AMD: R9xx Speculation

[no-X]

jaredpace: I don't believe they would make 256bit part with 960 SPs, especially if 6GHz modules are available.

As for Cayman, 6GHz modules would boost bandwidth by 25% while staying with 256bit bus. I think that could be enough for 1920 SPs / 32 ROPs part.

But how many TMUs? There are 3 basic alternatives:

1. SI will change ALUs from 5D to 4D, ratios will stay the same. That would mean 640 SPs (160 4D ALUs) and 40 TMUs per block. Bart would consist of two blocks (1280 SPs + 80 TMUs), Cayman of three blocks (1920 SPs + 120 TMUs)

2. ALUs will be 4D + ALU:TEX will be boosted from 4:1 to 5:1 to compensate. It would make the GPU smaller and maybe more effective in terms of performance per area, but branching granularity would be worse (and batch size too, if I am not mistaken). Single block would consist of 640 SPs (160 4D ALUs) and 32 TMUs then. Bart ~ 1280 SPs + 64 TMUs, Cayman ~ 1920 SPs + 96 TMUs.

3. ALUs will stay 5D, ratio unchanged, block will be smaller: 640 SPs (128 5D ALUs) + 32 TMUs. Bart ~ 1280 SPs + 64 TMUs, Cayman ~ 1920 SPs + 96 TMUs

 

[Kaotik]

The Eyefinity improvement will be the removal of the current requirement to run Eyefinity in anything over 1920x1080.

Huh? You can run Eyefinity at lower resolutions than that currently, too. There's several users at MuroBBS running at 5040x1050 (3x 1680x1050)

 

[neliz]

Huh? You can run Eyefinity at lower resolutions than that currently, too. There's several users at MuroBBS running at 5040x1050 (3x 1680x1050)

trying to run over 3x 1920x1200 limits you to use the DP port, that's the limit I'm talking about "removing."
Sapphire's Flex model is only achieving it's "do not want DP!" because one port is converted to Single-Link.

 

[Kaotik]

trying to run over 3x 1920x1200 limits you to use the DP port, that's the limit I'm talking about "removing."
Sapphire's Flex model is only achieving it's "do not want DP!" because one port is converted to Single-Link.

Oh, understood you wrong. That would mean adding 3rd TMDS transmitter on the chip, then.
And the DP limitation is always there regardless of resolution on current HD5-series, unless you count adapters as a way to get around it?

 

[iMacmatician]

1. SI will change ALUs from 5D to 4D, ratios will stay the same. That would mean 640 SPs (160 4D ALUs) and 40 TMUs per block. Bart would consist of two blocks (1280 SPs + 80 TMUs), Cayman of three blocks (1920 SPs + 120 TMUs)

I forgot about the blocks thing. If so then Turks could easily be 1 block of 640 SPUs (and Caicos some fraction of that).

And given the TDP ranges from that report I wonder if Antilles could be two Barts rather than two Caymans.

 

[no-X]

iMacmatician: Too many things are unknown. Will be performance per SP ratio better? Will the number of TMUs per block change? These variables can place Bart's performance between 5830/5850 as well as at 5870's level. Making a dual-GPU board makes sense only if it would be faster than HD5970, which means at least 5870's performance per GPU... dual Bart would be likely cheaper to produce than dual Cypress (Hemlock).

 

[Dave Baumann]

Sapphire's Flex model is only achieving it's "do not want DP!" because one port is converted to Single-Link.

Sapphire have put a DP reciever/TMDS transmitter chip on the board. Effectively they have put an active adapter directly on the PCB.

 

[LordEC911]

I forgot about the blocks thing. If so then Turks could easily be 1 block of 640 SPUs (and Caicos some fraction of that).

And given the TDP ranges from that report I wonder if Antilles could be two Barts rather than two Caymans.

Or more likely, IMO, Antilles is two cutdown Caymans, ala 5970 being a full 5870 at 5850 clocks.
Seeing as how the highend Cayman is probably ~200w, something ~160-170w range is needed for the dual GPU card to be under 300w.

I would be surprised if the highend Cayman really has a 384bit bus, ~250-288Gbps seems a bit much. 256bit w/ 6Gbps means bandwidth is just under 200Gbps, ~20% more over Cypress.

 

[GZ007]

I would be surprised if the highend Cayman really has a 384bit bus, ~250-288Gbps seems a bit much. 256bit w/ 6Gbps means bandwidth is just under 200Gbps, ~20% more over Cypress.

