My guess is high end will have 3 quad's(12 pipes); mid level will have 2 quad's(8 pipes); and low end will have 1 quads(4 pipes).
ASUS - ATi - Roadmap
- Thread starter mboeller
- Start date
Fred da Roza
Newcomer
Joe DeFuria said:
Unfortunately ATI doesn't get a DX9 part into the value segment until Q2 2004. My impression is the 5200 is killing them there. I guess ATI expects the 9100 IGP to gain them maket share.
A IGP 5200 should do quite well. Remember, the compared fx5200 was the bandwidth-challenged part. So just as with the 9200se vs igp9100, the igp 5200 would beat the 64bit fx5200. The transistor count of the fx5200 is probably too high for an igp, but maybe nvidia this time unlike with the nforce2 (full mx440) doesn't use the full core but strips it. Maybe they follow ATI's lead and kill the vertex shader (which is slow anyways), I'm not sure if it would easily possible to sacrifice half the "pixel pipes" (whatever the number actually is) without changing the architecture too much. Transistor count would still be higher than for the 9100igp (9200 vs fx5200 is ~36million vs ~45million) of course. But if ATI is only doing P4/P-M chipsets and NVIDIA only Athlon chipsets, they don't compete directly and so it probably wouldn't matter much if it costs 2-3$ more to produce.reever said:
Joe DeFuria
Legend
RussSchultz said:
Given the hints that Dave is dropping, I'm inclined to agree.
There is only one potential issue I have with that line-up, and that's if the 4 pipe chip can be made cheaply enough to have decent margins at the low-end.
IIRC, the FX 5200 and Radeon 9200 are roughly 35 million transistors on 0.15u. (Someone correct me if I'm wrong.) Details on the RV350 trasnsitor counts are sketchy, but I beilieve it is nearly double, at 65-70 million? So I have my doubts that RV370 could in effect be "only" a respin of the current RV360, and still be cheap enough for the value market.
So if the RV370 is a 4 pipe chip, I do anticipate (as discussed earlier) a few things chopped off relative to the RV350/RV360. (Maybe HyperZ).
Then again, it's also understood that UMC charges less per wafer than TSMC...so perhaps if RV370 is being fabbed exclusively at UMC...that might be cheap enough to keep the entire functionality of RV350 in tact.
It's also possible that ATI might be taking the step to "ignore" the extreme low end "discrete" products, and instead attack that from underneath with the integrated RS400 chipset. That would mean that ATI is willing to make a slightly more expensive chip for the "value" segment, because it won't have to scale down quite as low as low end discrete chips traditionally would.
So the complete ATI line-up for 1H '04 would look something like this:
1) Low-end Value: RS400 (2 pipe DX9)
2) High-end Value: RV370 on 64 bit bus. (4 pipe DX9)
3) Low Mainstream: RV370 on 128 bit bus (4 pipe DX9)
4) High Mainstream: RV380 on 128 bit bus (8 pipe DX9)
5) Low Performance: RV380 on 256 bit bus (8 pipe DX9)
6) High Performance: R420 on 256 bit bus (12 pipe DX9, PS 3.0)
7) Enthusiast: Same as number 6, but higher clocks.
Fred da Roza
Newcomer
Joe DeFuria said:
I believe the 5200 has ~50 million.
I hope you are wrong about this. A ~70 million transistor chip with a 64 bit bus seems like a waste. I really hope that ATI doesn't follow the nVidia model in the low end.
Joe DeFuria
Legend
Fred da Roza said:
Transistor count to bandwidth ratio doesn't mean anything. It's pixels / sec per bandwidth that matters.
And with the move to pixel shaders, higher pixel rate / bandwidth ratios actually start to make more sense...not less. (assuming shader throughput is tied to pixel pipelines).
Was the Radeon 9500 Pro a "waste?" That is still TODAY "the card to beat" in terms of price / performance. A 4 pipe 64 bit card has the same pixel / bandwidth ratio as the 8 pipe / 128 bit Radeon 9500 Pro.
In the low end, it's all about meeting a fixed cost first and foremost. This means less transistors and less bandwidth than the higher end parts.
A 4 pipe DX9 card on a 64 bit bus will likely perform similarly in DX9 titles (without AA) vs. a 4 pipe board on a 128 bit bus.
Fred da Roza
Newcomer
Joe DeFuria said:
The 9500 Pro was not a money maker. It was a stop gap measure, hence the 9600. And it did suffer performance deficiencies over the 9700.
