ATI R350 any info?

[runsalone]

Does anyone have any info on the new R350 GPU? I've heard that if nvidia ever gets their geforce fx on the shelves, ati will be shipping their R350 and RV350 graphics accelerators.

[Default]

they will be shipping soon anyways :D Just to leave nvidia in the dust and distroy thier luanch

[PreservedSwine]

RV350... .13 micron design, probably a little worse performer than the R300...

R350... .15 micron design, souped up R300 core, with 500mhz DDRI or DDRII ram, higher core clock...a real beast of a card..
Though ATI may not release it if the NV30 doesan't really outperform the R300...

[pip]

both the cards will sport 256 megs of RAM

[Cisco Kid]

Originally posted by pip
both the cards will sport 256 megs of RAM



and it really is not needed,but it will be there in reserve LOL

cisco kid

[augustwest]

i sit here with my radeon 7500 and whimper.......:drool: .........9700's are considered acceptable donations!

[Evnas]

Originally posted by pip
both the cards will sport 256 megs of RAM

I doubt it. 128mb card are but 6 months old, and 64mb has been used since the GF2...we will have 128mb cards for the next while. Especially with DDR2, as thats just way to expensive right now, and Samsung is the only company who has their design done.

[raven]

Originally posted by augustwest
i sit here with my radeon 7500 and whimper.......:drool: .........9700's are considered acceptable donations!

i sit here with my 8 ati meg rage pro turbo.......:drool:........7500's are considered acceptable donations! :D

[raven]

Originally posted by PreservedSwine
RV350... .13 micron design, probably a little worse performer than the R300...

R350... .15 micron design, souped up R300 core, with 500mhz DDRI or DDRII ram, higher core clock...a real beast of a card..
Though ATI may not release it if the NV30 doesan't really outperform the R300...

the R350 will be DDRII. and will totally destroy the NV 30

but about the micron design.... wouldnt it be .09 for the R350? i dont know much about this... but intel moved form .13 to .09 so that they could get them faster... so i would asume this would be the same for ati chips?

raven

[pip]

what the hell you talkin bout boy, intel isn't moving to 90 nm til prescot(oh sweet prescot when will you be released)

[PreservedSwine]

Originally posted by raven


the R350 will be DDRII. and will totally destroy the NV 30


raven

It has already taped out @ .13, and neither ATI nor Nvidia is even close to making anything @.09. It seems that TSMC has only just recently made tooling changes for Nvidia on .13,, and they are no where near prepared for a forway into .09. Think about it, the NV30 on .13 already needs at least a 10 layer PCB......

As far as 500mhz DDRII, hold onto your hat on that one....
Yes, the R300 core supoorts DDRII, but ATI already has deals w/ Hyunix, and they have recenty come out w/ 500mhz DDRI SDRAM. The perfromance between DDRI and DDRII is negligable, and the DDRI is *much* less expensive. Yes, the R300 *may* use DDRII, but don't think just because it can, that it will. Just alittle something to chew on...

[Cisco Kid]

Originally posted by augustwest
i sit here with my radeon 7500 and whimper.......:drool: .........9700's are considered acceptable donations!

funny thing is your 7500 will do the trick a while longer, what have you got it overclocked to?? Is it a BBA card.

Nice thing though is that card can support 1024x768 HDTV /TV out!!

Mine would do 340/270 and it scored decent marks in 3D MArk2K1SE just a hair under 6000

cisco kid

[raven]

Originally posted by pip
what the hell you talkin bout boy, intel isn't moving to 90 nm til prescot(oh sweet prescot when will you be released)

oops.... i meant will move to 90nm ... my bad:D

Originally posted by PreservedSwine
It has already taped out @ .13, and neither ATI nor Nvidia is even close to making anything @.09. It seems that TSMC has only just recently made tooling changes for Nvidia on .13,, and they are no where near prepared for a forway into .09. Think about it, the NV30 on .13 already needs at least a 10 layer PCB......

