R420 have 160M transistors

[AthlonXP1800]

TSMC's sources (http://www.theregister.co.uk/content/3/34607.html) claimed R420 have 160M transistors and will ship late Q1 2004. What I find strange is about R420 transistors numbers, it much smaller than sources told it will have 200M transistors and ATI told the next generation R420 chip will contain hundreds of millions of transistors, that 200M or over.

So what happened to 40M transistors and is ATI panicked 2nd times this time?

When ATI developed R400 and they got some information about NV40, decided that R400 will not good to compete NV40. ATI do not want to go into NV30 disaster path if they launched R400 and they realised that R400 project is very high risk strategy, they moved R400 up to R500 team and started a low risk strategy, R420 project based on the most successful R300 technology.

What happened to 40M transistors, here are 2 possible answers:

1) ATI send R420 design to TSMC to print out the prototype and send back to ATI. All test chips were probably failed, maybe 200M transistors are far too big for 0.13 micron process. ATI went back to the drawing board and cut off 40M transistors and redesign R420.

or

2) ATI got some final information about NV40 with 175M transistors and decided that the last 40M transistors are all worthless and useless, cant compete with NV40 and dont want to lose the crown to NV40 badly so ATI went back to the drawing board and cut off 40M transistors and redesign R420 reduced from 200M transistors to 160M transistors.

[-=DVS=-]

ATI told the next generation will contain hundreds of millions of transistors, that 200M or over in real R400 not R420 !

Heh you speculating like real fan boy :rolleyes: wait untill both products arrive , all info is humor/rumor mill to begin with....

[AthlonXP1800]

Originally posted by -=DVS=-
ATI told the next generation will contain hundreds of millions of transistors, that 200M or over in real R400 not R420 !


Well that was said in around March 2003, that was during R420 development, dont forget that R400 project was scrapped in late 2002.

R420 is the next generation after R300.
R400 is moved up to R500, it not the next generation after R300.

[Ninja Prime]

That article that mentioned 200m+ transitors was on the R500. LOL, your fanboi-ish specualtion amuses me though. :p

[AthlonXP1800]

Originally posted by Ninja Prime
That article that mentioned 200m+ transitors was on the R500. LOL, your fanboi-ish specualtion amuses me though. :p

At Rage3D forum and other sources (http://www.vr-zone.com/Home/news198/news198.htm), it said that R500 contain 300M transistors and to be manufacturing on 0.09 micron technology and it will be using DirectX 10. I could be surprise if ATI actually did cut back R500's original 300M transistors size to 200M transistors and the same thing with R420's original 200M transistors size to 160M transistors.

[-=DVS=-]

All of this is marely speculation and rumor sources are hardly correct esspecialy , looking at time frame R500 is years away and R420 is still some months..... and transistor count doesn't mean much at all :rolleyes: chip can be slow with billion transistors running at 10 MHZ ;) its actually good if you can make complex chip with as little as possible transistors means LESS HEAT AND FASTER CLOAKS :cool:

[MUYA]

hmmm 40 Million extra trannies (not the erm..lady boy variety). from 120/110 million. So that might mean that a rumoured 12X1 pipeline maybe unlikely for the r420. Maybe the 420 is just a r3XX with extra vertex shaders (dunno the amount) with VS3+ capabilities and an extra TMU on each of the pipeline and a boost in PS capaiblities ala PS3+ and also higher clocks? So in essence it is a 8X2 pipeline configuration with more whistles and bells than a 8X1 r3XX

Seeing that there were rumours that ATI got wind of NV's nv40 plans, might suggest that NV40 is now more definately a 8 pipeline GPU with 2 TMUS for each pipeline as well or however it's internal mechanisms are but u get the gist, it performs like 8X2 in all ops. (maybe 8 zixels? and 16 pixels? or somehting?) Only the rumour boyz can speculate more accurately.

