ATI R420 Specs & Hot News (R4xx Series)
July 30, 2003, 1:14 AM CST by Fusion
http://www.nvnews.net/ - Nv news reports that the R420 (ATI's next chip after the upcoming R360) will use a 12x1 or 6x2 pipeline configuartion. Also to add to this Nvidia's NV40 (R420's main competitor) will supposedly use a 8 pipeline config.. i doubt it will be 8x2, so this puts ATI in a large advantage. but then again, here are the current speculated specs..

12x1/6x2 architecture (using general purpose FP/texturing units a la NV3x/R500 and an R3x0-type pixel pipeline + backend) 0.13u (k>3.0) ASIC @~450Mhz 256MB GDDR-2@~750MHz DX9 (+PS/VS 3.0) compliant.

While theres some nice new stuff in there (13nm, +PS/VS 3.0) the GPU and memory really are quite low for a top end 2004 card. I feel that if the NV40 has a GPU over 550 mhz and memory over 1ghz (256-bit memory bus ofcourse)they may still be able to compete with this even with only using 8x1 architecture (which NV40 is supposedly going to use). However i have seen speculation about nvidia using a 12x1 config for NV40.. so its not entirely out of the question..

However i feel the next gen may be a repeat of the current gen. Nvidia will have the higher clocked GPU/Memory, but ATI will have better technology all round, and still be able to keep up, or even better Nvidia. Looks like it will be interesting, i dont think ATI will go back to second place any time soon. Nvidia will have to fight hard to regain the position they held around a year ago.
July 30, 2003, 1:21 AM CST by Pichu
However, that "speculated specs" is based on a logical guess, which most likely will be inaccurate. Here's the original quote:
R420, my best guess:

12x1/6x2 architecture (using general purpose FP/texturing units a la NV3x/R500 and an R3x0-type pixel pipeline + backend)
0.13u (k>3.0) ASIC @~450Mhz
256MB GDDR-2@~750MHz
DX9 (+PS/VS 3.0) compliant

http://www.nvnews.net/vbulletin/showthread.php?s=&threadid=15200&perpage=25&pagenumber=4
Surely, all new technologies are based on older technologies. So, if the delta remains a constant, it's no doubt of the repetition, but future may tell!
July 31, 2003, 5:53 PM CST by pat777 to Fusion
i doubt it will be 8x2
Why not? The NV30 was already 4*2. R300 is already 8*1. I'll be surprised if it wasn't 8*2.
July 31, 2003, 11:44 PM CST by Gitaroo
yeah, I would be really suprise if they dont have 2 texture per pieline for the next Geforce.
August 1, 2003, 8:33 PM CST by kenneth9265
Think and hope that both Nvidia and ATI will go the 12 pipe route.
August 2, 2003, 10:00 AM CST by Fusion to Gitaroo
really surprised? It doesnt matter at all, in current games all that matters in the result of the pipelinex*texture untits per pipe.. in multitexturing games 8x2 would be the same as 16x1, you dont need 2 texture units at all...

I doubt it will have that many because its alot to feed through the pipelines, the same as a 16x1 config, that is very large, and is not suitable for current cards.

The inquirer had speculation about this, and they also concluded that 8x2 or 16x1 will probably not happen for this genration.. but they speculated that the NV40 will most likely have a 12x1 setup.. that would be the smart thing to do, but nvidia havent made all the smart decisions lately.

Also around 3-4 months ago nvidia stated that they will not use a setup which only benifits fully from multitexturing apps. So they would use only one texture unit per pipe. Is that enough reason for my assumption?
August 5, 2003, 10:27 PM CST by chairmansteve
Let's count some transistors.

I'm going to assume DX9 to DX9.1 compliance for instruction decoding/processing won't require a big difference in transistors count, so that allows us to analyze the pixel/vertex pipes chip real estate independently from the rest of the chip. I'm also assuming most of RV350 and R350 is similar, so we can make this relative comparison.

R350 (Radeon 9800)
110 Million Transistors
8 Pixel Pipes
4 Vertex Pipes

RV350 (Radeon 9600)
80 Million Transistors
4 Pixel Pipes
2 Vertex Pipes

110 - 80 = 30

So, 30 million extra transistors allows 4 more pixel pipes and 2 more vertex pipes.

30 / 2 = 15

15 million trans. are needed for 2 pixel pipes and 1 vertex pipe. I'll assume that the amount of cache is relative to number of pipelines, so that is included in this transistor count.

