AMD: R8xx Speculation

[neliz]

We'll see. In any case, they seem to have a large headstart on the other green guys - so IMHO no need to hurry that much.

We should all be thankful to Jen-Hsun Huang, for giving us about half a year of quality time with RV870 before GT300 shows up!

 

[Jawed]

So Charlie reckons...

... ~350mm² Cypress. That's so much larger than RV770 it seems to imply that something's gone wrong with sweet-spot, D3D11 is way costly. How's it supposed to be a sensible price, e.g. $300?

Jawed

 

[LordEC911]

... ~350mm² Cypress. That's so much larger than RV770 it seems to imply that something's gone wrong with sweet-spot, D3D11 is way costly. How's it supposed to be a sensible price, e.g. $300?

Jawed

He actually stated exact numbers, 19mm per side, which is strange.
Why I am thinking that isn't the actual die size, while 300mm2 sounded nice, I'm starting to think it is somewhere in between, 320-325mm2.

 

[no-X]

... ~350mm² Cypress. That's so much larger than RV770 it seems to imply that something's gone wrong with sweet-spot, D3D11 is way costly. How's it supposed to be a sensible price, e.g. $300?

Jawed

Radeon X1950PRO / RV570 was 330mm². I think it was a very successful product, which was the key source of income in the times of delayed R6xx parts.

On my opinion ATi really needs a high-performance GPU, which could justify significant boost of prices. Hypotetical product 50% faster than HD4870 for +50% price wouldn't change current weak economical situation. I think they need a product, which could be sold for at least double price of HD4870/4890. And performance must correspond to its price, or nobody would buy it.

 

[itsmydamnation]

Radeon X1950PRO / RV570 was 330mm². I think it was a very successful product, which was the key source of income in the times of delayed R6xx parts.

On my opinion ATi really needs a high-performance GPU, which could justify significant boost of prices. Hypotetical product 50% faster than HD4870 for +50% price wouldn't change current weak economical situation. I think they need a product, which could be sold for at least double price of HD4870/4890. And performance must correspond to its price, or nobody would buy it.

so long as they have a SKU thats + 100% perf of the 4870/4890 at the same price i'll be very happy. If they have a moster that has a 100% on that as well, i guess they will be able to charge whatever they like

maybe the sweet spot for DX11 is 100nm bigger .

If:

GT300 is 500nm+
both use GDDR5
AMD continue to seem to have better memory management/useage
Nvidia continue with odd buses (384) but AMD stick at 128-256bits so you have 1024 vs 1536 etc

then AMD could still have higher margins and therefore room to price lower then NVIDIA, lots of IF's there

 

[Jawed]

Radeon X1950PRO / RV570 was 330mm². I think it was a very successful product, which was the key source of income in the times of delayed R6xx parts.

http://www.beyond3d.com/resources/chip/119

230mm² though I will admit there are wildly varying figures for the size of RV570 out there. Also it's worth remembering that RV560 was the same chip, destined for X1650XT.

On my opinion ATi really needs a high-performance GPU, which could justify significant boost of prices. Hypotetical product 50% faster than HD4870 for +50% price wouldn't change current weak economical situation. I think they need a product, which could be sold for at least double price of HD4870/4890. And performance must correspond to its price, or nobody would buy it.

Sweet spot is effectively $200 for the Pro version of the biggest chip - HD4850's launch price - though it's worth remembering that HD4850 came out faster than AMD would normally have made it (Dave's guiding hands). You could argue about the value for money of HD4870 at ~50% more cost for ~30% more performance, back when they launched - that's because HD4850 was "too good". What that actually means is AMD sacrificed some margin on RV770. Though volume should have made that up, I guess, in terms of absolute profit.

AMD still needs to hit the $200 and $300 price points. Juniper's going to be ~$100, I reckon. So the Pro version of Cypress needs to hit $200.

Unless there's a chip between Juniper and Cypress.

Or, unless there's something about the architecture that makes Cypress pricing much more palatable, e.g. that it can easily come in 3 speed grades (SE, Pro and XT, as it were).

