ATI to launch Radeon 9500 and 9700 on the 18th

[Sabastian]

http://www.theinquirer.net/?article=5779

ATI to launch Radeon 9500 and 9700 on the 18th

Non-PRO's tip up

By Fuad Abazovic: Friday 11 October 2002, 10:22


DOCUMENTS SEEN BY THE INQUIRER show ATI's upcoming offerings are still scheduled to appear on 18th -- next Friday -- if all remains on target.
The Radeon 9500 and 9700 cards where announced back in July when ATI presented the Radeon 9700 PRO, its fastest offering. But now we have a few details that might be interesting to you.

Both cards will run at 275MHz, while the memory will run at 550MHz the same clock speed as seen on last year's ATI's fastest 8500s. The number of pipelines will be 8 for the Radeon 9700 and only four for the 9500, contrary to many reports. There will be a faster version of the 9500 called PRO that should have 8 pipelines but this card will not be introduced with this bunch of non-pros, at least not for some time.

The main difference between the cards will be the memory interface. The Radeon 9700 will use a 256-Bit interface and is planned to compete with Nvidia's Geforce 4 TI 4600. The card will feature 128MB of DDR memory working at 550MHz, an AGP 8X interface, Dual 400MHz Ramdac and all the things that you already saw with Radeon 9700 PRO. But you can expect that this card will be a bit slower than the fastest 9700 PRO. Of course, the card will have full support for DirectX 9.0

ATI is planning to introduce this card at 299USD a price that is sure put some extra pressure on rival Nvidia.

The second card, the Radeon 9500 will cut down the pipeline number from 8 to 4 which will dramatically affect the pricing of these cards. These cards will work at 275MHz for the VPU and 550MHz for memory and will have 64MB or Ram on board, except in some special cards for "special" customers who will be able to get these cards with 128MB of memory. You may assume that some big OEMs will go for these cards.

The Radeon 9500 will use a 128-bit memory interface and synchronous memory and the GPU speed is something that ATI prefers when it clocks its cards. These cards will have full DirectX 9 compatibility with Dual 400MHz Ramdac, Smartshader 2.0, Smoothvision 2.0, Videoshader, Fullstream, Truform 2.0, HyperZ III -- meaning it will feature all the features of 9700 and 9700 Pro cards.

The amazing part is the price, since ATI aims at 179 USD for this card, placing it up against the very successful TI 4200. ATI's 9500 will be first affordable sub-200 USD card with full DirectX 9 support. That might be attractive to many people.

The memory used on this cards will come from Infenion or Hynix depending on qualification.

Our guess is that next week will be very interesting and we would like to see these babies running in our test beds. µ

 

[Joe DeFuria]

That actually makes a lot more sense to me than the 9500 Pro (8 pipe, 128 bit interface) rumors. I hope it's true.

Once the products hit the pipeline in volume, we should see the 9700's selling at around $250 U.S. and the 9500's around $150.

That would make a very good line-up for ATI IMO. The products (From the 9000 up to the 9700 Pro) are differentiated enough to make a clear and justified separation in price-points.

 

[Sabastian]

Yeah. I agree. What I am wondering is how the 9500 series cards will actually stack up against the ti4600 cards. I know that they should outpace the ti4200s but if ATi is able to take nvidias high end part with their midstream 9500 cards then they certainly are playing a smart game. Further AFAIK nvidia has no intentions of releaseing any other .13um part this year. This really equates to nvidia having no answer for the midstream till next year. Granted the NV30 will prolly beat the R300 in performance to some degree ... but by how much and will ATI be able to manage some sort of refresh in the near term to answer the NV30 in terms of performance?

 

[Dave Baumann]

If these are all from the same R300 core what types of margins must they be running??

 

[Joe DeFuria]

Yeah. I agree. What I am wondering is how the 9500 series cards will actually stack up against the ti4600 cards. I know that they should outpace the ti4200s

Actually, keeping in mind that the 9500 is a 4x1 architecture (assumption), compared to the GeForceTi's 4x2, I'm guessing that in non AA and non aniso situations, I would not be surprised to see the 4200 even beat the 9500 in certain situations.

