r520 performance speculation

[KimB]

There is nothing like SM3.0+.
And you cannot use things that are not written in the D3D specs.

Until DirectX is updated. It's not like sub-versions of DirectX haven't been released before.

Additionally, SM 3.0 does have a number of optional components, such as the maximum number of instructions, the maximum number of branches allowed, etc. In this respect, the GeForce 6x00 could already be considered SM 3.0+, since it does offer more than the absolute minimum for SM 3.0 support.

 

[ChronoDog]

I don't think that anything majorly new is going to be released before DirectX Next hits the streets... I doubt it that even such a thing as SM4.0 will be released, because even if half of what they said about DirectX Next is true, the whole architecture and principe of videocards will have to be quite different.

So I'm assuming that R5x0 is going to be nothing more than an SM3.0 card, with high hopes that DX Next is out by the time they'll have to push R600 on the shelves

 

[davepermen]

if it will be some SM3.0+, i hope it will add.. uhm.. noise().

that'll be cool (and good if ati wants to state they support gl2.0 with glsl.. as glsl has noise in).

what else could be useful?

well, bilinear for vertex textures, but thats not part of the shader feature, i think..

nothing i could imagine besides those two points for now.. anyone else?

 

[KimB]

Nah. I think the #1 thing you'd want to improve in SM3 vs. nVidia's implementation is branching performance.

Edit:
Well, I guess I should say you'd want to improve branching performance in the pixel shader, and texture read performance in the vertex shader.

 

[Calavaro]

I doubt they can up the core speed much more on this old ageing r300 core.
I bet they will have to add some more pipes (most likely 24 but 32 is a possibility) and keep the core speed the same as they have now (minimal core speed increase to keep with ATI "tradition") to increase "speed".
There will be SM3.0 support.
They will increase float precision to 32-bit. (also another reason to increase pipe count)
It will be called X900 (with the added XT, XT-PE, PRO et al)
They will use further and more agressive "opimizations" to make up for the lack of bandwidth, since highspeed RAM will be very scarce.
There will be dualcore cards to compete with nVidia's SLI.

They will have a X900XT-PE-Special Ed. that nobody can get a hold of that "wins" all benchmarks.

exit

 

[whql]

I doubt they can up the core speed much more on this old ageing r300 core.

How did you reach that one, Sherlock? Clock speeds have scaled well with process change so far.

They will increase float precision to 32-bit. (also another reason to increase pipe count)

Hoe do these have any bearing on each other?

It will be called X900 (with the added XT, XT-PE, PRO et al)

Doubtful, since this is likely to be a series of products based on the r5xx architecture.

They will have a X900XT-PE-Special Ed. that nobody can get a hold of that "wins" all benchmarks.

Funny that this critism seems to be levelled at ati, when the 6800 Ultra seems to be just as much of a phanton card by any metric (and thats before we look at Ultra Extreme / OC editions).

Here's one such metric:

http://www.steampowered.com/status/survey.html

ATI Radeon X800 XT Platinum Edition 4,042 0.33 %
NVidia GeForce 6800 Ultra 3,976 0.33 %

 

[Calavaro]

I doubt they can up the core speed much more on this old ageing r300 core.

1. How did you reach that one, Sherlock? Clock speeds have scaled well with process change so far.

They will increase float precision to 32-bit. (also another reason to increase pipe count)

2. Hoe do these have any bearing on each other?

It will be called X900 (with the added XT, XT-PE, PRO et al)

3. Doubtful, since this is likely to be a series of products based on the r5xx architecture.

They will have a X900XT-PE-Special Ed. that nobody can get a hold of that "wins" all benchmarks.

4. Funny that this critism seems to be levelled at ati, when the 6800 Ultra seems to be just as much of a phanton card by any metric (and thats before we look at Ultra Extreme / OC editions).

Here's one such metric:

http://www.steampowered.com/status/survey.html

ATI Radeon X800 XT Platinum Edition 4,042 0.33 %
NVidia GeForce 6800 Ultra 3,976 0.33 %

Let me remind you of the topic title and the answer your 4 points.

Topic title: r520 performance speculation

Answer 1 through 4: See topic title.

Stop being an ass and take it for what it is. SPECULATION. Sherlock.

 

[Mark]

600/1300
16-pipes
SM3.0

basically a faster R420 with SM3.0

90nm maybe?

 

[hovz]

600/1300
16-pipes
SM3.0

basically a faster R420 with SM3.0

90nm maybe?

if this card only has 16 pipes id question the value of even releasing it. seeing as its based on the r300 it will probably have a similar pipeline structure. that would rule out any noteworthy performance increases except those under certain shader situations.

 

[Smurfie]

1. 550/1200
2. 16 pipes
3. SM3.0
4. 32 bit precision (Will we finally have a stop to that 24bit nonsense and FUD?)

