David E. Orton's talk

Gee, that rather sounded like R420 is 32-bit instead of 24-bit, or maybe *in addition* to 24-bit, tho I don't know that it would make sense to support both.
 
This is what I was pointing at re: 32-bit:

< Q > To take in technology operation to GPU side, operational precision becomes necessary. Lowest being, 32bit floating point precision is necessary. But the internal operational precision of Pixel Shader of GPU of ATI is insufficient with 24bit.

< Orton >The present product is 24bit precision certainly. But, as for us here (24bit) there is no intention of remaining.

< Q > The notion that where with the following step it designates also Pixel Shader as 32bit precision?

< Orton >Extremely, immediately (pretty soon) is, (laughing).
Though, that it reaches operational precision above 32bit, story is another. With technology operation, also 64bit precision is required. But, when with graphics of picture indication was thought, as for the reason where precision above 32bit becomes necessary you see excessively and do not hit. In case of graphics, operational precision means the dynamic range. But, when of quality was thought, there is no aspect where the dynamic range above 32bit is necessary. It is sufficient with 32bit. "Never there is no" I, with we would like to call, we are not, but probably there are no times when in the near future the graphic processor becomes 64bit.

64bit in regard to mounting is difficult. I myself came from super computer side, (the origin SGI), has remembered that 64bit is big problem. Not only enlargement of the address space, operational unit (mounting) it was difficult even from the surface.
 
what's the estimated transistor count of R420? I remembered Uttar said it's a 12X1 architecture,and if everything is done in FP32, they may need at least a 200M die.
 
991060 said:
what's the estimated transistor count of R420? I remembered Uttar said it's a 12X1 architecture,and if everything is done in FP32, they may need at least a 200M die.
Click to expand...

The R420 is no more fp32 than the r3xx series. It is just that internally, many things are done at higher precisions. For example, on x86 cpus, doubles can be done at 80 bit precision, but the final value is truncated to a 64bit double precision fp value.
 
lost said:
The R420 is no more fp32 than the r3xx series. It is just that internally, many things are done at higher precisions. For example, on x86 cpus, doubles can be done at 80 bit precision, but the final value is truncated to a 64bit double precision fp value.
Click to expand...

That doesn't seem to fit the translation provided, tho I would certainly like to see someone try a more graceful translation before leaning too heavily on it.
 
991060 said:
what's the estimated transistor count of R420? I remembered Uttar said it's a 12X1 architecture,and if everything is done in FP32, they may need at least a 200M die.
Click to expand...

110M-150M are the numbers I see most often. I have seen people suggest 200M, but that may be sourced from before the rumored-to-be radical R400 became R500, and R420 based more closely on existing tech arrived on the scene (if I'm remembering the scenarios correctly).
 
110 million transistors for R420 is absurd since the R300 / R350 / R360 already has 107~110M.


remember the quote some months ago about R420 being in the hundreds of millions of transistors. I say that 200M or close to that, is what R420 will have. ATI roughly doubles their transistor count each generation since R100.
 
what's the estimated transistor count of R420? I remembered Uttar said it's a 12X1 architecture,and if everything is done in FP32, they may need at least a 200M die.
Click to expand...

It wont need that many, the complexity of core in the means of "pipelines"
don´t double in number of transistors.
As both R420 and NV40 already taped out and atleast one of them for quite some time, there is some doubt about one of them regarding transistor count but nothing more.
 
overclocked said:
It wont need that many, the complexity of core in the means of "pipelines"
don´t double in number of transistors.
As both R420 and NV40 already taped out and atleast one of them for quite some time, there is some doubt about one of them regarding transistor count but nothing more.
Click to expand...

From B3D current review, an R360 is 107M. An RV360 is 75M. Assuming that the difference between the two is almost entirely the 4 extra pipes, that gives 32M for the 4 extra pipes, and thus R420 would weigh in at about 139M (107+32)? This assumes (big assumption) they didn't add anything else of significance, like whatever it would take to go from FP24 to FP32. The other assumption is that the "R420=3 x RV350" rumor is more or less correct.

Gee, that's damn near substantive for me; wonder where I screwed it up. ;)
 
we forgot the increased number of on-die registers.

vs_3_0 requires 32 temp registers, R300 has 16 now. And the 512 instruction slots is 2 times of what R300 has now.

In ps_3_0, we need 32 temp registers which are 20 more than what R300 has now. Constant registers also increased from 32 to 224, and insrtuction slots will be 512 instead of 96.
 
From B3D current review, an R360 is 107M. An RV360 is 75M. Assuming that the difference between the two is almost entirely the 4 extra pipes, that gives 32M for the 4 extra pipes, and thus R420 would weigh in at about 139M (107+32)? This assumes (big assumption) they didn't add anything else of significance, like whatever it would take to go from FP24 to FP32. The other assumption is that the "R420=3 x RV350" rumor is more or less correct.

Gee, that's damn near substantive for me; wonder where I screwed it up.
Click to expand...

It´s hard too guess but also consider that RV350/360 has two vertex shaders less then R3xx class of chip.
I would think R420 being around ~150, 160 million transistors.
 
geo said:
From B3D current review, an R360 is 107M. An RV360 is 75M. Assuming that the difference between the two is almost entirely the 4 extra pipes, that gives 32M for the 4 extra pipes, and thus R420 would weigh in at about 139M (107+32)? This assumes (big assumption) they didn't add anything else of significance, like whatever it would take to go from FP24 to FP32. The other assumption is that the "R420=3 x RV350" rumor is more or less correct.

Gee, that's damn near substantive for me; wonder where I screwed it up. ;)
Click to expand...

You have to count the Hierarchical-Z engine that is, officialy, missing from 9600. And, I think, they must have some transitors on r3x0 dedicated to distributing quads from one set of 4 pipes to the other, which might grow in complexity on r420.

Also I think the vertex engine is brand new (vertex texturing?)
 
I would be careful about that site. They had all sorts of quotes from ATI guys saying the R300 or R350 was going to be .13 micron. Their information turned out to be quite inacurate.
 
I don't know, it just don't feel right to me and if it is true this would be the first and only place I've heard it from.

It doesn't make sense though, it just doesn't. I've heard from some ATi peeps that they didn't plan to go to 32FP 'til it was required which won't be 'til dx10 which won't be for a while yet...or is dx10 coming soon?
 
You must log in or register to reply here.

Similar threads

C
Kutaragi Ken: "WAN Cell Computing"
2 3 4
Replies
75
Views
7K
randycat99
R
C
Kutaragi Ken: "Cell Computing In RealTime Ops"
Replies
4
Views
3K
Megadrive1988
M
C
Kutaragi Ken: "Relation Between Cell & PS3"
Replies
11
Views
3K
Natoi
N
C
Kutaragi Ken: "PSX Concept aims at post VHS"
Replies
3
Views
2K
Paul
P
C
Kutaragi Ken: "The Semiconductor Strategy Of Sony"
2
Replies
23
Views
4K
Paul
P
Back
Top