However, turn on AA and / or Ansio, and the 9500 should easily beat the 4200. That, plus being "future proof" (cough) with DX9 support should make the 9500 the new "best bang for buck" card on the planet, relieving the 4200 of that title.

I'm guessing the 9500 should also beat even the 4600 in most cases with AA and especially aniso. However, I expect the 4600 to beat the 9500 routinely in non AA / non Aniso benchmarks. The 9500 of course has a good price advantage compared to the 4600.

If these are all from the same R300 core what types of margins must they be running??

Well, if it is the R300 core (with just pipelines disabled), then I assume the chips differ at least in the sense that the 9500 (128 bit) does not have the expensive packaging that the 9700 does, which is probably a significant cost reduction. (I would expect the 9500 reference boards to look more like the 8500, than the 9700).

That being said, I would also not be surprised if we find out that it was in fact a different chip though, perhaps around 80 million transistors or so. Will be interesting to find out.

Also, note that ATI will cut some costs by shipping most 9500s with "only" 64 MB of memory.

The question on my mind is: what margins will nVidia be making on their chips once they have to be lowered in selling price to compete with ATI's offerings?

 

[LeStoffer]

Actaully, keeping in mind that the 9500 is a 4x1 architecture, compared to the GeForceTi's 4x2, I'm guessing that in non AA and non aniso situations, I would not be surprised to see the 4200 even beat the 9500 in certain situations.

Ack, I think people give to much credit to that extra TMU unit (we see the same discussion with R9700 and NV30).

I truely have my doubt about the efficience of multiple TMU today with a given bandwidth.

My best example is the GeForce DDR against the GeForce 2 GTS. The design moved from 4x1 to 4x2 (and increased Texels/Second 480 from Million to 1600 Million). Granted the memory bandwidth increase was nearly non-existant (from 4.8GB/Sec to 5.3GB/Sec) but this makes for an interesting showdown between a 1 TMU and 2 TMU design:

http://www.nvnews.net/previews/geforce2_gts/introduction.shtml

Is there any real world increase because of the second TMU? It's hard to find, isn't it?

 

[Joe DeFuria]

Ack, I think people give to much credit to that extra TMU unit (we see the same discussion with R9700 and NV30).

To be clear, I'm really not giving all that much credit for the extra TMU. But again, in certain situations, it is a benefit. (See Radeon 8500 vs. Radeon 9000).

 

[Sabastian]

quote="DaveBaumann"]If these are all from the same R300 core what types of margins must they be running??[/quote]

That is a good question Dave. The cost of producing the chips though shouldn't be too bad as they are being created at the tried and true .15um process and that recently TSMC has lowered the cost of producing chips on this process substantially. For the fully functional Radeon 9700pro core ATI would prolly have to offset the price of these chips as the 256bit interface makes for an expensive PCB. But on the other hand the lower clocked and cut back R300 core running on a 128bit interface is a cheaper card for ATIs partners to produce. So I suspect that the margins are quite similar no matter what the third parties sell. So in effect ATi is prolly getting comparable margins on the high end as they will get on the low end whatever they may be. This would explain how ATIs third parties have been able to give competitive prices on the Radeon 9700pro in comparison with the TI4600. With this in mind the greatest indicator as to how much money ATi is making would almost solely depend on what volume of these chips are sold, not so much which ones. A mid range solution would give ATi parallel margins as a high end solution.(This is all speculation on my behalf so if anyone has some other sort of conclusion or twist please by all means intervene.) So ATi saves on R&D and also doesn't have to struggle with different designs and is able to move on to the next product considerably faster.

 

[DadUM]

So let me get this straight. Ati will now have the 9000, 9500, 9500 PRO, 9700, 9700 PRO plus all the AIW versions. Further, from what I gather, there are 4 different silicon chip designs to cover these (only the 9700 and 9700 PRO have the same architecture). I officially give up trying to figure out how graphics companies make money. I have no idea how they pay for the development of 4 concurrent architectures, write high quality drivers for 4 architectures, market 5 products within a <$300 market, and repeat this every 6-9 months.