Which brings me to this....any ideas or thoughts if the new gen q1 2004 GPUS will have 8 pipeline bretherens in the mainstream market?About time this happened we have been stuck with 4 pipelines since 20001?2002 GeForce ti200 in 2001? and Gf4 ti 4200 in 2002. Even the 57--U is a erm hybrid 4X1 pipeline architecture. 4X1 with some ops and 2X2 config with erm zixels? Will the new mainstreams still be 4 pipeline architecture like todays nv26 and rv360 etc? I understand that most of the costs of a 3D in high end cards is the memory...if that is not true please someone let me know. Anyway If they use low NS rated mmeory with a 256 bit memory interface and still retain a 8 pipeline GPU with some bits cut from the highend "GPU"for mainstream..wonder if it will be below $199 in costs?

I would hope so..but signs are showing probaly not. The Geforce FX 5900SE/XT are a good example of how it should be. The GeForce FX 5900SE are essentially the same your geforce FX 5900U ie same nuber of trannies...only its them chips that cannot clock as fast as the 5950U. So with that as an example, I hope for r4XX and nv4X mainstreams to be the same vein. That is if their yeilds are good I suppose too. But wouldn't it be nice to game in 4X aa as a default?

[StealthHawk]

The original GeForce from 1999 was a 4X1.

[Viral]

I still think R420 will have a transistor count in the high 100 million's.. around 180m IMO.

16x1 is possible, and the core looks to be between 450 and 550, the latter is fairly optimistic though..

I diddn't think 16x1 would happen afew months ago, but now i see it as a 50/50 chance between it being 12x1 or 16x1. If there are only 160mil transistors then 12x1 is probably the go, but i think ~180 could justify 16x1.

As for memory.. well, 1ghz DDR-II sounds about right, nothing mind blowing, but with 12x1 or especially 16x1, core and memory speed wont matter as much. Besides, 32gb/s memory bandwidth is still huge.

[MUYA]

Originally posted by StealthHawk
The original GeForce from 1999 was a 4X1.

I am well aware that the original geforce 256 wasa 4X1 card. In fact during the launch of the geforce 256 I was around nv news and rivazone for any speculations etc :D

Rivazone certainly had lively discussions didn't it? ;)

My cousin had it while I was struglling with my tnt2. It certainly wasn't a maintreams card...it was around $300 or so. Thats teh price my uncle paid for it in Singapore and that was only a few months after its launch. When I refer to mainstream cards I am refering to pricing of $200 or less. Like the gf 5700U and 9600XT etc


MUYA

[MUYA]

Originally posted by Viral
I still think R420 will have a transistor count in the high 100 million's.. around 180m IMO.

16x1 is possible, and the core looks to be between 450 and 550, the latter is fairly optimistic though..

I diddn't think 16x1 would happen afew months ago, but now i see it as a 50/50 chance between it being 12x1 or 16x1. If there are only 160mil transistors then 12x1 is probably the go, but i think ~180 could justify 16x1.

As for memory.. well, 1ghz DDR-II sounds about right, nothing mind blowing, but with 12x1 or especially 16x1, core and memory speed wont matter as much. Besides, 32gb/s memory bandwidth is still huge.

16X1 is probably just to complicted and a die hog for 130nm. I mean in terms of die size and yields. The more the trannies in a die as I understand it, the more chance of errors during the manufacturing process. Probably won't happen till 90nm and even then the fabs will be sweating bullets to get it right most probably. The jump to 90nm will also be a n30 ish one. Loads of problems getting yeilds upto acceptable levels. The more trannies u try to fit it, the more complicated its getting getting the yeilds right. smaller manufacturing process doesn't automatically mean you can stuff as much as you want :D

The nv30 according to Uttar and Mufu and other rumour hounds was meant to be 12X1 and 6X2 pipeline thing. It just didn't work out and it rumoured by them that NV took the decision to cut the pipelines to 8X1 and 4X2. All rumours of course. :D

I wish I could be optimistic of the chances of ATI or NV breaking into the more pipelines of a GPU theory but...doesn't look too good u nless TSMC/IBM/USMC have all managed to sort out with NVIDIA and ATI their yeild problems for big tranny count dies. If they do do that, I will be pleasantly surprised! :)


I dunno if ATI can do a similar approach to what AMD and INTEL are rumoured to be doing...dual core stuff. WOuld be intersting if NVIDIA and ATI did dual core stuff to increase the pipeline thing. Then again it would probably mean you having to re-mortage you house or sumat. ;)

Memory wise...bandwidth is king as we all know :D oooh I long for days when 4X aa can be had for games such as Doom III etc can be played with 4X aa out of the box :D I likey the idea of 38GB/sec memory bandwidth that is being tossed around. Wonder though if them memories will call for the erm NV30 ish cooling solution?

