English): while you may plan for some cost, $200-$300 bracket at launch for example, you have absolutely no guarantee that this will be the market cost of you solution in the end. The market may force you to lower your costs to $100-150 if the competition have better products in the same $200-$300 bracket.
That means that there is no sweet spot strategy, there are only current market conditions with sweet spots always moving and changing. Since a GPU can't be made in a week you can't possibly know if your chosen sweet spot for it will end up being sweet or bitter.
With RV770 you've got lucky, with R600 you missed the spot pretty significantly. There is no guarantee that you won't miss it again with RV870. No strategy can guarantee something like that.

 

[PatrickL]

You seem to confuse failure to have a good strategy and no strategy....

 

[neliz]

With RV770 you've got lucky, with R600 you missed the spot pretty significantly. There is no guarantee that you won't miss it again with RV870. No strategy can guarantee something like that.

That makes it sound like no-one has a strategy. Apple just got Lucky with the iPhone, Microsoft just got Lucky with Windows and intel just got lucky with Core2.

 

[no-X]

DegustatoR: If I understand well, key of the sweet spot strategy is to use majority of development resources to create primarily a GPU for the ~$250-300 (250-300mm²) segment. The important think is, that if you develop primarily a GPU for this segment, final performance/mm² can be better, than if you develop architecture suitable for ~$600 / 600mm² GPU and use a cost-down derivate for the ~$300 segment.

Every big (compared to competition) GPU was followed by unsuccessfull mainstream products: R5xx->RV530, R6xx->RV630, G80->G84, NV30->NV31. But almost all smaller GPUs including cost-down derivates were more successfull, despite the high-end part was criticised: G71 (successful) -> G73 (succesful) -> dual G71 (failed), G92 (successful) -> G94 (succesful) -> dual G92 (failed). RV770 (successful) -> RV730/RV740 (reviewers are happy with it, it will be probably a successful product) -> dual RV770 (not great, but more successful than previous multi-GPU products).

This strategy seems to be better - even if the enthusiast product fails, the most important segments are OK.

 

[trinibwoy]

The important think is, that if you develop primarily a GPU for this segment, final performance/mm² can be better, than if you develop architecture suitable for ~$600 / 600mm² GPU and use a cost-down derivate for the ~$300 segment.

Yep, that's the hypothesis. But one that doesn't have enough empirical evidence to back it up. We've only had one round of this and it's not only die-size that matters. How do you discount the transistors spent by Nvidia on CUDA for example when doing this sort of analysis based on gaming perf/mm^2?

Every big (compared to competition) GPU was followed by unsuccessfull mainstream products: R5xx->RV530, R6xx->RV630, G80->G84, NV30->NV31.

Those failures were due to architectural deficiencies, not die size. G84 might have been unsuccessful from a retail/enthusiast standpoint but it still sold well. Lately when it comes to Nvidia hardware there seems to be a big disconnect between (negative) popular opinion and actual sales.

 

[ChrisRay]

Everyone has a strategy. From my understanding of how degustator is talking is you can only make a "Target guess" as to where the market goes. Nvidia/ATI may have a general idea of what the other is doing. But its not like they dont surprise each other either.

But market changes based on whats available. And Nvidia and ATI have both misread the market and have ended up way off target.

 

[Silent_Buddha]

Actually I completely disagree.

What I'm inferring from what Dave has said and from previous interviews...

Manufacturer's have a general idea of what a product will cost to produce. They can also control how much is spent for R&D...

With that you can have a fair grasp of the total cost of an architechture.

In ATI's case with Rv770 they wished to target the sweet spot (as defined internally) 200-300 USD retail bracket. To achieve this, they had an internal target for a product that would only cost them X dollars to produce (manufacturing and R&D inclusive).

I would assume that sweet spot in sales is an inflection point where sales volume + margin is greater than other price segments. IE greater than the sales volume + margin of enthusiast.

I don't think anyone in their wildest imaginings can deduce that R600 or G80 is a "sweet spot" product.

Sure G80 ended up being hugely successful for an enthusiast product, but doesn't make the lions share of profit.

I'm pretty sure GT200 for example wasn't designed with the intention nor the R&D budget and manufacturing costs to compete primarily in the 200-300 price bracket. As witnessed by their initial price offering. And if not for Rv770 they would have most likely had another stellar enthusiast success similar to G80. And had quite good sales with G80 owners moving to GT200 en masse.

