R350 hype is already starting!

[Megatron]

Reactor Critical seems to be starting the hype ball rolling on the next radeons.
No real info/specs..just some "its coming" hype, and some speculations.

http://www.reactorcritical.com/cgi-bin/news.cgi?templ=single.http&id=1297

[-=DVS=-]

Soon we will hear *wow next year will be so awesome NVxx , Rxxx* :p , i hear similar statements every year :D

[Grrrpoop]

Both ATI and nVidia made a decision not to go with 8x2 pipes with their latest Gen of cards, I certainly don't see ATI's refresh card (R350) offering such a complex update of features.

Maybe R400, but imo I think R350 is just a R300 with Go-Fasta stripes. :D

Faster core, maybe DDR2, but I don't think we'll be getting more TMU's per pipe. Parhelia demonstrates that more TMU's does not a good card make ;) Tbh even DDR2 seems unlikely as I read a recent-ish ATI quote saying DDR2 equipped cards were in their opinion too expensive right now. The card supports DDR2 so I think Hercules or Gigabyte or whoever could bring out a DDR2 card but would the benefit over DDR1 be worth it?

I have a feeling they might be introducing some 256mb cards (to take advantage of UT2k3 and games based on that engine which in theory could use beyond 128mb of textures). Again, I doubt ATI themselves will do this, but nothing's stopping Herc doing it, they might be making their watercooled R350 some kind of OC'd uber-card :bounce: lol

I think ATI is just bringing out R350 to deny nVidia top spot again, but it was all the rumours of higher than expected specs which pissed a lot of ppl off over GFFX, I hope this isn't gonna happen with R350. IIRC it was probably RC who were one of the first stating GFFX would be 8x2, and we know how accurate that turned out to be :rolleyes:

[Gar]

I would bet a lot of money the R350 does not have 2TMUs as it would really not provide that much more increase in performance in "most" games. ATI most likely enhanced certain bandwidth saving techniques, maybe added some newer ones and increased the clock. They have a lot of room for improvement with the R300 core (considering how much of a performer it was in the first place) especially with the clock speeds. I bet we will see something different with the R350 making it more than an Ultra version of the R300.

[tamattack]

Actually, I think that a second TMU would provide a sizeable performance boost in today's games where multi-textured fillrate is still far more important than pixel shader speed.

Such a design would probably own the benchmarking crown vis-a-vis GFFX or R9700. That factor alone might make it worthwhile to engineer.

NB: not that I believe R350 to be 8x2, just that there could be justification for it.

[Gar]

Definitely would be in certain games and benchmarks, however, with the games and benchmarks of today (most aren't multitexture heavy) it probably won't be worth the investment by adding the 2nd tmu as you and I know adding another texture mapping unit isn't quite so simple and requires some extensive changes to the core. I think we can use the Pharelia (spelling?) as an example of how the extra tmu doesn't mean better performance.

[Grrrpoop]

Originally posted by Gar
I think we can use the Pharelia (spelling?) as an example of how the extra tmu doesn't mean better performance.
Yeah, like I said in my previous post, the Parhelia is the perfect example of why more TMU's don't automatically equate to better performance.

Parhelia has 4 pipes with 4 TMU's PER PIPE. It's still beaten hands down by the Ti4600 which is 4x2, and more importantly, the Radeon 9500 which is 4x1.

The deciding factor against Parhelia is its low core/mem clockrates.

If ATI can pump the R350's core up to at least 400Mhz, preferably 450MHz (500Mhz??) then I think it can beat GFFX by the rumoured 10% (lol stating a rumoured 10% more than the rumoured abilities of an unreleased card, tenuous..).

[StealthHawk]

Parhelia's weakness is its lack of bandwidth savings techniques, whereas its competition has some. it's really as simple as that. also i'm sure ATI and nvidia both have better(read: more optimized for performance) drivers than Matrox.

[zakelwe]

With the ease of overclocking video card core chips it will be interesting to see how overclockable the FX core is. If they have followed the first AMD 0.13 micron Tbreds then not much, but if they have got it right and followed Intel and AMD B revision then it might be quite considerable. Compelx chip of course which does not help.

