Beyond3D Forum

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)

Galduta 02-May-2007 12:08

The two pictures have grass shadows, softshadows , I think . Iis a DX 9 render feature

neliz 02-May-2007 12:21

Quote:

Originally Posted by aeryon (Post 980301)
or maybe you think it will be in front of G71. so what's the point ? :roll:

There are no newer drivers for a couple of cards in that bench. especially at the time of that bench.
Besides, the x2900's tested now aren't tested with release drivers as well...

IbaneZ 02-May-2007 12:44

Quote:

Originally Posted by Kaotik (Post 980281)
http://www.hisdigital.com/html/produ...d=304&view=yes
Was that one posted already? HIS has 2900XT on their site now

That's sweet. :smile:

My two latest cards have been HIS, and I'll probably buy one this time too.

Quote:

VIVO Cable
HDTV Output cable
DVI to VGA Dongle x 2
DVI to HDMI Dongle x 1
Crossfireâ„¢ Cable x 1
Nothing about a 8 pin "converter" though. I hope they'll include one.

fallguy 02-May-2007 13:00

Quote:

Originally Posted by Unknown Soldier (Post 980307)

I know what the black box is, but there was no confirmation till just late last night, that it came with the 2900. Its now on HIS' site. After I posted..

Interested in the 5.1 audio though. I dont understand how they have will audio, with a DVI to HDMI adapter.

neliz 02-May-2007 13:02

Quote:

Originally Posted by fallguy (Post 980327)
I know what the black box is, but there was no confirmation till just late last night, that it came with the 2900. Its now on HIS' site. After I posted..

Interested in the 5.1 audio though. I dont understand how they have will audio, with a DVI to HDMI adapter.

audio on the hdmi connector. there is more than enough bandwidth for video + audio on the dvi connector used.

Dalton Sleeper 02-May-2007 14:05

Fudzilla wrote that 2900XT in CrossFire scores 13100 in 3DMark06, I would expect more, alot more.

nicolasb 02-May-2007 14:30

Quote:

Originally Posted by neliz (Post 980329)
audio on the hdmi connector. there is more than enough bandwidth for video + audio on the dvi connector used.

Yes, but there won't be an HDMI connector. There will be a DVI connector with a dongle. And DVI doesn't carry audio. How does the audio signal manage to get as far as the dongle?

fallguy 02-May-2007 14:34

Yes, that is my question. Id rather just have a HDMI connection already. I bought a DVI>HDMI adapter years ago, hurry up an adopt it.

vertex_shader 02-May-2007 14:37

Quote:

Originally Posted by Dalton Sleeper (Post 980341)
Fudzilla wrote that 2900XT in CrossFire scores 13100 in 3DMark06, I would expect more, alot more.

1600x1200 8xmsaa? :wink:

Arnold Beckenbauer 02-May-2007 14:38

Quote:

Originally Posted by nicolasb (Post 980347)
Yes, but there won't be an HDMI connector. There will be a DVI connector with a dongle. And DVI doesn't carry audio. How does the audio signal manage to get as far as the dongle?

Quote:

Originally Posted by Xmas (Post 967048)
It works because HDMI is backwards compatible to DVI, and the graphics card can query device capabillities. So the link will start in "DVI compatibility mode", and when a HDMI capable device is detected the graphics card will add audio output (which is interleaved in the digital stream, so no additional pins are required).

Using a DVI-I connector on the card is the best choice because it can drive DVI single and dual link, HDMI single and dual link, and analog VGA.

:smile:

Chalnoth 02-May-2007 14:44

Quote:

Originally Posted by Nebula (Post 980313)
Interesting, the 8800GTX ss has better image IQ, the fire has black borders (like it is not translucent) on 2900XT. Also the 8800GTX ss seems sharper even though the imagefile is much more compressed. :???:

Well, I haven't played Stalker at all, but from playing Oblivion, that difference really looks like the difference between HDR and no HDR.

neliz 02-May-2007 14:47

Edit: whatever arnold said! :)

Jawed 02-May-2007 14:50

Quote:

Originally Posted by silent_guy (Post 980187)
Excellent slide, Topman!

Ouch, it looks like they abandoned 32-bit MCs! That would go a long way explaining the horrible GDDR4 XTX benchmark numbers of DT. I suppose they had no choice with a 512-bit bus?

Return to 64-bit channels is definitely very curious... I presume this means that each pair of chips in a channel is spread over both sides of the board, i.e. back-to-back, instead of sat side-by-side. I say this because the 3-3-2 arrangement looks awkward for 2 (out of the three) side-by-side pairings.

As to explaining the GDDR4 XTX numbers, weren't there rumours of a broken memory controller? If there are no actual XTX boards in the wild now, and all that's out there are old, partly broken prototypes, then...

