Beyond3D Forum - The NEXT LAST R600 Rumours & Speculation Thread

Beyond3D Forum (http://forum.beyond3d.com/index.php)

- Pre-release GPU Speculation (http://forum.beyond3d.com/forumdisplay.php?f=51)

- - The NEXT LAST R600 Rumours & Speculation Thread (http://forum.beyond3d.com/showthread.php?t=39173)

I just hate this chip -- too much trade offs, viewed as a complete design.

Quote:

Originally Posted by Jawed (Post 979437)

Note the rates are for 1 iteration.

Also note I left R600 at 800MHz. Scale it down for whatever you think the real clock of XTX (or the tested XT) will be. And scale G80 up for whatever you think Ultra will be...

Jawed

That's exactly what I was saying. Multiply those rates by 6 instructions per iteration and you get 156.3 GInstr/s for R600XT and 105.6 GInst/s for G80. Tertsi's numbers were ~100 GInstr/s for R600XT and ~40 GInstr/s for G80.

Either something's not right in the measurement or the compilers aren't doing their job. Can't see any other reason for the disparity.

Quote:

Originally Posted by mao5 (Post 979611)

so u totally ignored the 3xfps from 3.52GHz E6400+660/2200MHz GTX 1280x960 in Test Driver Unlimited?

That picture from another area, the game has very unbalanced frame rates, what driver the guy using with the XT?:smile:

Quote:

Originally Posted by silent_guy (Post 979617)

It's funny: back in college, when they taught me about scan chains as a DFT methodology, I thought the wiring overhead for this must be too much. (That was with 2 metal layers and abusing poly to route some of the wires.) :wink:

Hehe, i was allowed 4 metal layers. :grin:

Quote:

Originally Posted by silent_guy (Post 979617)

One way to look at it, is that those large muxes are at their core nothing more than a whole lot of 1-bit 6-to-1 muxes. That's exactly how the tools look at it anyway. And 10(?) layers of metal clearly gives you quite a bit of freedom to indulge.

This is true. Thank god for place and route tools! On the other hand, you can't use ALL those metal layers (usually top 2 reserved for power/ground IIRC?), then you have your clock distribution and what not. Also the higher metal layers are much less dense than the lower ones... I really would like to see a layout of a massive xbar though!

Quote:

Originally Posted by silent_guy (Post 979617)

Don't get me wrong: I really accept the argument that a ring has advantages wrt back-end implementation and getting data around the chip. So if you have no other choice, it's a solution that deserves to be looked at.

But the inevitable price to pay is that you give up something at the system level. And that's a bad thing. In this case, that price means more over-dimensioning to reduce freak corner cases, over-dimensioning to limit the latency penalty (2 opposite direction rings), less control over scheduling etc. On R580, that was manageable because it was only used for read return data. But if R600 also uses this to transport write data, it gets really ugly.

A crossbar doesn't have those system level disadvantages.

And since we have proof that a crossbar implementation is practically possible, it's make little sense to cheer about the presence of a ring...

Hm... i'm not sure i'm seeing what disadvantages there are besides the overdimensioning. Wouldn't you just dump your data onto the stop closest to you and it will just go to its required stop? All the scheduling and whatnot should happen individually on each of the stops, no? Since the memory controllers are going to be on the stops i think? What am i missing here?

One day left..we need a leaked whole review to liven things up...

Quote:

Originally Posted by Rangers (Post 979643)

One day left..we need a leaked whole review to liven things up...

It's not may 14 st?

I don't think so. All the May 2nd rumors seem to be sourced from that one VR-Zone piece. At least the ones I've read anyway. I'd be happy if it were right though!

Fudo just said it'll be may 14th again for the XT launch, nothing before that.

I think nobody is expecting an AMD launch tomorrow anyway.

I see the FUD piece now. Bummer. Hope it's not true.

Quote:

Originally Posted by _xxx_ (Post 978428)

Hey, stop stealing my ideas!!! I have a bet running about the price tag of $399 already!

Already getting cold feet about your bet? ;)

Taken from the reviewers guide:

Quote:

Highlight: DirectX® 10 Ready, Avivo™ HD, UVD (full hardware decoding of Blu-ray and HD DVD), Built-in HDMI and 5.1 surround audio, Free “Black Box” game bundle, killing price/performance ratio over 8800GTS/GTX.

:cool:

Heh, I'm sure they wish that was "killing performance over 8800GTS/GTX". It'll be interesting to see Nvidia's response - a 2900XT at $400 will be mighty attractive.

Quote:

Originally Posted by trinibwoy (Post 979678)

It'll be interesting to see Nvidia's response - a 2900XT at $400 will be mighty attractive.

If they can keep up with the demand. :wink:

True, street price should have some upward pressure for the first few weeks after launch.

I hope it is a price/performance killer so forcing nvidia to react by reducing GTX and GTS prices. We need the 8800GTS 320MB to be forced into the mid mainstream to replace the rather dull 8600!

Some SKUs just can't drop below certain price point, I think.
Unless your company is called Sony. :lol:

Quote:

Originally Posted by Sobek (Post 979581)

I'm a little confused. Is it 140% as in 1.4x the speed (one x1950 + 140% additional 'performance' on top of that for the x2900), or just an additional '40%' (100% = x1950xtx, x2900 is 40% faster).

I hate percentages. Kill me.

140% = 2.4x faster not 1.4x.:razz:

First, it's "25 - 140% faster". So, the 140% case is probably some corner case.
Second, these percentages are PR-percentages, therefore I'd say these have the margin of error about 10% ;).

"25%" - in games, "140%" in HD video decode CPU utilization perhaps.

Quote:

Originally Posted by fellix (Post 979630)

I just hate this chip -- too much trade offs, viewed as a complete design.

Does no one remember "128 scalar? But that's less than G7's 24 * 4 * 2!!" ?

Obviously, I wished that the thing would come out and stomp the G80, but I'm just as interested in reading about the tradeoffs, and why they were made. There are likely to be interesting wins here.

-Dave

Not sure if I think I'm seeing it cause I kinda expect it but does the ground texture in those two shots look different?

Quote:

Originally Posted by tEd (Post 979716)

HL2 16x CFAA

http://img389.imageshack.us/my.php?i...c5bff91et2.jpg

http://img168.imageshack.us/my.php?i...3e0d643fd2.jpg

Something tells me that this card was not ment for high res.:wink:

Quote:

Originally Posted by trinibwoy (Post 979721)

Not sure if I think I'm seeing it cause I kinda expect it but does the ground texture in those two shots look different?

I was thinking the exact same thing.
It actually looks sharper in the No AA example.

blur effect of CFAA?

Quote:

Originally Posted by Razor1 (Post 979727)

blur effect of CFAA?

Yep, that's probably why "24x" has a edge detect resolve filter

Quote:

Originally Posted by INKster (Post 979725)

I was thinking the exact same thing.
It actually looks sharper in the No AA example.

So do the glove textures..

anyone recognize the level? or is this the part where you first get the rocket launcher in hl2