[nobie]

Noone ever said R420 had 200M transistors...this wasn't even a rumour just pure speculation. So, the speculation was wrong. It's not the first time that's happened.

[Viral]

Current info shows both ATi and Nvidia will move to 110nm before they go 90nm.. they seem to stop off at more tech nodes than the CPU guys do:p remember 150nm, whereas both AMD and Intel went from 180 to 130.

Oh and as for bandwidth.. well there was an article some months back showing nvidia licensing 1600mhz GDDR-II.. on a 256-bit bus thats 51.2gb/s. So yes, i would say thats enough to make your dream come true:D

[Uttar]

Muya: 160/110 = 1.45 -> 1.5 due to efficiency improvements, I guess, and that not everything is +45%. Your 120-130 number is for the NV3x, not for the R3xx. Thus, 8x1->12x1. Not 6x2 though. Anymore questions? :p

Viral: Yes, they bought 1600Mhz GDDR2, but they're going to clock it at 1500Mhz AFAIK, unless their memory plans changed due to the November->February delay...


Uttar

[StealthHawk]

Originally posted by MUYA
My cousin had it while I was struglling with my tnt2. It certainly wasn't a maintreams card...it was around $300 or so. Thats teh price my uncle paid for it in Singapore and that was only a few months after its launch. When I refer to mainstream cards I am refering to pricing of $200 or less. Like the gf 5700U and 9600XT etc


MUYA

How about the r9500Pro then, it had 8 pipelines and was a mainstream card priced at $200 :D

[Nutty]

Anything over 100M is "hundreds of millions" anyway. Hundreds just means more than 100. 140 is still hundreds of millions.

[Maverick123w]

Wouldn't hundreds mean anything 200 and over :angel2: Seeing that there is only 1 hundred in 100-199

[MUYA]

Originally posted by StealthHawk
How about the r9500Pro then, it had 8 pipelines and was a mainstream card priced at $200 :D

Ya :D Thats what I am talking about, but those things disappeared real quick :(

[Nutty]

Wouldn't hundreds mean anything 200 and over Seeing that there is only 1 hundred in 100-199

No. 150 is 1.5 hundreds, which is more than 1, hence plural.

[Uttar]

150 < 160 < 175.
I wonder if NVIDIA knew this...


Uttar

[-=DVS=-]

Everyone talking how much R420 will have transistors how about NV 40 ?

[Ninja Prime]

Originally posted by -=DVS=-
Everyone talking how much R420 will have transistors how about NV 40 ?

Nvidia is afraid to tell us about NV40, they might lose their secret powers of... BENCHMARK CHEATING! Oh noes! :D

[Uttar]

Originally posted by -=DVS=-
Everyone talking how much R420 will have transistors how about NV 40 ?

You ARE fricking kidding me, right? You MUST be fricking kidding me! This is ridiculous!
How many times will I have to say 175M? :( And we knew that wayyyy before we knew that 160M number for the R420.


Uttar

[-=DVS=-]

Originally posted by Uttar
You ARE fricking kidding me, right? You MUST be fricking kidding me! This is ridiculous!
How many times will I have to say 175M? :( And we knew that wayyyy before we knew that 160M number for the R420.


Uttar

And you will bet your life on it that NV40 is 175 ? :rolleyes:

[Ninja Prime]

Originally posted by Uttar
150 < 160 < 175.
I wonder if NVIDIA knew this...


Uttar

If NV40 is 175million, that doesn't exactly mean its gonna be better/faster. Look at NV30... 15million more (~12%)than R300, but waaaay slower. 175million is 9% more than 160million of R420... R420 might still be twice as fast like NV30-R300. Especially since NV40 is supposed to use a similar tech as NV30-35.(or so I hear)