AMD: R8xx Speculation

[FrameBuffer]

a bit OT but I LOL'ed when I read your sig Florin:

Hemlock - When crushed, the leaves and root emit a rank, unpleasant odour

kinda reminds me of the Fermi Paradox*: "The apparent size and age of G300 die suggests that many technologically advanced GPUs ought to exist.
However, this hypothesis seems inconsistent with the lack of observational evidence to support it."

*edited to fit subject at hand ..


sorry couldn't resist

 

[green.pixel]

Good one.

 

[trinibwoy]

They would never do that! It is bad corporate mojo. Now hiring some little teardown house a few blocks from their office to buy one, tear it appart to the chemical level, and make a report which they then buy, that is different. That is OK.

Ah, but what's the shame in that? Imitation is the sincerest form of flattery after all

 

[neliz]

AMD's mobile lineup:

* Broadway XT -> HD5870, 128-bit, GDDR5, 45W-60W
* Broadway Pro -> HD5850, 128-bit, GDDR5, 30W-40W
* Broadway LP -> HD5830, 128-bit, (G)DDR3, 29W
* Madison XT -> HD5750, GDDR5, 20-30W
* Madison Pro -> HD5730, (G)DDR3, 20-25W
* Madison LP/Pro -> HD5650, 128-bit, (G)DDR3, 15-20W
* Park XT -> HD5470, 12-15W
* Park Pro -> HD5450, 10-12W
* Park LP -> HD5430, <8W

HD5800 Series - Gaming Enthusiast Segment
HD5700 & HD5600 Series - Performance Segment
HD5470 & HD5450 - Mainstream Segment
HD5430 - Thin & Light Segment

Broadway: Juniper based, 800SP's for HD5800, GDDR5, GDDR3 and DDR3 1Billion+ Xtors
Madison: Redwood based 400SP's 57/5600, GDDR5, GDDR3 and DDR3 600M+ Xtors
Park: Cedar based 80SP's 80SP's 5400, GDDR5, GDDR3, DDR3 and DDR2 240M+ Xtors.

Broadway/Madison will support 6 Display Pipelines for Eyefinity, Park (as Cedar based) will only have 4 Display Pipelines. DP Ports will most likely be based on mini-DP and not normal DP like the discrete cards.

 

[mboeller]

Park: Cedar based 80SP's 80SP's 5400, GDDR5, GDDR3, DDR3 and DDR2 240M+ Xtors.

Is that the great surprise? Cedar/Park GPUs are only as big as the RV710. Nearly exactly the same transistor budget. Therefore the performance should be the same too, or? Thats a little disappointing.

 

[neliz]

Is that the great surprise? Cedar/Park GPUs are only as big as the RV710. Nearly exactly the same transistor budget. Therefore the performance should be the same too, or? Thats a little disappointing.

Indeed, I myself was hopin for a little more, like 120SP's now the gap is just too big..

 

[mczak]

Indeed, I myself was hopin for a little more, like 120SP's now the gap is just too big..

Considering IGPs got faster the chip seems almost pointless indeed. Should be nice for building cheap, low-profile 4 output cards (or notebooks with those proposed mini-dp ports), but otherwise there seems to be little point for not just using a IGP.
Oh, and definitely no "mobile" Cypress?

 

[neliz]

Oh, and definitely no "mobile" Cypress?

Seems so eh? considering nvidia isn't going to launch anything past GT215/G92 in the mobile space either we've seem to have run into a wall.

 

[Kaotik]

Indeed, I myself was hopin for a little more, like 120SP's now the gap is just too big..

Considering IGPs got faster the chip seems almost pointless indeed. Should be nice for building cheap, low-profile 4 output cards (or notebooks with those proposed mini-dp ports), but otherwise there seems to be little point for not just using a IGP.
Oh, and definitely no "mobile" Cypress?

If I'm not mistaken, the Evergreens are all 16-wide SIMD * 5-wide VLIW? So the chips should all have shader count dividable by 80, and having for example 120 would mean moving (back?) to 8-wide * 5-wide, which might require different things from the drivers compiler compared to other Evergreens?

 

[digitalwanderer]

* Madison XT -> HD5750, GDDR5, 20-30W
* Madison Pro -> HD5730, (G)DDR3, 20-25W
* Madison LP/Pro -> HD5650, 128-bit, (G)DDR3, 15-20W

I can't believe they named a part after my daughter like that, I'm just sincerely touched!

 

[EduardoS]

Park: Cedar based 80SP's 80SP's 5400, GDDR5, GDDR3, DDR3 and DDR2 240M+ Xtors.