Jawed

 

[3dilettante]

... ~350mm² Cypress. That's so much larger than RV770 it seems to imply that something's gone wrong with sweet-spot, D3D11 is way costly. How's it supposed to be a sensible price, e.g. $300?

Not knowing the particulars of the product, I don't know.
The sweet spot strategy, at least how I interpret it, is that the design would focus on the mainstream 200-300 segment, and that the economics of the chip's manufacturability would be appropriate for a board of that segment.

The sweet spot could still hold (at least in an abstract sense, a third more area sounds pretty big), even at that die size. It would require some adjustment to other factors in the cost/price formula.

AMD may have achieved or hoped to have achieved some manufacturing cost reduction such that the chip could still be priced appropriately.
This would involve something like a difference in pricing from the foundry, or some kind of significant yield improvement. I'm not sure there's enough slack from TSMC for price, or enough room for yield improvement.

Another factor would be if the chip somehow enabled cost savings elsewhere on the board, or AMD projected that certain components would decline enough in price so that the GPU could cost more without pushing the end product over the line.

Or AMD's sweet spot strategy is good PR when you don't have a halo product, dunno.

As far as blaming DX11 for this, I suppose that could be at fault. I haven't seen commentary that the smaller chips have a similar bloat problem.

 

[no-X]

Jawed: Oops, you are right. My memory gets weak, I confused two different parameters.

 

[Silent_Buddha]

Wasn't there a rumor that TSMC was going to give AMD much better pricing due to the problems they had at 40 nm?

Also could there have been some cost decreases with regards to board costs? IE - the chip may cost more but the boards will be less?

Either way, I won't be a happy camper if they abandon the Sweet Spot and go over 300 USD significantly.

Regards,
SB

 

[LordEC911]

AMD still needs to hit the $200 and $300 price points. Juniper's going to be ~$100, I reckon. So the Pro version of Cypress needs to hit $200.

Unless there's a chip between Juniper and Cypress.

Jawed

I would expect that Juniper would be targeting the $100-$149 range but maybe a lowend Cypress would cover that segment.

Wasn't there a rumor that TSMC was going to give AMD much better pricing due to the problems they had at 40 nm?

Regards,
SB

Yes but some forumites that are supposedly "in the industry" suggested that any sort of deal/discount/reimbursement would be worked into the next contract...

 

[Jawed]

Not knowing the particulars of the product, I don't know.
The sweet spot strategy, at least how I interpret it, is that the design would focus on the mainstream 200-300 segment, and that the economics of the chip's manufacturability would be appropriate for a board of that segment.

I agree with that and the rest of what you say, I just can't think of any significant factors.

The only other thing I can think of is that JuniperX2 will be the $200 product. I think a $200 AFR board would kill sales.

Overall board cost can't change substantially as a 256-bit GDDR5 bus isn't going anywhere. So we're left with power-regulation and cooling - which do promise to be cheaper than HD4850, I guess. But, offsetting such a significantly larger die?

As far as blaming DX11 for this, I suppose that could be at fault. I haven't seen commentary that the smaller chips have a similar bloat problem.

Well, in theory the smaller the chip the more bloat the API-improvements add as a percentage of the overall die, if you work on the assumption that the ALUs don't have much (if any) API-improvement cost, and that the smaller dies have less percentage area of ALUs.

We know 32KB shared memory is coming, so that will add a few percent to the cost of a SIMD. A core concept of D3D11 is getting data into/out-of the ALUs by non-TEX/RBE means (or, if you prefer, "these aren't texels coming in and they aren't pixels going out", they're gather/scatter operations). Some of that will scale proportionally with ALU count, some will scale with MC count and some will be just the "new feature bloat". Scheduling the additionally kernel types, HS and DS, adds cost too, a cost that has a price of entry as well as a scaling element

Historically ATI GPUs offer reduced per-unit performance on the smallest GPUs, e.g. reduced capacity for hierarchical-Z, reduced MSAA performance/capability (e.g. no 8xMSAA), no double-precision, reduced ALU:TEX - in a bid to cut fat, and I guess offsetting API-bloat as well as general architectural entry-costs...

