AMD: R8xx Speculation

[Shtal]

RV770 were 2.5x more power then RV670.....

lol: Is it possible to predict RV870 will also have 2.5x more power then RV770.

SPECULATION
ATI RV870
40nm
(480 *5D) = 2400 Stream Processors
120 TMU's
32 ROP's
354m㎡
512bit GDDR5 memory

 

[CarstenS]

Don't forget the 60% increase in clock rate, plus shader hotclock and of course, the boost to 80 fat ALUs per SIMD - so you can just promote everything to FP64 and do not have to make a living on the obsolete FP32.

One of the above was actually serious. Go figure...

 

[rpg.314]

RV770 were 2.5x more power then RV670.....

lol: Is it possible to predict RV870 will also have 2.5x more power then RV770.

SPECULATION
ATI RV870
40nm
(480 *5D) = 2400 Stream Processors
120 TMU's
32 ROP's
354m㎡
512bit GDDR5 memory

I am skeptical about this. AMD's strategy is to make mid range gpu's. 354 mm is closer to high end of moderate than lower end. (nvidia wont mind doing that thought, they thought even 576 mm is ok and worth $650, :smile

I would imagine somewhat smaller die size. In the 230 - 270 mm range. Having said that, rv8xx 's successor will in all probability be made on 32 nm as TSMC is said to get 32 nm up and running by end of next year. So it might actually end up around 300mm2.

As for 512 bit memory, My gut feeling is that they wont do it.

Just my 2 cents

 

[hoom]

'Gut feeling'?
They had to throw in 1/5th of the final processing power of RV770 (200GFLOPS worth!) just to fill space because they were pad limited for a 256bit bus.

 

[rpg.314]

I gotta ask this. What is meant by pads? And how do they limit you?

'Gut feeling'?

'Gut feeling' - They didnt go for 512 bit bus with rv770 when nvidia did. Their must be some advantages to it (256 bit + gddr5) over (512 bit + gddr3). I mean while they might increase the bus width, I am doubtful they'll double it outright.

 

[trinibwoy]

I gotta ask this. What is meant by pads? And how do they limit you?

The pads are the connections between the die and the package. They have a certain size and cannot be made smaller. So for a given number of connections you need a certain minimum die size to accomodate all of them. The original RV770 design was too small to accomodate a 256-bit bus so they had to add more SIMDs to fill it out a bit and make it big enough. It's a really nice problem to have.

 

[CarstenS]

The original RV770 design was too small to accomodate a 256-bit bus so they had to add more SIMDs to fill it out a bit and make it big enough. It's a really nice problem to have.

I'm having a hard time believing this, actually. After all, the RV670 was quite a bit smaller and had a 256 Bit interface as well. Or did GDDR5 require the pads to become bigger (as clocks increase)?

 

[fellix]

I'm having a hard time believing this, actually. After all, the RV670 was quite a bit smaller and had a 256 Bit interface as well. Or did GDDR5 require the pads to become bigger (as clocks increase)?

Probably... At last, the only hint here is that a GDDR5 device is up with 34 pins per chip from GDDR3, if that matters. And the CF side-port thingy eats some more real estate there, too.

 

[nyt]

Probably... At last, the only hint here is that a GDDR5 device is up with 34 pins per chip from GDDR3, if that matters. And the CF side-port thingy eats some more real estate there, too.

So the 4850 could have been equipped with a smaller 'RV760' having 600SPs, no CF/GDDR5 support, because it wouldn't be pad limited using the same size as RV670?

 

[fellix]

I think that won't fit into the ATi's strategy, e.g. few ASIC designs satisfying the widest range of board SKUs. The diversification of the product line is pushed outside the chip, to put it that way:

* cheap board/memory (4850)
* performance design with a new fast memory (4870)
* boutique high-end (4800 X2)

It's all around a single GPU, not even bothering with fused parts, like the 8800GT is.
The same ordering is valid for the previous HD3000 line.

 

[turtle]

After a night out at the pub, I came home and felt the gpu speculation flow through my fingers as if from a higher power.

I thought I'd revive this thread by throwing something at the wall and seeing what sticks, and because when you come home by yourself at 3am, there's nothing better than gpu speculation.


