The NEXT LAST R600 Rumours & Speculation Thread

Still not official?

But here you see that the low texel-fillrate can broke R600's neck.

 

Rumors are true its coming officialy with catalyst 7.5, what is few days away, so the beta/build version drivers has it too i think.

 

Looks like it was a pretty decent time although I thought more people would have shown up. I'm relieved to see that Kyle looks and acts like a regular guy. Funny how the AMD PR guy hasn't even gotten used to saying "HD" as yet - he started off saying "X"

 

They made buggy nVidia drivers The way that performance scales from GTS to GTX is odd...unless I`m mistaken(too lazy to look now), the GTX isn`t 2x shading power WRT GTS, even though that test suggests it. Very interesting.

 

Kyle isnt a bad guy to meet and talk too. He's pretty forward and reasonable to talk too. But he's definately got the same aura he does on the forums. Extreme confidence in himself. But not a bad guy at all.

 

I know nothing's totally concrete...but i'm supremely unimpressed with the HD2900. Supremely. Un. Im. Pressed. Don't know what else to say...I truly expected more for all this time spent waiting. I'll still buy one though, perhaps two and go for Crossfire. But I was really expecting just a *tad* more

 

These numbers don't suggest that, but the possibility the G80 could be the direct successor of the NV30 due to the SIMD design and perhaps some other flaws.

Correct me if I'm wrong, but the GTS only has 6 out of the 8 G80's SIMD units, which reduces the flexibility even more, thus the 33% lower theorical mathematical power has a chance to become 50%.

If this is exact, nVidia will have won his bet by selling thousands of "NV30 like" DX10 GPUs before any use of the API is made.

 

Comparing NV30 with G80 is like comparing a Pentium 4 with a Core 2.
Especially so when a "feature" suddenly turns into a "flaw", just because it didn't work on a previous and completely different main architecture design...
Ever heard the saying "What doesn't kill us makes us stronger" ?

 

Why do you think it's that bad in dx10?

 

You know, NV30 had lots of wonderfull features... according to nVidia

The SIMD design of the G80 made me think there was a risk performance-wise since day 1 as it isn't that flexible and requires alot of work and cache to be efficient if the instructions are not excessively similar. Removing 2 SIMD units out of 8 removes even more flexibility, leaving only 6 distinct instructions to be handled simultaneously and thus reducing efficiency.

Don't take all this as truth as I'm not stating this is the case, but these numbers suggest it.

 

Maybe the Core 2 could be compared to a Pentium 3 since it belongs to the same pedigree tree?

 

Aha - this seems to be confirming my theory that if R600 is faster anywhere, it will be in geometry shading. Granted G80 still has a lot of space to improve in that area: a recent test that I did indicated that even having *any* geometry shader present (I used an "identity" geometry shader that just passed through the triangles) cut the performance of the shader in *half*.

It will certainly be interested to see... in any case I think the Cubemap examples are a bit contrived to begin with because arguably separately rendering the faces with a good frustum culling algorithm is going to be just as fast or faster except in the most degenerate cases.

What I really want to see is how well R600 runs cascades...

 

He's a very handsome man too.

 

What about G80's SIMD design is inflexible, especially compared to R600's? Instructions need to be similiar? Only six instructions? Reducing efficiency? What?

 

Hold on, you created a geometry shader that did no work and it still halved performance?

 

Well, by that order or thinking, we could trace much or AMD's K10/Barcelona's basic design back to the original Athlon and Athlon XP, right up to the similar L1 and L2 cache size and configuration...

And what do the actual performance numbers suggest when you compare a G80 to a NV30 ?

 

Well I agree he certainly doesn't look like the typical socially castrated tech-geek but handsome? That's pushing it a little too far (j/k) Does look like he works out, although that could just be fat and bad lighting playing tricks on me.

 

8 16-wide SIMD units are far less flexible than 128 independant SP.

Each unit has to execute 16 times the same Op to be efficient, leaving only 8 different branches, and even 6 in the case of the 8800GTS. To do that, you must reorder the instructions (out of order execution, heavy task) and store temporarily unuseable data.

Now, if we consider R600 to have 8 8-wide SIMD units (8 clusters of 5 scalars using the same Ops combination), there are 8 different "paths" too, but up to 40 different Ops could be executed at the same time, reducing the reordering task.
If R600 has 16 4-wide SIMD units, it's even more flexible and the fact each cluster could be used to treat a particular pixel adds a little more too.

 

Well the 8.37

FEAR


1600x1200 FA@16x, FAA@4x SS
2900XT/8800GTS:
min:17/20
med:35/37
max:64/70

1920x1200 FA@16x, FAA@4x SS
2900XT/8800GTS:
min:16/16
med:30/32
max:55/57


2900XT - 8.37- 8800GTS - 158.19 -

1600x1200 FA@16x, FAA@4x no SShadows,
2900XT/8800GTS:
min:20/30
med:59/58
max:124/130

1920x1200 FA@16x, FAA@4x no SShadows,
2900XT/8800GTS:
min:19/27
med:51/51
max:111/104


STALKER 1600x1200 all máximun, AFMax y noAA - C2 3200

Timedemo1: Av: 37,55, Min: 19,79, Max: 123,84, Mid: 39,52
Timedemo2: Av: 60,14, Min: 37,60, Max: 291,21, Mid: 65,48

.8,36

Timedemo1: Average: 33,12, Min: 17,51, Max: 88,29, Mid: 35,51
Timedemo2: Average: 50,50, Min: 31,28, Max: 282,81, Mid: 56,67

With the 8.37 , disappear in the control panel the setting "high quality anisotropic " ....

 

Currently it seems to be that if any GS shader is scheduled, the hardware will decrease the number of threads running for the other types indiscriminately. That's the driver doing that, though (as far as I can tell), rather than something at the hardware level.