One or two SIMDs? Full DX11? How is LDS and GDS? Will be the next integrated graphics?

 

[mczak]

If I'm not mistaken, the Evergreens are all 16-wide SIMD * 5-wide VLIW? So the chips should all have shader count dividable by 80, and having for example 120 would mean moving (back?) to 8-wide * 5-wide, which might require different things from the drivers compiler compared to other Evergreens?

I don't think that makes much of a difference to the driver.
I haven't seen any confirmation yet Redwood is 16 wide, so we only know this for Cypress/Juniper. Doesn't matter though, with 80 shader units I strongly doubt it's got only one cluster as that would imply 4 texture units hence it's probably 8 wide anyway just like rv710.

Seems so eh? considering nvidia isn't going to launch anything past GT215/G92 in the mobile space either we've seem to have run into a wall.

From nvidia, I'm more interested what the fastest fermi-based mobile part will look like... GT200 didn't cut it for mobile, and since the next part down the ladder was GT215 in the GT2xx line with an obvious huge performance gap not many conclusions can be drawn imho.

 

[rjc]

Notebookcheck has reviewed the 5650M:

http://www.notebookcheck.net/ATI-Mobility-Radeon-HD-5650-Graphics-Card-Review.24035.0.html
According to our test the ATI Mobility Radeon HD 5650 can be about classified at the level of the GTS 250M from Nvidia, maybe even above it. While the 3DMark benchmark tests still claim the opposite, it can convince in many games in practice. In many of them the GTS 250M is outperformed, some results stay at its level or slightly below it. Considering that our test sample came with DDR3 memory, the performance with GDD3 memory should be even better.

It includes a die shot.....

(Package is supposed to be the same size, pin compatible with the RV730 package, should be able to figure out die size relative to that)

 

[Tchock]

Wow, if that's just Madison...

5650M probably uses the same amount of power as the 4650M, probs 15-25W. 4670M's on par with the 250M in TDP terms.

looking forward to 5670/5750 + Arrandale.

 

[AnarchX]

Is that the great surprise? Cedar/Park GPUs are only as big as the RV710. Nearly exactly the same transistor budget. Therefore the performance should be the same too, or? Thats a little disappointing.

Same transistor count sounds a bit weird for equal SPs count plus D3D11, GDDR5 and Eyefinity.
Or is Park/Cedar now 80SPs/4 TMUs and has only a 32-Bit MC?

 

[Pressure]

AMD's mobile lineup:

* Broadway XT -> HD5870, 128-bit, GDDR5, 45W-60W
* Broadway Pro -> HD5850, 128-bit, GDDR5, 30W-40W
* Broadway LP -> HD5830, 128-bit, (G)DDR3, 29W
* Madison XT -> HD5750, GDDR5, 20-30W
* Madison Pro -> HD5730, (G)DDR3, 20-25W
* Madison LP/Pro -> HD5650, 128-bit, (G)DDR3, 15-20W
* Park XT -> HD5470, 12-15W
* Park Pro -> HD5450, 10-12W
* Park LP -> HD5430, <8W

I'm a bit put off by the naming convention. They should really just have called the Juniper part for what it is, the Mobility Radeon HD 5770 and so forth for the rest of the GPUs.

 

[Alexko]

I'm a bit put off by the naming convention. They should really just have called the Juniper part for what it is, the Mobility Radeon HD 5770 and so forth for the rest of the GPUs.

I agree that it would have been clearer and more honest, but when Nvidia calls a GT215-based card GTS 360M, what can you do? Mobility Radeon HD 5770 would sound like something similar despite being vastly superior. Besides, Juniper is ATI's best mobile GPU, so it makes sense to give it the biggest number.

 

[mboeller]

Same transistor count sounds a bit weird for equal SPs count plus D3D11, GDDR5 and Eyefinity.

...and it gets even weirder if you think about the rumored performance of the Cedar with up to P~2000. Link: http://gathering.tweakers.net/forum/list_messages/1385125

If Cedar is really so small than maybe the speed is rather high. So Cedar could really be something special with >1GHz

For comparison, the 4550 get a vantage score of ~P1100. (at least in the only review I was able to find which tested vantage with performance preset).

 

[Silent_Buddha]

Interesting...

We're on the eve of turning from 2009 to 2010, I could help but look at the poll that was created for this thread.

Only 44.52% of voter's expected GT300 to launch next year.

Although I suppose if you take paper launches and paper specs, then it could have been one of the earlier choices.

Regards,
SB

 

[neliz]

In the olden years, you could see the names of the votees.. oh how fun that would've been. (maybe still for the threadstarter?)