ASUS - ATi - Roadmap

[mboeller]

found @ 3DCenter.de :

It seems the R420 will be available exactly one year after the R350 in march/april 2004. Also it seems the RV280 will be replaced as an lowend-solution only in Q1/2004 with the DX9-part RV370. I wonder were ATi hides the AGP-version, or will they have no lowend-DX9 solution for the AGP-port at all?

 

[Joe DeFuria]

Interesting...thanks for the link.

There are lots of options open to ATI in Q1, so it's hard to speculate on what each of these parts might be. Time to put on the speculation hat...

1) R420: the most common rumor is that it is R-300 based, (with possibly VS/PS 3.0 bolted on), 12 pipeline chip. 0.13u....may or may not be low-k. Assumption is also 256 bit memory interface. I'm guessing would probably come in two flavors...one (Pro) with bleeding edge GDR-3 @ 700+ Mhz actual, and one (non-Pro) with more "normal" GDR-3 at 500-600 MHz.

2) RV380: Could just be just a RV360, but with PCI Express support (and likely GDRR-3 support IMO), possibly with a bump up in core clock speed. One version with GDDR3 @ 500 Mhz is a real possibility IMO. I'm thinking the RV380 has a chance at finally bringing radeon 9700 performance to the <$200 MSRP segment. Other (though less likely, IMO) possibilities include increasing the pipeline count to 6, and/or increasing memory interface to 256 bits.

3) RV370: My guess: 2x1 (or maybe 2x2) version of the RV360. Option to use either 64 or 128 bit DDR memory. Built on 0.13u (possibly UMC), and running at 500+Mhz core speed. Memory speeds / width configuration will vary wildly depedning on exact price target for the OEM.

(edited for a couple typos.)

 

[Natoma]

re: #2

That's if the 9600XT doesn't beat it to that title first.

 

[Joe DeFuria]

re: #2

That's if the 9600XT doesn't beat it to that title first.

I just don't see that happening...the 9600XT just doesn't have the raw bandwidth....by a long shot. (Only about 1/2 the bandwidth of the 9700 Pro.) The only reason why the rumor is floating around that the 9600XT is a 9700 Pro killer, is because of one statement by Anandtech...which I'm guessing is simply not correct.

 

[Natoma]

Wait. Are you talking about the 9700 Pro or the 9700 Non-Pro? In your first post you stated radeon 9700 performance and in your reply you stated radeon 9700 pro performance. If you meant pro in your first post then I'd be inclined to agree. But 9700 performance, which is the caveat I was replying to, just might be within reach by the 9600XT.

 

[Joe DeFuria]

Either the 9700 or 9700 Pro.

The 9600 XT only has 55% of the memory bandwidth of the 9700 non-pro.

While the 9600 XT can be very comparable in terms of fill-rate and shader performance (only 10% short of 9700 non-pro fill-rate), the memory bandwidth limitations will prevent the 9600 Xt from competing in high-resolution, high AA situations.

In other words...it's more or less the equal of the 9500 Pro. As good as the 9500 Pro is...it's not in the same category as the 9700 non-pro.

To be clear, I expect the 9600 XT to be a clear price/performance leader, and one killer card....just don't expect it to perform "across the board" as good as any of its 256 bit cousins.

 

[Tim]

Interesting...thanks for the link.

There are lots of options open to ATI in Q1, so it's hard to speculate on what each of these parts might be. Time to put on the speculation hat...

1) R420: the most common rumor is that it is R-300 based, (with possibly VS/PS 3.0 bolted on), 12 pipeline chip. 0.13u....may or may not be low-k. Assumption is also 256 bit memory interface. I'm guessing would probably come in two flavors...one (Pro) with bleeding edge GDR-3 @ 700+ Mhz actual, and one (non-Pro) with more "normal" GDR-3 at 500-600 MHz.

If the R420 is 12 pipes at 500MHz 664MHz GDR-3 would provide the bandwith/fillrate ration as the 9800XT. I would be suprised if Ati increased bandtwith more than fillrate.

2) RV380: Could just be just a RV360, but with PCI Express support (and likely GDRR-3 support IMO), possibly with a bump up in core clock speed. One version with GDDR3 @ 500 Mhz is a real possibility IMO. I'm thinking the RV380 has a chance at finally bringing radeon 9700 performance to the <$200 MSRP segment.

I agree.

Other (though less likely, IMO) possibilities include increasing the pipeline count to 6, and/or increasing memory interface to 256 bits.

I think that 6 pipeline design is extremely unlikely the R3x0 uses quad-blocks designing a 6 pipe design would require a bit too much work and only have a few less transistors than the 8 pipe chips.

3) RV370: My guess: 2x1 (or maybe 2x2) version of the RV360. Option to use either 64 or 128 bit DDR memory. Built on 0.13u (possibly UMC), and running at 500+Mhz core speed. Memory speeds / width configuration will vary wildly depedning on exact price target for the OEM.

I think there is a chance that the RV370 is still a 4 pipeline design I believe that can reduce the 75 million transistors of the RV350 quite a bit without reducing the number of pipelines.

 

[Joe DeFuria]

If the R420 is 12 pipes at 500MHz 664MHz GDR-3 would provide the bandwith/fillrate ration as the 9800XT. I would be suprised if Ati increased bandtwith more than fillrate.

Agreed. Target memory speed probable depends on a couple things:

1) Power consumption
2) Core clock speed.

500 Mhz is a good guess for R420 core clock, but it could very well be closer to 600 Mhz...at least for the "pro" version.

(Recall that ATI improved both transistor count and clock speed going from the R-200 to R300...and that was staying on the same 0.15 process.)