As far as 500mhz DDRII, hold onto your hat on that one....
Yes, the R300 core supoorts DDRII, but ATI already has deals w/ Hyunix, and they have recenty come out w/ 500mhz DDRI SDRAM. The perfromance between DDRI and DDRII is negligable, and the DDRI is *much* less expensive. Yes, the R300 *may* use DDRII, but don't think just because it can, that it will. Just alittle something to chew on...


i guess that my info has gone out of date:( i just got confused when you said the RV350 will have a .13 micron and preform slower, and the R350 will have a .15 mocron chip and will run faster

hope i didnt get anyone's hopes too high:p

[PreservedSwine]

hope i didnt get anyone's hopes too high

Actually, I completley agree...

a 500mhz DDR (whether it be DDR1 or DDRII doesn't really matter, it's still DDR) core enhanced R350 is a monster of a card. With it's robust 256 bit DDR bus, and efficient design, would absolutley clobber the NV30, if released benchmarks on the NV30 are anywhere near accurate....:)

[DaddyB]

Actually true DDR2 has double the bandwidth compared to DDR1 since it uses a 4bit data prefetch compared to DDRs 2bit prefetch. Unfortunetly both ATi and Nvidia are using DDR2 in a "DDR1 compatability mode" so that it only uses a 2bit prefetch and performs on par with DDR.

As far as I have heard R350 will be a revamped R300 core, clocked much higher and still using a 0.15u core along with DDR2.

[Rafterman223]

They've already said it'll have a .13 core, and DDR2. it may have 256 ram but only to be used as a marketing tool

[Evnas]

Originally posted by DaddyB
Actually true DDR2 has double the bandwidth compared to DDR1 since it uses a 4bit data prefetch compared to DDRs 2bit prefetch. Unfortunetly both ATi and Nvidia are using DDR2 in a "DDR1 compatability mode" so that it only uses a 2bit prefetch and performs on par with DDR.

As far as I have heard R350 will be a revamped R300 core, clocked much higher and still using a 0.15u core along with DDR2.

DDR2 does NOT have double the bandwidth...that would be a totaly different type of ramd (QDR - Quad Data Rate)...DDR is STILL Double Data Rate, as in 2, not 4. Its not like RDRAM

Originally posted by Rafterman223
They've already said it'll have a .13 core, and DDR2. it may have 256 ram but only to be used as a marketing tool

ATi never said it would have DDR2...infact, Samsung (the only company producing DDR2 right now) has said nVidia is the only company that has placed any orders for DDR2

[DaddyB]

I read the JEDEC white papers on DDR2 long ago and again a few weeks ago, I know I am correct. True DDR2 does have double the bandwidth of DDR1 and if you reread my post you will see that I said DDR2 use a 4 bit data prefetch, DDR uses a 2bit data prefetch, I didnt mention how many transfers per clock it does as it is a little obvious (Double Data Rate RAM). QDR is completly different, it transfers data twice on the rising and twice on the falling edge of a clock cycle.

I read the withepapers because a lot of people, such as yourself, dont know much about DDR2 so I thought I would find out about it and share my knowledge. Just because you dont know much about new/upcoming technologies don't assume no one knows anything.

If you want to know about DDR2 then I suggest you go to JEDEC.com and read about it yourself, as I did. If you don't want to spend a few hours learning then just click here (http://www4.tomshardware.com/business/20021120/comdex_day_2-01.html#nvidia_geforce_fx_memory_bandwidth) for verification that I am right. Just read the few paragraphs below the "Geforce 4 Ti" heading.

With all of the new RAM standards coming out (DDR2, QDR, GDDR3...) I know it can be confusing to understand them all, which is why I chose to learn about them rather then listen to people who only know half the story.

[PreservedSwine]

SO, just because the prefetch is larger, that somehow = twice the bandwidth??
I disagree as well, otherwise, they would not call it DDR RAM. It would be QDR Ram. As far as I can tell, the pre-fetch changing merely allows for a much higher clockspeed, not doubling the already doubled bandwidth..