 

[Bigus Dickus]

I think everyone had assumed the 9500 would be 4 x 1 with a 128 bit memory controller. It's nice however to have some rumors of clockspeed now.

The 9500pro 8 x 1 with 128 bit memory sounds like it's still firmly in the highly speculative rumor category. This could be an interesting product, and I am more curious than anything as to what its performance would be.

What I'm most excited about though with this Inq. speculation is the possibility of a Ti4200esque R300. That is, same core, same memory controller, just lowered clock speeds. It was assumed at first that there would be an LE or OEM version of the 9700 running at 300/300 or 275/275, but none of those surfaced. Thus far, it's been all 9700pro's. If they leave off the floppy power connection, that should be easy enough to rectify, and the core will probably clock similarly to the pro version. Very interesting indeed, as I had thought there would be no reduced clock 256 bit version, only the supposed 8 pipe 128 bit 9500pro jobby.

 

[Joe DeFuria]

DadUM,
No, according to this rumor, there is no 9500 Pro (at least not yet).

There is the following:

9000: DX 8.1, Sub $100
9500: DX 9, mainstream performance, $179
9700: DX 9, enthusiast performance, $299
9700 Pro: DX9, top-of-the-line performance, $399.

That's a reasonable product line-up to me. All-in-wonder products are a different segment alltogether.

That's really only 2 architectures for drivers and development, not 4. And one of those, the 9000, draws heavily from the past architecture, the 8500.

Also, ATI repeats this every 9-12 months. Not 6-9.

I do agree that this is a torrid pace! However, it's not as exceedingly impossible as you make it out to be.

 

[Sabastian]

So let me get this straight. Ati will now have the 9000, 9500, 9500 PRO, 9700, 9700 PRO plus all the AIW versions. Further, from what I gather, there are 4 different silicon chip designs to cover these (only the 9700 and 9700 PRO have the same architecture). I officially give up trying to figure out how graphics companies make money. I have no idea how they pay for the development of 4 concurrent architectures, write high quality drivers for 4 architectures, market 5 products within a <$300 market, and repeat this every 6-9 months.

Hrm, I think you have missed something here. There are not 4 different architectures. ATi is finally doing what nvidia has been doing for the last couple of years. There is only one architecture being split 3 ways. The 9000 series is however a different architecture consisting of about 40 million transistors. This is based on the Radeon 8500 core. Further it is exactly the same sort of idea nvidia has been reproducing with its MX series only considerably better at this point.

 

[GetStuff]

<200$ DX9 card already? ATi is sure on a roll...

 

[DadUM]

Hrm, I think you have missed something here. There are not 4 different architectures. ATi is finally doing what nvidia has been doing for the last couple of years. There is only one architecture being split 3 ways. The 9000 series is however a different architecture consisting of about 40 million transistors. This is based on the Radeon 8500 core. Further it is exactly the same sort of idea nvidia has been reproducing with its MX series only considerably better at this point.

Architecture was a poor choice of words because that really does imply some extra stuff. Let's think in terms of silicon. For a 9500 (assuming the specs are correct), Ati had to order a different chip production from the foundry than for the 9700. The fact is, they still had to pay for that design, it's validation, it's production, and a whole new PCB to go with it. And since the 9000 part fills in a lower market line and the 9700 part fills the higher line, they won't order as many chips of that design reducing the "economy of scale" effect.

As for comparisons to Nvidia, my statement was I don't understand "how graphics COMPANIES make money." Nvidia has a lot of lines also, although they speed bin more to separate mainstream from enthusist, cutting down some of those costs.

 

[demalion]

You missed some info, DadUm....the 9500 is the R300 core, with a set of 4x1 pipelines crippled. There have been several snippets of info that lend support to it being the same chip design.