Reasons.
From observations so far, Intel's haven't been able to ramp the speed up much from their transition of 130nm to 90nm. I doubt ATI or Nvidia can muster the resources into transistor research to outdo Intel. Given that ATI will be going Sm3.0, it means more transistors needed to support SM3.0 features, including upping precision to 32 bits. I guesstimate a 25-33% increase in transistor count just on the 24->32bit increase. Since transistor count will probably shoot up, along with the die size, the switch to 90nm will not help much. Heat will be an issue at the forefront, and as a result, clockspeed will not ramp up. Also, with the need for more transistors needed for 32bit precision and SM3.0 features, it might be unreasonable to expect more than 16 pipes.

 

[nelg]

Stop being an ass and take it for what it is. SPECULATION. Sherlock.

Then I speculate that it will have hairy balls.

 

[nobody]

There is nothing like SM3.0+.
And you cannot use things that are not written in the D3D specs.

There wasn't anything like SM2.0+ until nVidia launched the FX series and dubbed their version SM2.0+... and MS added stuff like PS_2_0_a to the specs to prevent major disaster for nV.

No, you are wrong here. SM2.0+ is well defined in the D3D9 specs.

 

[nobody]

There is nothing like SM3.0+.
And you cannot use things that are not written in the D3D specs.

Until DirectX is updated. It's not like sub-versions of DirectX haven't been released before.

Additionally, SM 3.0 does have a number of optional components, such as the maximum number of instructions, the maximum number of branches allowed, etc. In this respect, the GeForce 6x00 could already be considered SM 3.0+, since it does offer more than the absolute minimum for SM 3.0 support.

I don't believe there will be another DX Update before longhorn and WGF arrive. And i don't consider anything that's within the specs but exceeding the minimum specs being worth a "+".
SM2.0+ is a different thing as is exceeds the maximum specs of pure SM2.0.
Of course, marketing is again a complete different thing, thus i'm quite sure we will here something about SM3.0+...

 

[whql]

Stop being an ass and take it for what it is. SPECULATION. Sherlock.

I would have thought there would have been some reasoned thought process though!

4. 32 bit precision (Will we finally have a stop to that 24bit nonsense and FUD?)

Whats "FUD" about FP24? This is the specification for ps2.0 and its not like there has been anything that has shown any issues from it.

From observations so far, Intel's haven't been able to ramp the speed up much from their transition of 130nm to 90nm. I doubt ATI or Nvidia can muster the resources into transistor research to outdo Intel.

Graphics aren't up at 3GHz though.

Given that ATI will be going Sm3.0, it means more transistors needed to support SM3.0 features, including upping precision to 32 bits. I guesstimate a 25-33% increase in transistor count just on the 24->32bit increase. Since transistor count will probably shoot up, along with the die size, the switch to 90nm will not help much.

The suggestion is that r520's die size will be smaller.

 

[nobody]

1. 550/1200
2. 16 pipes
3. SM3.0
4. 32 bit precision (Will we finally have a stop to that 24bit nonsense and FUD?)

That sounds very familar to me. I think it's the most realistic assumption about R520 we can make so far.

SM3.0
550MHz core
600MHz ram
16 pipes

 

[nobody]

Whats "FUD" about FP24? This is the specification for ps2.0 and its not like there has been anything that has shown any issues from it.

FP24 is the minimum specification for SM2.0 full precision. And probably it would not be FP24 if ATi had support for FP32 in their first DX9 chip.

 

[whql]

FP24 is the minimum specification for SM2.0 full precision. And probably it would not be FP24 if ATi had support for FP32 in their first DX9 chip.

So what? How is that "FUD"? Its still the specification. Clearly ati gave a persuasive argument to Microsoft for that to be the case. The balance of evidence would suggest that is not a decision that was bad in any way since it clearly enabled faster first generation dx9 chips, which in itself would have promoted the use of dx9 to developers, and there has been nothing that shows any issues with fp24. How is it "FUD"? Especially if it has enabled faster uptake of dx9 usage?

 

[Tim]

1. 550/1200
2. 16 pipes
3. SM3.0
4. 32 bit precision (Will we finally have a stop to that 24bit nonsense and FUD?)

From observations so far, Intel's haven't been able to ramp the speed up much from their transition of 130nm to 90nm.

Intel more than doubled their transistor count with Prescott (both logic and cache), and has increased the clockspeed 12% compared to Northwood so far, if Ati does just as bad the R520 would be 600MHz core and 400 milion transistors.

 

[Sage]

you guys are forgetting the new memory architecture.

i think 16 pipes (24 or 32 is way too much with FP32 SM3 pipelines for just another R300-based chip, even going to 90nm) and a reasonable core clock increase. I'd say 600-650.

The thing you guys are forgetting is balance- what will going to a 24 or 32 pipe chip running at 600MHz do without associated increase in memory bandwidth? As it is, high-end chips are really starving for more bandwidth. The transition to 256-bit and GDDR3 yielded some huge benefits but do we see anything close on the horizon? 1300 MHz memory on a 256-bit bus will not cut it for the fill-rates you guys are talking about. What's needed is something more intelligent.

 

[KimB]

The GeForce 6600 GT seems to do pretty well with its 128-bit bus and 500MHz clockspeed with eight pipelines. So I don't think we've hit a memory bandwidth wall yet. Current technologies have pushed that wall back quite far.