110 - (15 * 4) = 110 - 60 = 50

50 million transistors covers everthing other than the pixel/vertex pipes. That covers the memory controller (well, R300 256-bit controller may take up a lil more space than 128-bit from RV350), 2D engine, video/DVD processing, display interface, anti-aliasing, Hyper Z, compression, etc.

Thus, we have the following formula.

B + (K * V) = T

B = base number of transistors = 50 million
V = number of 2+1 pixel/vertex unit groups
T = total transistors
K = transistors per 2+1 pixel/vertex pipes = 15 million

For 16 pixel pipes + 8 vertex pipes:
50 + (15 * 8) = 50 + 120 = 170

Would at least 170 million transistors be a likely possibility for 0.13-micron R420? Sure, it could be done.

But! Maybe R420 will increase color precision to 32-bit per component.

15 * (32/24) = 20

That brings K up to 20 million (at most, since the vertex units already support 32-bit per component).

50 + (20 * 8) = 50 + 170 = 210

210 million transistors might be pushing it, but still possible.

If we get 16 pixel + 8 vertex pipes, what sort of performance numbers shall we expect at 400MHz?

6.4 Billion Pixels/Sec
800 Million Vertices/Sec

That's 2.105x the power of Radeon 9800 Pro. They better match that with 1.0-1.2GHz GDDR2.

Anyhow, 210 million trans is cutting it close, so maybe 12 pixel pipes would be more reasonable. However! 8x2 could require fewer transistors than a full 16x1, so 8x2 may not be out of question.
August 5, 2003, 11:54 PM CST by Inane_Dork
ZZOOOOOOOOOOOOOOOOOMMM!!!

:p


Seriously, any way you figure it, that's a heck of a lot of processing power.
August 7, 2003, 12:00 AM CST by TheBoondockSaint to Fusion
I was under the impression that the nv30 was really a 4x2 design but then could perform like a 8x1 in certain situations. Considering that the nv30 is at 4 actual pipes now id doubt that they're going to be able to jump to a 12x1 or a 6x2 for the nv40.

"While theres some nice new stuff in there (13nm, +PS/VS 3.0) the GPU and memory really are quite low for a top end 2004 card."

if it turns out to be only $400 (yea i know, -only- $400) i wouldnt say it would be much of a let down cause it would still be higher clocked and the same amount of ram as the $500 ones today. now if turned out to be 500 then i'd be a different story
August 7, 2003, 2:23 AM CST by Inane_Dork to TheBoondockSaint
Look down in this thread for schematic diagrams and quotes indicating the issue (if you're interested). Mr. 1337 Leader chairmansteve found them.

http://forum.pcvsconsole.com/viewthread.php?tid=6181
August 8, 2003, 4:00 AM CST by Fusion to TheBoondockSaint
Indeed the NV15,NV20,NV25,NV30 and NV35 have 4 pipelines and all push 2 TMU's per pipe. But how does this make it unlikely for them to make a 12x1 or 6x2 architecture?? The only difference between Nvidias and ATi's current technology as far as pipelining is concerned is that ATI use a higher pipe config, with only 1 TMU per pipe, while NVidia use less, but 2 TMU/pipe. How does this mean ATI can make 12x1/6x2 cards but Nvidia cant.. i dont see the logic there, seriously i think you are overestimating the few weaknesses having less pipelines but more TMU's has. Also recent rumors show that NV40 should be now released July/August, and that NV38(NV35 revision) will be released not to far from now. The reliability of this is also rated as very high. So is this not long enough for Nvidia to implement a new architecture? The NV40 now looks to directly compete with R420, because of its delay from its unrealistic Q4 04 release.

http://www.notforidiots.com/GPURW.php

Expect R420 to be a top of the line card just as the 9700pro and 9800pro were. So yes most likely it will be $500. Especially because its new architecture, and that higher pricing will be more accepted/expected after this generation. But seriously... do you think specs below Nvidias current top contender will seriously be able to take on NVidia's next gen card?? Even with more pipelines i dont see it happening. 750mhz GDDR-2 isnt that much pushing power for 12pipelines+

With the delay of NV40, i would expect a 600mhz+ core, 1200mhz+ GDDR-2, 12x1/6x2, as well as some nifty new features and improved AI with perhaps a true 16x AF, and more AA modes. In fact even with only a 8x1 setup i can still see a card with these type of specs beating the R420 with its given specs.
August 12, 2003, 9:12 PM CST by Synth
R4xx to be "Hundreds of Millions of Transistors" - 4:26 am EST - ben6
Rick Bergman, ATI's VP of Desktops presented at the Pacific Crest Securities Investors Conference Forum. Some highlights

1. 9600 is shipping in volume with no yield issues.
2. R4xx will be the next generation chip next year and have hundreds of millions of transistors on .13um
3. Integrated will be important going forward in desktop segment


http://www.rage3d.com/#1060676768
August 12, 2003, 11:20 PM CST by chairmansteve to Synth
Yep, 200+ million trans is what I've been expecting for ATI R4xx VPU.