Also, I've heard rumors the R350 is going to be 8x2.......seems alot for .15, anyone else hear that?

[DaddyB]

Sigh..... Again, GO READ THE WHITE PAPAER. OK aparently you didn't even bother to click the link or go to JEDEC.com to see that you are wrong before posting so Ill just post the relevant parts of the white paper here along with a link to the white paper on JEDEC's website.

Survey results from Dram vendors on CAS cycle time
Assumptions: ~2002-2003 time frame
256mbit device
Achieve high yield for x4, x8, x16 devices
-----------------------------------------------------------------------------------
Supplier | Cycle Time | -----Max pin-- | -----Max pin-- | ----Max pin
--------------------------------frequency ----- frequency ---- frequency
--------------------------------prefetch 2 ----- prefetch 4 ---- prefetch 8


A ----------- 7.5ns -------- 266 mbit/sec ---- 533 mbit/sec ---- 1066 mbit/sec
B ----------- 6.6ns -------- 303 mbit/sec ---- 606 mbit/sec ----1212 mbit/sec
C ----------- 6.0ns -------- 333 mbit/sec ---- 666 mbit/sec ---- 1333 mbit/sec
D ----------- 6.7ns -------- 298 mbit/sec ---- 597 mbit/sec ---- 1194 mbit/sec
E ----------- 6.5ns -------- 308 mbit/sec ---- 615 mbit/sec ---- 1230 mbit/sec
F ----------- 6.5ns -------- 308 mbit/sec ---- 615 mbit/sec ---- 1230 mbit/sec
G ----------- 6.5ns -------- 308 mbit/sec ---- 615 mbit/sec ---- 1230 mbit/sec
H ----------- 6.5ns -------- 308 mbit/sec ---- 615 mbit/sec ---- 1230 mbit/sec

As you can see from that chart changing the prefetch from 2 to 4 doubles the bandwidth at any given speed, doubling the prefetch again from 4 to 8 would indeed also double the bandwidth. So now we know that changing the prefetch to 4 WILL double the bandwidth and thats all well and good but how do we know that DDR2 is supposed to use a 4 bit prefetch? Again, its in the white paper:

• Dram core utilizes a prefetch of 4
• Decreases frequency of dram core relative needed to support
increased data rate
• Dram core runs at 1/4 data bus frequency


Also note the part about the core running at 1/4 of the data bus frequency. Think of what we now call 266mhz DDR or pc2100, we know it doesnt run at 266mhz - it runs at 133mhz but writes twice per clock cycle, effectivly doubling the badwidth. So the core runs at 1/2 the data bus frequency, when we increase the prefetch to 4 it again doubles the bandwidth allowing a 133mhz core to have an effective bandwidth of 533mhz (4x the 133mhz core).

The quotes are taken from page 6 of the JEDEC white paper on DDR2, which you can see here (http://www.jedec.org/DOWNLOAD/pub21/HotDDR/DDR_II_evolution.pdf) (it is in pdf format so you will need adobe acrobat or another pdf viewer).

And once again I must say just because it has twice the bandwidth of DDR it DOES NOT MAKE IT QDR - QDR writes 4 times per clock cycle - DDR2 still writes twice per clcok cycle. Think of it this way, DDR2 is writting twice per clock cycle but the chunks of data it's writting are twice the size of the chunks DDR1 would write.

So if you still disagree with this there is no point in posting in here as anyone who understands this stuff or has taken the time to actually read the links I posted now knows that I am right and why. If you disagree with it then just email JEDEC and ask them why they didnt make the DDR2 standard the way you wanted them to make it.

Does anybody else think I should post all of this in the Memory forum in the hopes it will be made a sticky so I won't have to keep having this same discussion?

[raven]

any knwo approximately when these cards will be comming out? om personally just interested in the AIW version of these cards.. anyone hear of any approx dates?