Regards,
SB

 

[Kef]

My Prediction:

ROPs: 32
TMUs: 64
SPUs: 1280
Speed: ~750-800Mhz

 

[LordEC911]

My Prediction:

ROPs: 32
TMUs: 64
SPUs: 1280
Speed: ~750-800Mhz

2XRV740

Where is the fun with that speculation?

 

[Dave Baumann]

I don't think anyone in their wildest imaginings can deduce that R600 or G80 is a "sweet spot" product.

These would be more consistent with the (to term it badly) "Monolithic die and trickle down strategy". And this is one of points of the "sweet spot" strategy - traditionally it has taken in the region of 3-8 months for derivative products to come out with this approach, with RV770 we effectively addressed 3 major product segments in a very short order - Performance, with 4850, Enthusiast, with 4870 and then a month or so later, the Ultra Enthusiast segment with the X2.

 

[Blazkowicz]

The 4870X2 was supposed to be the revolutionary MCM too.

They did build the sideport in. I guess it failed to deliver on time or was broken in some way, but they still put some R&D in, and it would be a selling point (hell they can charge like 50€ more for more PCIe lanes and pretty heatsinks on a motherboard)

 

[Pressure]

They did build the sideport in. I guess it failed to deliver on time or was broken in some way, but they still put some R&D in, and it would be a selling point (hell they can charge like 50€ more for more PCIe lanes and pretty heatsinks on a motherboard)

Actually Baumann talks about the sideport in his most recent interview.

By the time the product were to launch they had improved their drivers to an extent where the benefit of the sideport were negligible.

 

[LunchBox]

What do you guys think about this spec?

• GPU: RV870 @ 900MHz / 40nm
• Shader cores: 1200 (240 vec5 cores)
• Memory: GDDR5 @ 1100MHz / 512-bit
• Texture units: 48
• ROPs: 32
• 3D APIs: DirectX 11 / OpenGL 3 / Shader Model 5

it came from this site
http://www.geeks3d.com/?p=3997

 

[Lukfi]

The 512-bit figure is certainly wrong.

 

[crystall]

What do you guys think about this spec?

• GPU: RV870 @ 900MHz / 40nm
• Shader cores: 1200 (240 vec5 cores)
• Memory: GDDR5 @ 1100MHz / 512-bit
• Texture units: 48
• ROPs: 32
• 3D APIs: DirectX 11 / OpenGL 3 / Shader Model 5

it came from this site
http://www.geeks3d.com/?p=3997

1200 ALUs means 15 SIMDs (each one 5x16), that doesn't match well with 48 TUs.

 

[no-X]

It could be 12 SIMDs + 5:1 ALU:TEX.

//ed.: like this:

 

[KonKort]

It are the same specs ATI-Forum and Hardware-Infos posted last week. Only the memory interface is different. 512 instead of 256-bit.

 

[Arnold Beckenbauer]

It could be 12 SIMDs + 5:1 ALU:TEX.

//ed.: like this:

I'd like to see smaller SIMDs. Two or three 8-way-SIMDs per cluster.

 

[Rangers]

If those specs are real, very boring. 50% more shaders and 20% more texture units, plus 100% more rops. Real performance increase would seem to be less than 50% as well as they seem to be trending towards being more texture limited again (although hopefully this time they actually know what theyre doing).

Doesnt seem as ambitious as the supposed GT300 specs. Way early to speculate though.

 

[Jawed]

A comparison of RV730 and RV740 performance would prolly be useful:

http://www.gpureview.com/show_cards.php?card1=579&card2=612


RV740 is:

• 100% TEX
• 200% fillrate
• 200% GFLOPs, using 4:1 instead of 2:1 ALU:TEX
• 160% GB/s

This page shows performance in comparison with HD4670, though I don't know if it was tested with the latest driver:

http://www.computerbase.de/artikel/...adeon_hd_4770/19/#abschnitt_performancerating

excluding 8xAA, that averages to 180% performance. Worst case is 161% in Crysis Warhead. Best case is 200%+ in some games, making me suspect other things.

Seems that focusing on RBEs and ALUs was the right thing to do. Indeed it would be worrying if RV870 doesn't follow this trend with doubled-RBEs. Fingers and toes crossed.

Jawed