If it is of course then the DDR II will really be holding it back with all the goodies turned on. I get the feeling 2ns DDR II memory will not be able to overclock that well.

We'll see though, probably turn out to be terrible to overclock the core and easy to massively overclock the memory ....

Regards

Andy

[Gar]

I would be very very surprised if you can get more than 50mhz core overclock out of the FX. The chip itself is basically an overclocked version of what they initially intended to use. Why else would there be a gigantor dust buster heatsink/fan unit and an enormous copper heat spreader on the back of it. Basically you are receiving an already overclocked card, which I think anyone is going to have a tough time overclocking. Of course I could be completely wrong, but hey, what else is the 'rumor mill' for?

[suburbanguy]

What about the possibility that R350 is 8:1 - just 1 TMU per pipe, but each "TMU" can pump out 2 (or even 4) textures per clock? The team behind R300 and thus R350 have many of the people that made GC's Flipper chip. Flipper has a single TEV (an uberTMU) for the whole chip. the TEV can do 4 trilinear filtered texels per clock.

This is a bit of a stretch (but not so much so) but perhaps R350's
TMUs are more like Flippers TEV. but instead of 1 TEV in the Flipper producing textures for 4 pipelines, you have 8 TEVs, one in each of R350's 8 pipes. if you could get more texels per cycle out of a single texturing unit per pipe, you could save a large amount of transistors. that alone would justifiy this R350 refresh.

[lagadu]

Originally posted by suburbanguy
What about the possibility that R350 is 8:1 - just 1 TMU per pipe, but each "TMU" can pump out 2 (or even 4) textures per clock? The team behind R300 and thus R350 have many of the people that made GC's Flipper chip. Flipper has a single TEV (an uberTMU) for the whole chip. the TEV can do 4 trilinear filtered texels per clock.

This is a bit of a stretch (but not so much so) but perhaps R350's
TMUs are more like Flippers TEV. but instead of 1 TEV in the Flipper producing textures for 4 pipelines, you have 8 TEVs, one in each of R350's 8 pipes. if you could get more texels per cycle out of a single texturing unit per pipe, you could save a large amount of transistors. that alone would justifiy this R350 refresh.

my thought exactly, using something akin to flipper's TEV unit would probably be the best option, they already did it once for the flipper, it wouldn't be hard to make it happen for the R350, it'd result in something equivalent to a 8x2 (or maybe 8x3) with only 10-15% more transistors due to the TEVs being more complex than TMUs...
another option that i read somewhere that i think would be interesting would be some sort of "TMU pool" where each TMU could be independent of it's pipeline, thus being used where it'd provide the most benefits, this could efectivly deny the disadvantage of pixel shading VS multitexturing, as in this situation, each TMU would only come into use where really needed, without unused TMUs anywhere...

[tamattack]

Originally posted by Gar
Definitely would be in certain games and benchmarks, however, with the games and benchmarks of today (most aren't multitexture heavy) it probably won't be worth the investment by adding the 2nd tmu as you and I know adding another texture mapping unit isn't quite so simple and requires some extensive changes to the core. I think we can use the Pharelia (spelling?) as an example of how the extra tmu doesn't mean better performance.

The second TMU can make a significant difference (but obviously does not double performance), as many of today's games are dual texture aware. But anything beyond two will suffer the same problems as the original Radeon in that the third and/or fourth (as in Parhelia) are heavily under-utilised (if at all) in today's games.

Originally posted by Grrrpoop
Parhelia has 4 pipes with 4 TMU's PER PIPE. It's still beaten hands down by the Ti4600 which is 4x2, and more importantly, the Radeon 9500 which is 4x1.

The deciding factor against Parhelia is its low core/mem clockrates.

See above. If we factor out two of the Parhelia's four TMUs, then the performance difference vis-a-vis the Ti4600 is inline with what you would expect based on the clock speed differences alone.

As for R9500 (non-Pro), I'm having a hard time finding reliable benchmarks on it... so I'd like to see where you got that info.

Originally posted by StealthHawk
Parhelia's weakness is its lack of bandwidth savings techniques, whereas its competition has some. it's really as simple as that. also i'm sure ATI and nvidia both have better(read: more optimized for performance) drivers than Matrox.