I think that 6 pipeline design is extremely unlikely the R3x0 uses quad-blocks designing a 6 pipe design would require a bit too much work and only have a few less transistors than the 8 pipe chips.

Agree here too...just thought I'd throw the 6 pipeline thing out there.

I think there is a chance that the RV370 is still a 4 pipeline design I believe that can reduce the 75 million transistors of the RV350 quite a bit without reducing the number of pipelines.

Yes...they could get rid of Hyper-Z more or less completely, which could save a chunk of transistors. Would be interesting to see performance profiles of something like a 2x2 "with" HyperZ, vs. a 4x1 without it.

With ATI's experience of moving vertex shaders to the CPU (see IGP 9100), it's also possible that ATI could cut out Vertex shaders from the RV370 too, and relegate that task to the CPU.

 

[mboeller]

Again found @ 3DCenter.de :

If this chart is correct then it seems the RS300 has only a short life span. If the RS400 has an improved memory controller like the RS350 (according to Dave Baumann) then this could be an really interesting product. Will this be the first DX9 chipset, or will Nvidia be first? Would be nice if both come out with an DX9 chipset at the same time. I had not expected that ATi will have an DX9-chipset so early. I expected something like this only at the end of 2004 or even at 2005.

 

[Joe DeFuria]

If this chart is correct then it seems the RS300 has only a short life span.

Well, the RS300 looks to have about a 1 year life span. I suspect RS400 is being set for a "summer '04" release, much like the RS300 was summer '03. Resulting motherboards might not appear until very late in the year though.

(edit...looking at the graph again, you're right...clearly, it's indicating a launch of RS400 at the same time as the rest of the traditional spring products...that is in fact a bit earlier than expected.)

If the RS400 has an improved memory controller like the RS350 (according to Dave Baumann) then this could be an really interesting product. Will this be the first DX9 chipset, or will Nvidia be first?

I'm sure one of the driving factors for DX9 integrated chipsets is Longhorn. I'm sure both ATI and nVidia want motherboards with DX9 chipsets shipping by the time Longhorn hits the market, presumably sometime in early '05?

Would be nice if both come out with an DX9 chipset at the same time. I had not expected that ATi will have an DX9-chipset so early. I expected something like this only at the end of 2004 or even at 2005.

Well again, I expect that we won't actuall see RS400 products until near the end of 2004. ATI had already communicated that a DX9 chipset was in the works for 2004, so it's not too much of a surprise to see it on a roadmap.

What is interesting, is that this is the first "real evidence" that I know of that shows R420 as being a Shader 3.0 part. Even though that was a given assumption by most, I still had shader 2.0 as a possibility for R420.

 

[Joe DeFuria]

Now that I think about it some bit....

ATI might be trying to kill two birds with one stone with ther RS400. In addition to being DX9, it might also have PCI express support. (Completely unsubstnatiated speculation on my part.) That would help explain ATI's push to get it out at the same time when PCI express makes it debut.

 

[Dave Baumann]

Note.

The arrows on the ASUS slide are kind of misleading as to what goes to what...

 

[Joe DeFuria]

Note.

The arrows on the ASUS slide are kind of misleading as to what goes to what...

So, um....in what way? I'm going to go out on a limb and say that the R360 to R420 arrows are correct.

How, would you draw the arrows from RV280 (9200) and RV360? Hypothetically speaking, of course. 8) Don't worry, I won't quote you on that!

 

[hoom]

Will this be the first DX9 chipset, or will Nvidia be first?

Given that nVidia has failed to get a DX9 desktop graphics card out yet, I can't really see them having a DX9 chipset

Maybe they'll downgrade the DX9 effects to DX7 with the drivers?

 

[reever]

I think Nvidia is coming out with a 5200 IGP, but seeing as how the 9100Igp is already on par with the standalone 5200 , i dont thinkit would be very good

 

[Dave Baumann]

So, um....in what way? I'm going to go out on a limb and say that the R360 to R420 arrows are correct.

Well, no actually.

What philosophy does ATI use...? The trickle down approach.

So, where do you think the current R360 spec will go when R420/3 is around...?

 

[Joe DeFuria]

Well, no actually.

What philosophy does ATI use...? The trickle down approach.

So, where do you think the current R360 spec will go when R420/3 is around...?

Hmmm...

Well, in recent history, the prior "spec" was just replaced by the new spec. The spec didn't trickle down...the technology did.

Having said that...

It is conceivable that ATI could do something like the following:

1) R420 replaces the R350/R360 in the $300-$500 MSRP products range.
2) RV380 = exactly the R350, (8 Shader 2.0 pipes) but on 0.13u. Migrates to $150-$300 MSRP price range.
3) RV370 = $100 and under price bracket.

Am I getting warmer?

 

[hoom]

They are rough performance graphs for the different market segments.

Not tracking architecture.

 

[Dave Baumann]

Well, in recent history, the prior "spec" was just replaced by the new spec. The spec didn't trickle down...the technology did.

Actually Joe, there are two precidents. Yes, the spec changed from R200 to RV250, but did it from R100 to RV200? No, in fact performance went up. What is the difference between these two precidents? R100 to RV200 was a die shrink, R200 to RV250 was the same process. This time, however, we are transitioning to 130nm...

 

[Joe DeFuria]

Actually Joe, there are two precidents. Yes, the spec changed from R200 to RV250, but did it from R100 to RV200? No, in fact performance went up. What is the difference between these two precidents? R100 to RV200 was a die shrink, R200 to RV250 was the same process. This time, however, we are transitioning to 130nm...

Indeed.

So it looks like the line-up I presented in my previous post may be close to the mark.

ATI could the differentiate the low and high end RV380 parts by varying the bus from 128 to 256 bits.