If radeon the 256bit 4850 could be cheaper than the 128bit 5750 , than the "expensive" 384bit buss doesnt seems to me so expensive at all.(with the 5800 margins even less so)
What is the cost of 4 aditional mass production gddr5 memory chips ?

 

[Scrat]

If radeon the 256bit 4850 could be cheaper than the 128bit 5750 , than the "expensive" 384bit buss doesnt seems to me so expensive at all.(with the 5800 margins even less so)
What is the cost of 4 aditional mass production gddr5 memory chips ?

The additional cost is related to additional die area and development costs, not just those 4 chips.

 

[ferro]

The additional cost is related to additional die area and development costs, not just those 4 chips.

RV770 (260mm2) is larger than Juniper (166mm2). The total lack of DX11 competition for the 5750 probably plays an important role in its current inflated pricing.

 

[no-X]

The 4850's are likely old stock, which is cleared with imperceptible (if any) margins. It doesn't mean, that 5700 is overpriced. 5700 is often equiped by 1024MB memory (4850 had typically 512MB) and it must cover manufacturing and development expenses.

It's interesting, how many people care about expenses and selling price today... Just a few years ago nVidia sold 190mm² G71 for 6-time higher price compared to 170mm² 5770 and nobody cared

 

[Alexko]

On the other hand, no too long ago, ATI was selling HD 3850s (192mm² die, 256-bit bus) for way less than 100 € in Europe…

 

[ferro]

The 4850's are likely old stock, which is cleared with imperceptible (if any) margins. It doesn't mean, that 5700 is overpriced. 5700 is often equiped by 1024MB memory (4850 had typically 512MB) and it must cover manufacturing and development expenses.

Just for fun (my apologies that it's terribly off topic):

Price development for the Sapphire HD4850/512 since its introduction:

Price development for the Sapphire HD5750/512 since its introduction:

Blue is the average price. One goes down, the other goes up, they meet after about 10 months. That's the difference between competition and no competition, and has nothing to do with die size, memory bus width, and so on.

 

[no-X]

Alexko & ferro: They weren't supply constrained with 80nm and 55nm wafers. 40nm process is the most expensive manufacturing node of last years and TSMC reputedly raised its price at the beginning of this year. Memory modules got more expensive, dollar is stronger... Almost every factor, which can affect selling price, changed at the same way during this year...

 

[caveman-jim]

RV770 (260mm2) is larger than Juniper (166mm2). The total lack of DX11 competition for the 5750 probably plays an important role in its current inflated pricing.

How does it have inflated pricing if it has no competition? Am I misunderstanding you?

 

[ShaidarHaran]

How does it have inflated pricing if it has no competition? Am I misunderstanding you?

It has inflated pricing compared to the previous generation product which can be purchased for a similar price.

 

[Gipsel]

And given the TDP ranges from that report I wonder if Antilles could be two Barts rather than two Caymans.

According to the latest Catalyst 10.8 Antilles is Cayman based.

By the way, the driver reveals also the relations between marketing and engineering Codenames and all of the new "Southern Islands" GPUs have the prefix "NI" to it, e.g. "NI Cayman", "NI Barts", or "NI Caicos". So it appears SI is really NI in 40nm.
And some additional thing is that the mobile versions will have Pro, XT, and LP versions, whereas one has XT, Pro and sometimes LE versions for the desktop. And also the mobile chips sometimes appear as Gemini (X2?) variants.

 

[no-X]

It makes sense. ATi has never released 2 256bit GPUs for the same product line at the same process. It implies, that Bart was initially 128bit part, probably a direct replacement of Juniper, around 170mm² at 32nm. 170mm² GPU enlarged (what's the opposite of shrink? ) to 40nm would be 265mm² large. Maybe slightly larger due to the change of memory bus or slightly smaller if the architectural changes compensated it.

I'm quite curious if this GPU will be competitive with 100mm² bigger GF104...

 

[Kaotik]

According to the latest Catalyst 10.8 Antilles is Cayman based.

By the way, the driver reveals also the relations between marketing and engineering Codenames and all of the new "Southern Islands" GPUs have the prefix "NI" to it, e.g. "NI Cayman", "NI Barts", or "NI Caicos". So it appears SI is really NI in 40nm.
And some additional thing is that the mobile versions will have Pro, XT, and LP versions, whereas one has XT, Pro and sometimes LE versions for the desktop. And also the mobile chips sometimes appear as Gemini (X2?) variants.

Where did you find these in Cat 10.8's? the .inf's for sure don't have Cayman nor Barts nor Caicos - or NI anything for that matter
(edit: vista/win7 x64 .inf's anyway)