[Evnas]

March is the word on when they will come out.

[augustwest]

im assuming the prices of the 9700's will come down with this release as well as nvidias fx release. am i making a wise assumtion or should i just get a damn job?

[raven]

Originally posted by Evnas
March is the word on when they will come out.

aww.. thats sooo far away.. if only it came out in february... then id go with the AIW R(V)350 =)

[Evnas]

Originally posted by raven


aww.. thats sooo far away.. if only it came out in february... then id go with the AIW R(V)350 =)

Its the month after February, so that shouldnt be a problem, hehe...come on now...ATi has to give nVidia the crown back for at least a little while before they snatch it back ;)

[PreservedSwine]

Originally posted by DaddyB
Sigh..... Again, GO READ THE WHITE PAPAER.
And once again I must say just because it has twice the bandwidth of DDR ....

OK OK, I read it I read it
:)

It is interesting to note, unless I have misinterpreted NV30 information and DDRII info, if the NV30 core speed was doubled, using the same DDRII still at 500MHZ, would then be truly optimized to run in a true DDRII compatability mode, and offer double the memory bandwidth it currently offers, yet with no bump in memory speed?

What would then be the benifit to rasing the core speed to 750, and dropping the memory to 375 DDRII? I assume to need a 2-1 ratio in core speed to ram speed to use the DDRII in it's native mode...The DDRII, although running a much lower 375mhz, would then offer (2) times the bandwidth of 500mhzDDRII coupled w/ a 500MHZcore?

[DaddyB]

Well if the DDR2 was running at 375mhz in "DDR2 mode" it would have an effective badwidth equal to 1500mhz DDR, suitable for a 750mhz core. Thats supposed to be one of the advantages of DDR2, it should be cheaper 'cause you can use slow modules but have greater bandwidth (you won't have to use ultra fast 2ns RAM to achieve the bandwidth 1ghz RAM would provide. you can use slower 4ns RAM which would be much cheaper).

Imagine the next generation of cards with 256bit buses and true DDR2 (or GDDR3 for that matter), 50GB/s+ of real bandwidth (none of that 4:1 compression/effective bandwidth nvidia is talking about). Evidently the next limitation in vid cards will be the core speed and with the extra voltage (9700 + GFFX) and extra cooling (GFFX) required to reach the speed/performance of todays GPUs even 0.9u cores won't help much.

[Darkseid]

i really don´t know how are they going to keep rising speed of the cores, die shrink is not enough

[Evnas]

IMO, nVidia has always had very inneficient core designs. Hell, they said the .15 process could in no way reach 400Mhz...ive seen the 9700 Pro pushed to 430MHz...they really need to revamp their engineering team to something more efficient

[raven]

Originally posted by Evnas
IMO, nVidia has always had very inneficient core designs. Hell, they said the .15 process could in no way reach 400Mhz...ive seen the 9700 Pro pushed to 430MHz...they really need to revamp their engineering team to something more efficient

i second that... Nvida has to do something so that they can still put a lil pressure on ATI so that ATI prices can stay reasonable:D

[crotale]

Originally posted by DaddyB
I read the JEDEC white papers on DDR2 long ago and again a few weeks ago, I know I am correct. True DDR2 does have double the bandwidth of DDR1 and if you reread my post you will see that I said DDR2 use a 4 bit data prefetch, DDR uses a 2bit data prefetch, I didnt mention how many transfers per clock it does as it is a little obvious (Double Data Rate RAM). QDR is completly different, it transfers data twice on the rising and twice on the falling edge of a clock cycle.

I read the withepapers because a lot of people, such as yourself, dont know much about DDR2 so I thought I would find out about it and share my knowledge. Just because you dont know much about new/upcoming technologies don't assume no one knows anything.

If you want to know about DDR2 then I suggest you go to JEDEC.com and read about it yourself, as I did. If you don't want to spend a few hours learning then just click here (http://www4.tomshardware.com/business/20021120/comdex_day_2-01.html#nvidia_geforce_fx_memory_bandwidth) for verification that I am right. Just read the few paragraphs below the "Geforce 4 Ti" heading.