Yes, this is a big part of it. Combined with the leisurely clock rate and the lack of effective use of the 3rd and 4th TMUs.

I guess the point of this long message is that I still firmly believe that a second TMU will make a significant performance impact on today's games and benchmarks.

[rwolf]

"Mr Kwok Yuen Ho, the CEO and chairman of ATi Technologies has given a speech during the press announcement and presentation of the Radeon 9000 and 9700 at China today. Here are some of the interesting highlights during the Q&A session with Mr Kwok :

Based on the Radeon 9700 technology with parallel connection of up to 256, ATi is currently examining the possibility of having Dual Radeon 9700 VPUs on a card. Also Quad display outputs is possible with such card. However, there is no official schedule on when such card will be launched. Most likely the target segment for Dual VPU card is for simulation market.
Radeon 9000 for mobile segment will be released within 2 months and the mobile edition of Radeon 9700 will be taped out in time to come.
The Lite edition of Radeon 9700 which is possibly the Radeon 9500 will be targeting at the mainstream market will have the number of rendering pipelines and memory channels reduced by half.
ATi does not have plan to make Radeon 9700 to be based on 0.13 micron process but the next generation R350 will and it is scheduled to be launch by first half of next year and supports DirectX 9.1
ATi is currently partnering with TSMC as their main foundries but UMC will be producing for them as well." -- vr-zone

This is a news story from vr-zone ( http://vr-zone.com) that was dated about august of last year. All the stuff in the article has come to pass. (There is a company in the simulation market making a dual 9700 card) . The only thing that is inconsisted with the most recent rumors is that the R350 will be on the .13 micron process. Isn't it a coincidence that the chip will be coming out the same time as the nv30 or a bit sooner/later? ATI would ha

Is, but the number of transistors of the RADEON 9700 which stuffs this much function approximately 1 hundred million 1 10000000 or more, the RADEON of previous life generation 8500 (the R200) from 6,000 ten thousand increased to approximately 2 times. And, the ATI produces this, with current 0.15 mu m process. The reason the David E Orton of the ATI you explain the president and the COO as follows.

" The party of the processor always, architecture and process technology simultaneously does not have to change, you say. In other words, the architecture which is stabilized new process technology or new architecture it must be combination of the process technology which is stabilized, unless. Our strategy has paralleled to that. In other words, you start new RADEON 9700 with 0.15 mu m, the following generation, this is the R350, but it moves to 0.13 mu m there " - impress.co.jp (translated with babelfish

Here we have a quote from David Orton saying R350 is .13 microns.


The hurdle is high 0.13 mu m process where


"This time, the fact that the light and darkness of the NVIDIA and the ATI was divided is process technology. Because " them (the NVIDIA) that this time you have taken time it has been about to produce with 0.13 mu m. 0.13 When it is mu m, also the hurdle of the design becomes high. As for success of this corporation, prudent method (RADEON 9700 with 0.15 mu m is produced) you have taken ", the K Y Ho person emphasizes.

There is a special circumstance of 0.13 mu m process in this. With the latest 0.13 mu m process, many semiconductor enterprises, took time with yield rate improvement. The Edward K Chang of the TSMC which produces the tip/chip of the ATI and the NVIDIA the person (the Marketing Manager, the Platform Technology Marketing Division) you explain as follows. " With 0.13 mu m, () the learning curve of yield rate improvement was slow in comparison with the former process generation. Though this is the problem which everyone holds. The reason copper wiring which is the technical element which joins newly, low induced electricity (Low-k) ratio (during wiring the membrane) it was the material sees "

Though, as for the NVIDIA you insist with the NV30 which is produced with 0.13 mu m there is an advantage in the efficiency aspect. " As for the R300 because it is produced with 0.15 mu m process technologies, it is difficult to make high speed, (in present announcement 300MHz +)" " because the NV30 is 0.13 mu m process, it can make the directly fast clock probably will be, " (the NVIDIA, the David B Kirk person, the Chief Scientist)