EDIT: the crippled set of 4x1 pipelines is assumed to be from chips that simply failed validation of one of the 4x1 pipelines.

 

[Joe DeFuria]

To be fair to DadUM...we certainly don't know that for sure. It may be a R300 with 4 pipes disabled, it may be a different chip.

Also, it seems highly unlikely to me (and I believe Russ had a discussion on this), that ATI would "test to see if 4 pipelines are bad, and use those bad chips for the 9500."

If it is an R300 chip, it's more likely that it's simply a speed binned R-300 that didn't validate at 325 Mhz, and has 4 of its pipelines disabled...regardless of whether or not "they work".

 

[duncan36]

As mature as .15 is it seems highly unlikely that this is the case.

 

[Clashman]

To be fair to DadUM...we certainly don't know that for sure. It may be a R300 with 4 pipes disabled, it may be a different chip.

Also, it seems highly unlikely to me (and I believe Russ had a discussion on this), that ATI would "test to see if 4 pipelines are bad, and use those bad chips for the 9500."

If it is an R300 chip, it's more likely that it's simply a speed binned R-300 that didn't validate at 325 Mhz, and has 4 of its pipelines disabled...regardless of whether or not "they work".

Makes you wonder if there'll be a hack to enable those other 4 pipes, similar to, I think it was, the 3dfx Velocity products.

 

[LeStoffer]

Ack, I think people give to much credit to that extra TMU unit (we see the same discussion with R9700 and NV30).

To be clear, I'm really not giving all that much credit for the extra TMU. But again, in certain situations, it is a benefit. (See Radeon 8500 vs. Radeon 9000).

Okay Joe, fair point there! I forgot that R 8500 and 9000 PRO are very close (minus the second TMU). But I think that it'll be fair to say that any advantage that a second TMU have will be all but gone with FSAA and AF enabled (especially on an ATI card).

 

[demalion]

I will mention that it seems to me the maturity of the .15 process is not the only factor, as even achieving this design at these clock speeds on .15 was a surprise to many.

I can point out that ATi has done very similar before for the initial Radeon launch. I've drawn a parallel to this before and I don't see why it isn't viable now. They had both lower clock speeds (than the fastest variant of the chips available) and features disabled, and this seems to fit perfectly with trying to maximizing returns from fabrication.

I can also seek to recall to memory the discussion here about the initial rumor regarding this before the 9700 launch about a design consideration that made each 4x1 set of pipelines redundant for a functioning core (though not redundant for performance) that seems to be supported by the general trend of info leading up to the upcoming 9500 launch.

Also, it seems highly unlikely to me (and I believe Russ had a discussion on this), that ATI would "test to see if 4 pipelines are bad, and use those bad chips for the 9500."

Why? I don't recall this being a supported conclusion in anything I've read here, perhaps I've missed it.

I will also point out that at Radeon launch, there were "crippled" chips that could have the crippled features turned back on and they would function properly...it seems to me there are all sorts of statistics that could be used to aid in this type of testing rather than "testing every single core on the wafer", and that this statistic modelling could easily be used, as it seems to have been for these early Radeon cores I mention, to determine which chips are destined for the "9500".

I'll provide guesses based on what seems likely to me based on my own experience on testing criteria (for something pretty completely unrelated) for the convenience of Russ, and others, to correct and/or add to:

location on a wafer

which "assembly line" was used to manufacture, and what past history is known about that assembly line

correlation (i.e., when this easy to test criteria is observed, it can indicate something else less easily measurable is likely to be true)

direct test sampling... perhaps random, perhaps methodical



It would seem simple to me to have a "lower criteria" based on these where the analysis of the likelihood of a chip successfully running using only 4 pipelines at a clock speed would be different than even running at the same clock speed with 8 pipelines, and that this difference in likelihood would be a valid reason to justify using the core(s) in the 9500. This even before considering the cost savings on the 128-bit bus design of the card and the additional factors of 256-bit versus 128-bit stress on a core factoring into what it would be profitable for ATi to do.