With all of the new RAM standards coming out (DDR2, QDR, GDDR3...) I know it can be confusing to understand them all, which is why I chose to learn about them rather then listen to people who only know half the story.

According to this page, http://www.theinquirer.net/?article=3030 (http://www.theinquirer.net/?article=3030), DDRII do not have double up of bandwidth compared to DDRI
Only 50%

[DaddyB]

Well that article begins by saying: "READER SEB GEE has put together his thoughts on future plans for double data rate (DDR) memory." and goes on to say: "there's an element of speculation in his thoughts".

So it seems that it was just written by a reader and if you look at the date you will see it was written close to a year ago - before the standard was ratified.

Anyway I have provided a link above directly to the JEDEC whitepaper where it says it has double the bandwidth, the relevant parts I quoted and posted above but if you dont want to take my word for it just read the whitepaper yourself.

Getting the info straight from the horses mouth is more accurate then year old speculation IMO.

[PreservedSwine]

Originally posted by DaddyB
Well if the DDR2 was running at 375mhz in "DDR2 mode" it would have an effective badwidth equal to 1500mhz DDR, suitable for a 750mhz core. Thats supposed to be one of the advantages of DDR2, it should be cheaper 'cause you can use slow modules but have greater bandwidth (you won't have to use ultra fast 2ns RAM to achieve the bandwidth 1ghz RAM would provide. you can use slower 4ns RAM which would be much cheaper).

Imagine the next generation of cards with 256bit buses and true DDR2 (or GDDR3 for that matter), 50GB/s+ of real bandwidth (none of that 4:1 compression/effective bandwidth nvidia is talking about). Evidently the next limitation in vid cards will be the core speed and with the extra voltage (9700 + GFFX) and extra cooling (GFFX) required to reach the speed/performance of todays GPUs even 0.9u cores won't help much.

It appears we missed something....
It appears that both DDRI and DDRII have the same bandwidth after all...... Note that you must think of DDR and DDR-II as having three different frequencies. Memory Cell, I/O Clock, Data Rate. What's important to know is that the I/O CLOCK is the actual MHz of the ram chip....NOT THE memory cell array. 400 Mhz DDR-I has an I/O clock of 400 Mhz, and 400 Mhz DDR-II has an I/O clock of 400 Mhz.

(so in the above illustration, you're looking at a 100 Mhz DDR-I module, and a 200 Mhz DDR-II module. NOT a 100 Mhz DDR-II module!)

All DDR-II modules are in fact "quad-pumped" relative to the memory cell array. However, memory cell array frequencies have NOTHING to do with the stated frequency of the ram chip! It's an "internal" clock.

What does this mean?

It means that the "500 Mhz DDR-II" on the GeForceFX is running at an i/o clock frequency of 500 Mhz, for a data frequency of 1 GHz. This is exactly the same that a 500 Mhz DDR-I module would be. The difference is, the 500 Mhz DDRII module has in internal cell array frequency of 250 Mhz....HALF that of a 500 MHz DDR-I module.

So in short, it's helpful to think of it this way: DDR-I and DDR-II are both effectively double pupmped with respect to the reported chip clock frequency. And that is what is "most important" in terms of determining the bandwidth. However, DDR-II's "internal" cell array frequency is one half of the chip clock.

Sorry to cut-n-paste, but here's the link....
http://www.nvnews.net/vbulletin/showthread.php?s=&threadid=6015

DDR-II runs at half the clock and twice the bandwidth internally relative to DDR-I, but the I/O data clock is the same.

DDR-II 500MHz = DDR-I 500MHz as far as theoretical bandwidth is concerned.

