AMD: R8xx Speculation

Ok folks after that post, wrap the recent stuff up throw it into the closest trashcan and let's start all over again

 

That seems a lot of effort to simply press "New thread"

 

Does one of my multi-chip delerium posts count:

http://forum.beyond3d.com/showthread.php?p=1097622#post1097622



Maybe it was Shtal:

http://forum.beyond3d.com/showthread.php?p=1129866#post1129866

Jawed

 

I thought it was the stuff that was leaked from Vijay Pande back then...

 

I had something within this thread in mind along the line: "at the beginning there was light..."

 

Yes, Vijay Pande mentioned 480 SPs

 

Youre right here that even you dont now how much Ever since They created E6/F2 revision of their cpus they dont wanna create too much waste and everything is sold away. And for recent Deneb (C2) core it's pretty much f surprise they dont create athlon with just one core and all L2 cache of the surrounding cores enabled (2MB total ) They certainly wont create two mainstream cores just for fun.
And with all that speculations i see 1600SP (as most probable after my 2048SP speculations fails, cause everyone fence themselves how they need much more transistors just to make DX11 work over DX.10.1)

 

Prepare for a very cryptic month or two ahead.

 

The man has appeared! So Davey boy and hints about R8xx coming this year? This time hopefully AMD gets the heat output and power consumption down, at least in idle mode. Downclock the GDDR5 is a must! A 100mhz 1200SP cores and 100mhz GDDR5 in 256bit should be enough for even HD decoding, no?

There is one thing i dont get about GPU and DX, I just saw Resident Evil 5 benchmarks and it ran notably slower on HD4800 series than GT200 series. Is this a driver bug or something about the programming by Capcom is off? I would thought a DX game should at least run similarly on the same DX class hardware, that would be the purpose of a standard API?

 

The TWIMTBP logo is all over the game.

 

Haven't all modern Capcom console->PC ports run better on NV hardware whilst featuring the TWIMTBP logo? I certainly remember that to be the case with Lost Planet anyway.

 

Street Fighter IV was pretty equal on both. The OS and CPU seemed to be far more limiting factors than GPU for that one.

Haven't been able to run the Resident Evil one yet since I need to turn off Japanese for non-unicode programs. Stupid arsed demo program doesn't let you choose language, but automatically picks one. And I'm not rebooting my machine for just one demo.

Regards,
SB

 

DX capability has nothing to do with how fast the hardware is. GT200 is faster than HD4800 in general so it being faster in RE5 is no suprise. Different architectures favour different engines so the 4800 can be as fast or even faster in rare cases than the fastest GT200's but in most cases its slower, because its a slower (but much cheaper) architecture in general.

The fact that they are both DX10 (or 10.1 in the 4800's case) simply means that they support all the same hardware features, not that they implement them at the same speed.

Hence why an 8400 isn't as fast as a GTX 285

 

Exactly even 80SP @200MHz is more than enough for DHD (1080p) decoding so 1200SP @100MHz shoud be sufficient for 4 simultaneous QHD (1440p) streams . But maybe we finally saw UVD engine that works properly or they disband it in favor of finally reorganized Stream project. Does anybody knows where that simultaneous multiple overlays featured in X1900 series disappear from drivers, and why dont we see that in newer products ever since?
We gradually lost features whether they really function neat as that overlays or be an great HD decoding engine like UVA/UVD and being picky about the streams they want to decode. And yes it would be great to see a really green power down card from red monster. And some real alternative to Hybrid SLI on newer chipsets and gpus would be great (finally)

 

to clearly put down things, you realize that video decoding is done on a DSP, not run as software on shaders? (but maybe shaders are ujsed for some clean-up or post effects, I don't know)

 

Decompression should be performed by UVD (separated engine), while de-blocking is done mainly via shader core (some part were done via fixed hardware in past, I've no idea if anything changed). As I remember, UVD engine takes slightly over 3 mm² of 55nm GPU (around 10M tranzistors).

Current set of video features is a mix of former AVIVO features + UVD. As for AVIVO [X1 gen]:

 

http://we.pcinlife.com/thread-1211919-1-1.html

translation:


... With the new sources, huh, huh, but not the NDA sources, so true unknown

RV870 is a package with two RV840, Sideport enhanced high-speed Internet, so that from a practical application, performance-driven point of view including the RV870 looks like a GPU, the RV840 specifications for the double 1600SP/256bit, A11 XT version of 3Dmark Vantage at least P17000 +. Single RV840 core area is similar to RV670, but the impact of performance 128bit between 4850 and 4890 between, RV870 certainly can not go beyond the performance of the legendary two-out 512 of the GT300, but the poison code-named R800 (dual RV870) victory sure.

Feedback points:
1. A true strategy in mind and N have significant differences in mind, N core in mind, A mind is a multi-GPU technology, the Internet challenge
2. 256bit the GDDR5 can not be on the performance of



Separate downstairs replies:

In fact, the specifications of RV870 is the first yellow very violent ..... 1920SP ... 384bit .... .... unfortunately fail to live up to TSMC


TSMC not able to do not come, but the immature enabled TSMC 40nm temporary reduction ATI specifications, put it bluntly, 40nm is now the core of this case no one who is greater the more serious leakage problem, GT300 difficult to let go? Not because core of what mine ...

 

Are they saying RV840 has 1600SP/256bit or RV870?
If AMD is really going to do MCM for RV870 then how they will approach CF of two R800 cards? That will be 8 chips to handle The only way I can see it if they really managed to overcome AFR issues or did a design which can share 3D pipe between cores
I love and hate rumor-mills

 

God, what translator did you use? Sideport enhanced high-speed Internet? You mean Interconnect and not Internet.

Babelfish comes up with a more decent translation.

 

So the wafer containing 181mm² chips has chips whose spec is slightly beefier than RV740. 800 ALU lanes, 40 TUs, 128-bit bus and an enhanced sideport.

Separately it seems to be saying that the small chip was planned to be 192-bit with 960 ALU lanes, but that is a victim of 40nm woes. AMD, I suppose, could have known about 40nm woes one year ago, but is such a timescale likely to work. The small chip that was demoed just over a month ago would have taped out ~4 months ago+, I guess.

Alternatively, the 192-bit, 960-lane chip merely sounds like a refresh part that will be scheduled for March 2010, or something...

2 clusters more than RV740 is only about 11-14mm² I reckon, leaving 30mm² for D3D11 changes + enhanced sideport. Does RV740 have a sideport? Seems unlikely...

If this MCM is enhanced-sideport, is it AFR or is it something else? If something else, can the 4-chip board be anything but AFR based?

Jawed