我的中国朋友，

这些都不是真正的规格，但只针对投机的交谈。



GTX-300 - ~200mm, 40nm


1. Nvidia GTX 380

480 sp
2GB GDDR5
834/2085/4700
64 ROPs
160 TMUs
3 Teraflops
256-bit
150+ Gbps
<225W

$449


2. Nvidia GTX 360

384 sp
1536MB DDR5
780/1950/4200
56 Rops
128 TMUs
2.25 Teraflops
192-bit
100+Gbps
<150W

$349

3. Nvidia GX2 380

960 sp
834/2085/4700
64x2 ROPs (128)
160x2 TMUs (320)
6 Teraflops
256-bit x 2
150gbps+ x 2 (300gbps)
< 300W

$699

RV870 - ~140mm2 (similar size to Rv730) 40nm

1. ATi 5800 series

12 arrays
512MB/1GB GDDR5
192x5 (960 sp)
c650mhz/782mhz
m4700mhz/5700mhz
16 ROPs
48 TMUs
1.25/1.5 Teraflops
128-bit
75gbps/90gbps
<75W/150W
512MB-1GB GDDR5
$179/$219


2. 5800 X2 series (R870 dual core w/ shared memory via sideport)

24 arrays
1GB/2GB GDDR5
384x5 (1920 sp)
c650/782mhz
m4700mhz/5700mhz
32 ROPs
96 TMUs
2.5/3 Teraflops
128-bit per chip
150gbps/180gbps
<150W/<225W


$299/$399

3. 5800 X4 series (R870 dual core package x 2 connected with PLX chip?)

24x2 arrays (48)
2x1GB/2x2GB
384x5 x 2 (3840 sp)
c650/782
m4700mhz/5500mhz
32x2 ROPs (64)
96x2 TMUs (192)
5/6 Teraflops
128-bit per chip, each package can share mem
150gbps x 2 (300gbps)/ 175Gbps x 2 (350gbps)
<225W / < 300W
2GB/4GB

$499/$599

4. RV740 ~70mm2. Essentially 3800 series performance using a 128-bit controller and gddr5...

HD5670

64x5 (320 sp)
512MB/1GB
c800mhz/1ghz
m4000mhz/5000mhz
16 ROPs
16 TMUs
512Gflops/640Gflops
128-bit
64gbps/80gbps
<45W / <75W

$99/$129

 

[Tchock]

Let me guess, PCInlife.

No other site has such optimistic lopsided people!

 

[Kaotik]

Indeed, not to mention that who ever wrote 'em apparently skipped math in school
Ye, sure, <225 could be under 150W too but looking at the rest of it, it would be reported as <150W then, and <225+<225 definately ain't 00W in this case when everything else says nothing has been scrapped / downlocked or anything

 

[Tchock]

Plus they got their info wrong.

RV730 already IS RV670 performance compute-wise. RV740 from what I've heard is aiming to be much more of an ambitious midrange at 40nm.

 

[Wesker]

I've got a small, candle light sized flame of hope that R800 will follow a monolithic, single chip design.

Can I just say that I hope ATI's 5000 series doesn't go the way of Nvidia's 5000 series.

 

[_xxx_]

I gotta ask this. What is meant by pads? And how do they limit you?

Pads at the bottom of the die connecting it to the outer world, they just fill up all the available space. For more pads they would have to increase the area, making the whole thing way more costly.

'Gut feeling' - They didnt go for 512 bit bus with rv770 when nvidia did. Their must be some advantages to it (256 bit + gddr5) over (512 bit + gddr3). I mean while they might increase the bus width, I am doubtful they'll double it outright.

They just relized that 512-bit would do nothing significant for this chip except raising costs. Thus the use of faster memory which gave them enough bandwidth to avoid the need for a wider bus. Finally they started learning something from history.

 

[Ailuros]

NV would have IMHLO some serious brain damage if they'd arrive with something that sports 160TMUs and on top of that twice as much for a GX2. TMU can of course mean many things, depending on capabilities, but the majority of that GT3x stuff sounds like utter nonsense. Same goes for 64 ROPs@256bit.

What's a GT3x supposed to be anyway? NV's D3D11 core or what? If yes then I'd be mighty careful even thinking of any ROPs as we know them today heh....

 

[Chris123234]

GTX-300 - ~200mm, 40nm

Nvidia GTX 380 vs. Nvidia GX2 380
stats---------------------stats x 2
<225W-------------------< 300W




ATi 5800 series vs. 5800 X2 series
stats---------------------stats x 2
<75W/150W----------<150W/<225W

Why would doubling the the larger nVidia chip result in a much smaller percentage of power usage increase than doubling the smaller AMD chip??

 

[BRiT]

Why would doubling the the larger nVidia chip result in a much smaller percentage of power usage increase than doubling the smaller AMD chip??

Because in speculation land, anything is possible.

 

[Kanyamagufa]

Because in speculation land, anything is possible.

Hahaha, that was pure brilliance. Serious comedic points there, my friend.