A few months ago, this press release gave weight to my expectation.
ATI's visual processing unit (VPU) has more than 200 million transistors of digital logic. A design of this size and complexity requires not only extremely high stuck-at fault coverage, but also thorough testing for delay-related defects, the preponderant defect type in 0.13 micron process geometries and below. Using basic scan methods, excellent delay test requires up to 6X more tester time than required for stuck-at faults, which already is at an unacceptable cost of test.

http://www.synopsys.com/news/announce/press2003/ati_socbist_pr.html
August 15, 2003, 12:52 AM CST by Synth to chairmansteve
Ahh interesting. You know, some people are saying this is for R5xx, not R4xx. they think that R4xx will be around 150 million transistors, but that makes little sense to me, being as ATi has been (roughly) doubling their transistor counts every GPU generation.

I think R420 and R450 will probably wiegh in at around 200-220 million. then in 2005, on .09nm, I'd guess the R500 will be at around 400-440m. the next Xbox GPU (probably R5xx based) will probably exceed 500m.

That is about where PS3's EE3 CPU (not GPU) was originally going to be, circa 1999-2000. actual plans with Cell have no doubt changed somewhat, but PS3's Cell will likely be 500-800m transistors and the GPU could top 1 billion. even if PS3's GPU is not that large, the combined CPU and GPU should still top 1 billion.

All just simple estimations, based on where chip complexity seems to be headed ^__^
December 30, 2003, 8:26 AM CST by Mikefoundahobby to Fusion
I just hope that nVidia gives up their FX design and come with something they don't need to lie about.

I had several nVidia cards but this year I made the move to ATI,and so far I don't regret it (Radeon 9700 for 209 Euro).

nVidia is hurting the whole gaming industry,the Cg stuff takes up to 5x as long to get the sxit together,that means delays for releases and maybe even higher pricing for games!

They questioned FUTUREMARK,argued with VALVE and even John Carmack (pointed out that DX9 performance on FX design will be behind compared to R3xx no matter what driver they will develop).

The really sad thing about it all that (ATI cheated too in the past with their Radeon 8500 drivers benching Q3 in particular) we consumers are paying the price for all this crap.

Benchmark programs are important for developers to indicate/measure performance,it even became integrated in marketing strategy.

The pressure for having good benchmark results is so big that even a respected company (well till Q1 2003 at least) like nVidia gave away their credibility for being reliable for marketing sakes.

Please don't get me wrong I don't hate nVidia nor am I anti nVidia,I just want nVidia and ATI to make good products that compete to oneanother,so we consumers get a fair price for fair performance.

The thing I see now is that benchmark result for nVidia FX products are simply unreliable because you don't know when you read about a new driver release wether they cheated or improved their drivers (drivers were nVidia's strongest marketing/selling arguement.....up in smoke by the end of 2003).

You buy a gfFX for DX9 and they reset to 8.1 to get some more FPS cause in DX9 it won't really work,and face it who doesn't feel cheated when you find out about that ($500 ouch!)?

Another thing is that if ANY company wants to make consumers believe that a single driver update is gonna boost your 3d performance by anything more then 10% (and that already might be high) thinks that all their custumors are idiots.

Kudos ATI
January 25, 2004, 10:08 AM CST by SH64
A blob on anandtech about the R420 suggests the its going to support FP32.
I've read the latest rumors on FP32 support in R420, but I don't see how that's any major news. As far as I know, the DirectX 9.1 spec (with pixel shader 3.0 support) calls for full precision to be 32-bit floating point, so if ATI wants to be DX9.1 compliant they will have fp32 support.
Source : http://www.anandtech.com/weblog/index.html
under the title "wireless internet is great"
January 25, 2004, 9:17 PM CST by Fusion to SH64
Well i hope so, i think it should have enough power to use such precision, and if it's the DX9.1 spec, then i see no reason why ATI wouldn't use it.
© 2000-2005 pcvsconsole.com