Jawed

 

[digitalwanderer]

Enough of these esoteric details, do we know what colour it is yet? :|

 

[LordEC911]

Enough of these esoteric details, do we know what colour it is yet? :|

Green, Red and Black.

 

[Kaotik]

Enough of these esoteric details, do we know what colour it is yet? :|

I put my money on black high end PCB (X2) and red for the rest

 

[Jaaanosik]

I agree with that and the rest of what you say, I just can't think of any significant factors.

The only other thing I can think of is that JuniperX2 will be the $200 product. I think a $200 AFR board would kill sales.

Overall board cost can't change substantially as a 256-bit GDDR5 bus isn't going anywhere. So we're left with power-regulation and cooling - which do promise to be cheaper than HD4850, I guess. But, offsetting such a significantly larger die?


Well, in theory the smaller the chip the more bloat the API-improvements add as a percentage of the overall die, if you work on the assumption that the ALUs don't have much (if any) API-improvement cost, and that the smaller dies have less percentage area of ALUs.

We know 32KB shared memory is coming, so that will add a few percent to the cost of a SIMD. A core concept of D3D11 is getting data into/out-of the ALUs by non-TEX/RBE means (or, if you prefer, "these aren't texels coming in and they aren't pixels going out", they're gather/scatter operations). Some of that will scale proportionally with ALU count, some will scale with MC count and some will be just the "new feature bloat". Scheduling the additionally kernel types, HS and DS, adds cost too, a cost that has a price of entry as well as a scaling element

Historically ATI GPUs offer reduced per-unit performance on the smallest GPUs, e.g. reduced capacity for hierarchical-Z, reduced MSAA performance/capability (e.g. no 8xMSAA), no double-precision, reduced ALU:TEX - in a bid to cut fat, and I guess offsetting API-bloat as well as general architectural entry-costs...

Jawed

I am not sure if it's that much significant but wasn't Cypress supposed to be an MCM? If that's the case then yields are going to be better.

 

[Kaotik]

I am not sure if it's that much significant but wasn't Cypress supposed to be an MCM? If that's the case then yields are going to be better.

Depends on rumor you believe in.
One rumor said that Cypress = highest end, the "X2 product", using 2x Redwood on MCM, Redwood being the "RV870" at ~300mm^2, Juniper "RV830" at ~181mm^2. And then it leaves the "Cedar" and "Hemlock" names for "RV810" and then we end up with 1 codename too many.
Oh, and now the latest rumormongers spout out that there would be "Trillian", X3-card

 

[Tchock]

Enough of these esoteric details, do we know what colour it is yet? :|

I did lodge multiple complaints to ATI though, requesting them to use pink instead.

 

[digitalwanderer]

That would be simply faboo!

 

[keritto]

Hasn't the laser cutting been used quite a lot actually? I think Xx00 & 6x00 series were the last which could be softmodded, the later ones not due laser cutting?

These are laser cut traces from chip to pins inside the chip package after testing. Not to do anything with technology used for cutting silicon wafer itself.

I myself cant believe that they still use diamond cutters on silicon. Like its some burnt clay board. Seems pretty retarded way cause pressurized water cutting doesnt produce heat and cutting residues doesnt vaporize that could damage silicon.

From top to bottom:
Hemlock(R800), Cypress(RV870 300mm2), Cedar(RV840/~225mm2), Juniper(RV830/180mm2) and Redwood(RV810/120mm2).

As we lost ourselves in botanic garden i'd make a wild guess that according to its name Hemlock should be an IGP part as it represents small plant and Redwood a really wide style boulevard chip

 

[LordEC911]

As we lost ourselves in botanic garden i'd make a wild guess that according to its name Hemlock should be an IGP part as it represents small plant and Redwood a really wide style boulevard chip

Not all hemlocks are small bushes... some are pretty large trees.

Hmmm... IGP named Mistletoe?

Wiki FTW-
Hemlock
Redwood
Cedar
Juniper
Cypress