ATi's roadmap for 2004

[cristic]

For the entire story go here:

http://www.anandtech.com/video/showdoc.html?i=1966

We have been hearing about R420 for quite some time now and recently the name R423 has been in the headlines. As you can guess, R420 is the AGP successor to the R350 (Radeon 9800) while R423 is a PCIe version of the R420. The specs on the two GPUs are as follows:

* 0.13-micron low-k manufacturing process
* 160M transistors
* ~500MHz core clock
* 8 pipe design
* 6 vertex engines
* Improvements to all of the basic architectural features (shader engines, AA, etc...)
* 256MB 256-bit GDDR3 (~1GHz data rate)
* Single slot design

The R420/423 chips will offer twice the pixel fill rate and vertex throughput of the R350 core, as well as increases in memory bandwidth. Initial indications show that there may be two versions of the R420/423 with different memory clocks; one design calls for ~1GHz GDDR3 memory while the other calls for slower DDR1 memory.

8 pipe design ... hmmm...

 

[Tim]

8 pipe design ... hmmm...

Dave hinted 8 extreme pipes at some point.

 

[Dave Baumann]

Salt, pinch of.

 

[MistaPi]

Twice the pixel rate of the R350 with the same amount of pipelines? Is that possible?

 

[Martin Eddy]

8 pipe design ... hmmm...

Yeah but they also call NV30/35 8 pipe design. Anands at it again.

 

[Ailuros]

Twice the pixel rate of the R350 with the same amount of pipelines? Is that possible?

Depends with what kind of base filtering you calculate fill-rates or not?

 

[Waltar]

Twice the pixel rate of the R350 with the same amount of pipelines? Is that possible?

Nope. And there is no way ATi is staying with 8 pipelines, unless they're basically moving the r360 to .13 micron and doubling the vs units.

And I highly, highly doubt that.

 

[vb]

how will 6 vertex engines process a quad?

(or more quads)

Not getting it ........

 

[jpaana]

how will 6 vertex engines process a quad?

(or more quads)

Not getting it ........

Vertex processing isn't in any way connected to how the primitives are rasterized. Depending on the implementation they could be processing six or more vertices at the same time, which is nice as the shader lengths grow.

 

[Luminescent]

how will 6 vertex engines process a quad?

R300/50/60 had 4 vertex engines, therefore doubling the number would mean a total of 8 vertex engines for R420.

 

[Geo]

Salt, pinch of.

Dave, how long were you under NDA for R8500, R300, R350, R360?

 

[jvd]

sigh is it going to be worth it to move over to a pci express platform ?


I might have to pick up a 3000+ 64bit athlon for cheap and whatever is the cheapest nforce 3 mobo with pci .

 

[Mark]

how does Anand figure "twice the pixel fill rate" of the R350 when it has the same 8-pipe design and the core speed only goes up by ~120MHz.

core speed x pixel pipelines = pixel fill rate, right?

 

[jvd]

how does Anand figure "twice the pixel fill rate" of the R350 when it has the same 8-pipe design and the core speed only goes up by ~120MHz.

core speed x pixel pipelines = pixel fill rate, right?

perhaps its a defered renderer ?

Normaly you'd take the fillrate and times it by 2 for a base line of how much extra performance you'd get and most of the time it scale up from there .

 

[vb]

how will 6 vertex engines process a quad?

(or more quads)

Not getting it ........

Vertex processing isn't in any way connected to how the primitives are rasterized. Depending on the implementation they could be processing six or more vertices at the same time, which is nice as the shader lengths grow.

There was earlier speculation on this forum that R420 features R500 vertex engines. as R500 is suposed to have similar units for VS and PS, our very own, now somewhat silent DaveBaumann suggested that they may be capable of limited pixel processing. all that and suposed doom 3 enhancements leds to believe they may be used for a Z-pass (zixel, 8x0 or whatever) considering all that I asumed it will have 4 very advanced vertex units. on 40 mil more transistors I guess you it will be hard to have 2 more vertex units and all of them so advanced.

so, if this is true, how will six units work on a quad?

Luminiscent: 8 doesn't sound any better due to complexity involved

 

[vb]

anyway (8+6)X500= 7000Mpx
(8+4)X500= 6000Mpx ->double R350
R350 8X380=3040Mpx

 

[epicstruggle]

The only interesting part from anand's was this:

It isn't clear whether R423 (PCIe) based designs will eventually carry higher clock speeds than their AGP counterparts, but there is a definite possibility.

This could be a smart way of pushing PCIe.

later,
epic

 

[UPO]

how will 6 vertex engines process a quad?

(or more quads)

Not getting it ........

Vertex processing isn't in any way connected to how the primitives are rasterized. Depending on the implementation they could be processing six or more vertices at the same time, which is nice as the shader lengths grow.

There was earlier speculation on this forum that R420 features R500 vertex engines. as R500 is suposed to have similar units for VS and PS, our very own, now somewhat silent DaveBaumann suggested that they may be capable of limited pixel processing. all that and suposed doom 3 enhancements leds to believe they may be used for a Z-pass (zixel, 8x0 or whatever) considering all that I asumed it will have 4 very advanced vertex units. on 40 mil more transistors I guess you it will be hard to have 2 more vertex units and all of them so advanced.

so, if this is true, how will six units work on a quad?

Luminiscent: 8 doesn't sound any better due to complexity involved

even If vetex units are capable of some kind of pixel processing it doesn't necessarily mean working on quads (IMHO). maybe all the "pixel processing" is about texture look up?

 

[vb]

even If vetex units are capable of some kind of pixel processing it doesn't necessarily mean working on quads (IMHO). maybe all the "pixel processing" is about texture look up?

I don't see how this would help D3

 

[Guden Oden]

so, if this is true, how will six units work on a quad?

Easy. Vertex shaders don't work on quads, they work on vertices (the points defining the shape of a polygon).