[Evnas]

Originally posted by PreservedSwine


It appears we missed something....
It appears that both DDRI and DDRII have the same bandwidth after all......
Sorry to cut-n-paste, but here's the link....
http://www.nvnews.net/vbulletin/showthread.php?s=&threadid=6015

DDR-II runs at half the clock and twice the bandwidth internally relative to DDR-I, but the I/O data clock is the same.

DDR-II 500MHz = DDR-I 500MHz as far as theoretical bandwidth is concerned.

Thats exactly what i was looking for! DDR2 internally is twice as fast as DDR1, but when it comes to actual bandwidth, they are both exactly the same. I just never bothered to argue it cuz i didnt have the source.

[DaddyB]

Originally posted by PreservedSwine



It appears that both DDRI and DDRII have the same bandwidth after all......



Hmm, well I guess that depends on how you look at it. I just finished reading that other thread and the lost circuits article, the only thing that they are saying that disagrees with what I said is the naming convention (ie. 500mhz DDR).

I either read or assumed (read too much about it now, can't remember) that the GFFX would be using 2ns RAM, which would mean that it is running in DDR1 mode, it is possible that it is using 4ns RAM in true DDR2 mode - either way it works out the same (1ghz effective).

So the only thing that is different is that they are saying that 500mhz DDR2 is running at a core speed of 250mhz and has an effective bandwidth of 1ghz, whereas I was saying that 500Mhz DDR2 is running at a core speed of 500mhz, therefore it would have 2ghz effective bandwidth.

I am not sure which is the correct naming convention, after all what speed does pc2100 ram run at? of course you (or I) would answer 266mhz, which isn't exactly correct - it runs at 133mhz but writes twice per clock (266mhz effective). so what do you call RAM that runs at 1ghz, writes twice per clock but has a prefetch of 4? 250mhz DDR2? 500mhzDDR2? 1ghzDDR2? after all it has a core speed of 250mhz but writes twice per clock (500mhz) and has a prefetch double that of DDR (1ghz).

Other then that what they are saying is the same as what I am saying - 10ns (100mhz core) DDR runs at 200mhz, 10ns DDR2 runs at 400mhz - double DDR1.

Do you get what I am saying? even I'm confussed now...

[DaddyB]

Apparently I did read that the GFFX will have 2ns RAM, I just stumbled across an article which says so that I know I have read before. its from the inquirer so I don't know that it is really correct though. Linky (third paragraph) (http://www.theinquirer.net/?article=7021)

So if that is correct then what they are saying in that other thread is wrong, the GFFX's DDR2 would not be running in true DDR2 mode. If they are right then the GFFX would be using 4ns RAM.

4ns RAM = 250mhz core *2 (DDR) *2 (double prefetch) = 1000mhz

2ns RAM = 500mhz core *2 (DDR) = 1000mhz

Since we all know that the effective bandwidth of the GFFX is 1000mhz (16GB/s) it has to be one of the two configurations I posted above and if it is using 2ns RAM as that article says then it must not be true DDR2, otherwise it would be 2ghz effective and Im sure if that was true nvidia would have corrected all of the sites that reported the GFFX as having a mere 16GB/s of actual bandwidth.

[PreservedSwine]

That makes no sense...why shell out the X-tra $$ for DDRII if your going to run it at DDRI mode...there must be something else at work here.....

[DaddyB]

Well DDR2 uses less voltage, thus produces less heat so it is easy to clock it higher. I am not aware of any DDR1 RAM that can run at 500mhz.

[PreservedSwine]

Originally posted by DaddyB
Well DDR2 uses less voltage, thus produces less heat so it is easy to clock it higher. I am not aware of any DDR1 RAM that can run at 500mhz.
Hyunix.....

[Evnas]

Originally posted by DaddyB
Well DDR2 uses less voltage, thus produces less heat so it is easy to clock it higher. I am not aware of any DDR1 RAM that can run at 500mhz.

Hynix has released DDR1, and Hynix happens to be ATi's #1 memory supplier....thus the reason i highly doubt ATi will do with DDR2 at all, and just got to GDDR3 when they are done with it.