Jawed

Jawed 02-May-2007 15:01

http://resources.vr-zone.com/newzhunter/HD2900-10.jpg

I can now make out what's going on with the texture unit, working down through it, left-to-right:
  • 4x TA unit for vertex or texture fetch (prolly unable to do LOD/Bias, filter-specific calculations?)
  • 4x fp32 vertex fetch or fp32 texel fetch
  • 4x TF units
  • 16x fp32 texel fetch (might be usable for vertex fetches?)
  • 4x TA units for filtered texture addressing (cannot be used for vertex fetch addressing)
Hope that's right...

Jawed

Jawed 02-May-2007 15:05

Quote:

Originally Posted by silent_guy (Post 980184)
There should only be a few clock cycles delay between entering and leaving the crossbar itself. So in that case, assuming no stalls, the difference is entirely in the propagation delay. Worst case, that would be double for a ring? In practice, I expect it to be less. But even if it's only, say, 30% more instead of 100%, that's a lot of additional buffering.

OK, for the worst case, it would seem we're talking about 40 clocks + 200 for DDR = 240 versus 280 in total for the ring bus.

Is that reasonable?

Jawed

chavvdarrr 02-May-2007 15:09

Quote:

Originally Posted by Jawed (Post 980371)
I can now make out what's going on with the texture unit, working down through it, left-to-right:
  • 4x TA unit for vertex or texture fetch (prolly unable to do LOD/Bias, filter-specific calculations?)
  • 4x fp32 vertex fetch or fp32 texel fetch
  • 4x TF units
  • 16x fp32 texel fetch (might be usable for vertex fetches?)
  • 4x TA units for filtered texture addressing (cannot be used for vertex fetch addressing)
Hope that's right...

Jawed

And would you be so kind to compare vs G80 texturing power?

fellix 02-May-2007 15:10

Does that mean, that any order of vertex fetched texture filtering will be done through a Geometry Shader run code?
Much like the lack of FP16 box filtering in R5x0, were custom one have to be coded, then.

Sound_Card 02-May-2007 15:12

Quote:

Originally Posted by chavvdarrr (Post 980377)
And would you be so kind to compare vs G80 texturing power?

At the the same time, R580 too.

silent_guy 02-May-2007 15:51

Quote:

Originally Posted by Jawed (Post 980375)
OK, for the worst case, it would seem we're talking about 40 clocks + 200 for DDR = 240 versus 280 in total for the ring bus.

Is that reasonable?

You mean 40/80 cycles for propagation across the die and for all the MC related scheduling? According to Bob, the 40 would be too low, right? Where does the 200 come from?

Anyway, if your numbers are reasonable, it's a 16% buffering penalty just for transport.

Jawed 02-May-2007 15:57

Quote:

Originally Posted by fellix (Post 980378)
Does that mean, that any order of vertex fetched texture filtering will be done through a Geometry Shader run code?
Much like the lack of FP16 box filtering in R5x0, were custom one have to be coded, then.

Being a unified architecture, all kinds of fetch/filter are available to all kinds of shader.

So geometry or vertex or pixel shaders can perform vertex fetches or texture fetches or bilinearly (or better) filtered fetches. On Int8, fp16 or fp32 data formats.

I'm heading out now...

Jawed

CJ 02-May-2007 15:58

Quote:

Originally Posted by IbaneZ (Post 980324)
That's sweet. :smile:

My two latest cards have been HIS, and I'll probably buy one this time too.

I've been told that HIS is still playing with final clockspeeds. It looks like they're going to clock their HD2900XT a bit higher than reference clocks.

R600 is in full mass production at the moment and retailers should all have their cards at the 14th of May in time for the launch.

trinibwoy 02-May-2007 16:07

Quote:

Originally Posted by Dalton Sleeper (Post 980341)
Fudzilla wrote that 2900XT in CrossFire scores 13100 in 3DMark06, I would expect more, alot more.

The 8800 Ultra doesn't score much more than that on an X6800. Once you get in that range it's all about the CPU.

Robin B 02-May-2007 16:24

Quote:

Originally Posted by CJ (Post 980427)
I've been told that HIS is still playing with final clockspeeds. It looks like they're going to clock their HD2900XT a bit higher than reference clocks.

R600 is in full mass production at the moment and retailers should all have their cards at the 14th of May in time for the launch.

Could it be that Ati wants to clock it to 750 just to keep it at 225w ? Not many out there with a 8 pin psu.

Geeforcer 02-May-2007 16:51

You know, I for one am getting really tired of the whole "Here some scores with no specks or settings provided" or my new favorite "Here is one screenshot with FRAPS on". IT. IS. SO. POINTLESS.

fellix 02-May-2007 16:56

Well, the NDA is still too heavy to lift. :lol:

All I want now is a [very] high-detail shot of the R600 die. Sadly, don't have one of G80. :???:


All times are GMT +1. The time now is 16:29.

Powered by vBulletin® Version 3.8.6
Copyright ©2000 - 2013, Jelsoft Enterprises Ltd.