The K Y Ho person of the ATI recognizes concerning this simply. " So it probably is certain. As for their ones becoming fast as for wonder it is not. But, as for them shipment is late as the compensation. If the rough road (new architecture), the motorcycle which is agreeable to that (process technology) with we should go "

0.13 Mu M process this much the hurdle was high, it was difficult to expect in advance for 0.15 mu m generations to continue to here. 0.13 If mu m has improved yield rate favorably, as for the NV30 stuffing the gap, the wax which can surpass with efficiency. If it cannot put out the NV30 within year, the ATI wins betting, the NVIDIA means to be defeated. However, true efficiency race/lace starts, when both corporations line up with 0.13 mu m. And, the and enviroment for software development (the + compiler) fight becomes extremer.

The ATI does not have the plan which that way moves current RADEON 9700 to 0.13 mu m process. " 0.13 mu m, the following generation, probably become the following revision. If the next year first half you think " it becomes possible, the K Y Ho person says. The David E Orton of the ATI the president and the COO, the movement to 0.13 mu m calls " the next generation, per R350 (the kind of)" is. You can agree upon this logically. If in other words, from 0.15 mu m the physical design changes completely with the movement to 0.13 mu m, redoing from the logic design, because the one which also the itself expands, can save the development resource. In other words, the R350, with 0.13 mu m process, developing also architecture into the next year first half, is seen it appears. As for the ATI " as for all principal standards, because it supports (the Ho person) with you say in appropriate time ", as for the R350 as for the base the DirectX 9.1 (next year first half release schedule) you probably have become the generation. " - impress.co.jp (translated with babelfish

This article also talks about the R350 being a .13 micron chip.

Here is a copy of ATI's road map (If it is correct)

http://pc.watch.impress.co.jp/docs/2002/0909/kaiati.jpg

Notice how the more cost effective R350 will push high end performance into the main stream making way for the R400 to be the high end part.

[StealthHawk]

i thought it was agreed that R350 was .15?

[Gar]

Posted by CMKRNL at Beyond3d

In the past few weeks the following have taped out:

NV31 in 0.13.
R350 in 0.15.
RV350. I'm fairly certain 0.13, but not 100% positive.

I know this article is from the head guys at ATI, but I really doubt the flagship R350 is .13. CMKRNL hasn't missed on one prediction at beyond3d. He has predicted all of the Nvidia/ati tapeouts, delays, specs etc. When he doesn' know somehting he doesn't post. People say he works for TSMC and has the info way before anyone else. When saying the R350 is at .13 i think they aren't correct since it is somewhat ambiguous. There are 2 versions one called the RV350 and the R350. RV350 is at .13 to lower cost in production and will be the mainstream chip (also used as a mobile chip). If indeed they did tape out the R350 at .13 I really don't think its going to be out anytime soon.

[The One]

i wonder how the ati card will do in performance with a 0.13 micron chip and if they are going to have to use 12 layers as well ,and if they are goin to experience heat problems like nvidia has ?:rolleyes:


whats everyones opinion????

[StealthHawk]

R350 is .15

RV350 is .13

R350 is going to be the new high end product.

[Gar]

Originally posted by The One
i wonder how the ati card will do in performance with a 0.13 micron chip and if they are going to have to use 12 layers as well ,and if they are goin to experience heat problems like nvidia has ?:rolleyes:


whats everyones opinion????

Well considering the maturity of .15 right now and the fact that people are reaching 400mhz with standard cooling I highly doubt they will need a dustbuster to cool it. Also, they will probably use the same PCB as the 9700 since it is using the same traces (not completely, but very close to).

[PreservedSwine]

Originally posted by The One
i wonder how the ati card will do in performance with a 0.13 micron chip and if they are going to have to use 12 layers as well ,and if they are goin to experience heat problems like nvidia has ?:rolleyes:


whats everyones opinion????

ATI's approach to .13 is vastly different than Nvidia's.

Rather than make the trasition to .13 on a top of the line graphics card, they're doing it with their budget solution, the RV350. The RV350 will probably perform *near* R9700 levels, but I'm thinking/guessing it will be slightly slower, akin to the difference between an R9000 to an R8500.

Making a much less compicated .13 chip seems a great way to get their feet wet in the new process, as the card taped out months ago. I think it's all ready to market, but ATI is waiting until R350 is ready before an announcement is made.

Making thier first .13 card on their budget solution is a smart move on several fronts....

Top of the line performance is not expected, the magnifying glass reviewers us on top of the line cards will not be nearly as intense.

Also, a much less comlicated design makes the move much more simple, IMO. This is, after all, basically an R300 core with some features, and probably some pipes, removed, and the core speed increase should help make up the difference in lost performance.

I doubt the RV350 will require a 12 layer PCB like the NV30, it will be a much less complicated, less power hungry, cooler design...it will also be a bit slower. It's not aimed at NV30 performance.
All in all, a very smart way to bring this core to OEM's and Budget solutions:)

[Gar]

I am thinking ATI won't release the RV350 until the R400 is ready, much like how the 9000Pro was released during the R300 release. It will provide a fast, cheap and easy to produce solution once the 9700pro has started to dim out of the lime-light 1 year after its release. It will most likely be the bargin yet good performing, feature rich mainstream chip as an alternative to the R400 (much like the 9000pro was to the R300).

[netape]

I heard RV350 will replace R9500 (non pro), so the performance should be R9500 < RV350 < R9500 Pro. I'm not sure tho :)

[PreservedSwine]

Originally posted by Gar
I am thinking ATI won't release the RV350 until the R400 is ready, much like how the 9000Pro was released during the R300 release....

Close, the R350 and RV350 will be announced together...the R400 is still a long way away (6-10 months, IMO)
Rumors have it the RV350 is already finished..they're just waiting on the R350
:)

[Gar]

They were announced together but does not mean they will both be released together? If you check over at beyond3d, especially this post (http://www.beyond3d.com/forum/viewtopic.php?t=3299) many people feel that the RV350 is going to be released similarly to the way the RV250 was released. The 9000pro was delayed quite a bit until the R300 was released and was used as a bargin chip ($100.00) for the mainstream. I have a feeling, along with many others, that the RV350 is going to be used the same way. It isn't a replacement for the 9500pro or the 9700pro for that matter. It will merely be a budget card that offers a lot of the features that the R300 provides, expcept not as many pipelines or even the 256 bus, although nothing has been confirmed for its specifications. This is all speculation, infact ATI could use it to replace the 9700 using a 4 pipe setup with a 500mhz clock to equilibrate the 9700pros speed. Who knows...

Here are some quotes to support my theory:


Posted by MuFu
So I'm assuming RV350 is a mainstream counterpart to R400 (not a Radeon 9500 replacement, as such) and we may see a staggered launch a la RV250/R300 next year - remember that they "sat on" RV250 for ages until the R300 launch (it taped out many months prior).

Posted by DadUM
Given past numbering, R(V)350 should be a lower end R300 design cleaned for "mainstream" users.

Of course there are arguments for the otherside of the issue, but I think its notable that we use the prior RV lines to base what ATI will do with the current RV350.

[PreservedSwine]

Originally posted by Gar
They were announced together but does not mean they will both be released together? If you check over at beyond3d, especially this post (http://www.beyond3d.com/forum/viewtopic.php?t=3299) many people feel that the RV350 is going to be released similarly to the way the RV250 was released. The 9000pro was delayed quite a bit until the R300 was released and was used as a bargin chip ($100.00) for the mainstream. I have a feeling, along with many others, that the RV350 is going to be used the same way. It isn't a replacement for the 9500pro or the 9700pro for that matter. It will merely be a budget card that offers a lot of the features that the R300 provides, expcept not as many pipelines or even the 256 bus, although nothing has been confirmed for its specifications. This is all speculation, infact ATI could use it to replace the 9700 using a 4 pipe setup with a 500mhz clock to equilibrate the 9700pros speed. Who knows...

Here are some quotes to support my theory:



Of course there are arguments for the otherside of the issue, but I think its notable that we use the prior RV lines to base what ATI will do with the current RV350.

Thanks for the great link. Some *very* insightfull postings in there. You've just totally clouded my judgment:D