Joe DeFuria
Legend
Fred da Roza said:
Agreed. That doesn't mean such a configuration can't be a money maker if it's priced right. (9500 was 8 pipes on 0.15u. We're talking about a 4 pipe part on 0.13u process.)
Of course it did. Most significantly AA performance suffered (as I said). The 9500 Non-Pro suffered significantly worse performance deficiencies over the 9700 and 9500 Pro.
OEMs are selling 64 bit 9600 cards also.Fred da Roza said:
Joe DeFuria said:
Certainly great at the time, but easily beat price/performance-wise now by the 9800 non-Pro, which at this point is only ~$20 more, if that. Since the 9500 was discontinued it doesn't go down in price, while the 9800 non-pros have been steadily sinking due to all the myriad of new 9800's coming down.
And hey, if you get one with Samsung RAM you've got yourself a Pro right there!
(Soft-modding the 9500 was always great, too, which added many bonus points to it. Much love for the 9500.)Joe DeFuria said:
Actually, I don't think we'll 256 bits bus on a value/mainstream part just yet. Three reasons:
1. G-DDR 3 (?) around 600MHz wil offer the same bandwidth as the 9700 on a 128 bit bus.
2. unfortunately 64 bit bus isn't going anywhere
3. remember when we had new memory type (SDRAM->DDR SDRAM) bus widths remained the same for,what, 4 years? maybe 5
I just got wonder if there will be any slippage of the time line I gotta wonder why the XTs got made they really didn't need the preformace boost atleast not in the top end (9800 xt ) maybe they are getting much better yields I dunno.
Anyway I hope the R420 is as impressive as the 300 was if so I might even pay around 450 US for it. I guess we will have to see.
Hey is anyone working on PS 3.0 benchmarks?
Anyway I hope the R420 is as impressive as the 300 was if so I might even pay around 450 US for it. I guess we will have to see.
Hey is anyone working on PS 3.0 benchmarks?
About RS400..
It will come about 3 quarters later than RS300
RS300 is 2-pipe RV250/RV280 at 300 MHz, and is doing quite well now.
RV360 can be easily clocked to 500 MHz.
so RS400 could be a 2-pipeline RV350/360 clocked to some 450 MHz.
(and maybe without VS)
It would offer 1.5 x theoretical fillrate advantage over RS300,
but bigger performance in real situations due the architechtural differences between th R2xx and R3xx series, even if memory bandwith only increases 33 % from DDR400 -> DDR2-533.
That kind of performance increase would be kind of enough for 3 quarters,
I see no big need for more pipelines on this market segment.
It will come about 3 quarters later than RS300
RS300 is 2-pipe RV250/RV280 at 300 MHz, and is doing quite well now.
RV360 can be easily clocked to 500 MHz.
so RS400 could be a 2-pipeline RV350/360 clocked to some 450 MHz.
(and maybe without VS)
It would offer 1.5 x theoretical fillrate advantage over RS300,
but bigger performance in real situations due the architechtural differences between th R2xx and R3xx series, even if memory bandwith only increases 33 % from DDR400 -> DDR2-533.
That kind of performance increase would be kind of enough for 3 quarters,
I see no big need for more pipelines on this market segment.
Joe DeFuria said:
The NV34 has AFAIK 47 million transistors. A 55-65 million transistor RV370 would not be that bad assuming it is produced at a low-cost 0.13 process.
Yes, the price of 0.13 must obviously have dropped quite a bit since the introduction. With integrated solutions for the low-end value segment it might even make sense just to add PCI-express to the RV350/360.
The RV370 (or similar) will in time trickle down to the extreme low-end. But until that happen Ati will have to do with the RV280 and integrated solutions. I am pretty sure that no matter what we will see RV370 based products in the sub $100 range the question is if we will see sub $75 RV370s.
I pretty much agree but I think there H1 lineup will include some of the old R(V)3x0 cards for all the AGP users out there.
hkultala said:
<dream-mode on>
I would like to see an integrated 3D-chipset with the following features :
<dream-mode off>
The QBM-support could be possible cause VIA will introduce an chipset with 128bit interface and QBM support next year. At the moment QBM works only up to 166MHz but maybe they have already overcome this limitation (have not checked).
The transistor-count for an 4pipline RV360-style integrated 3D chip would be too high, but well I said "dream-mode".
Joe DeFuria
Legend
vb said:
I agree. That's why my "lowest" 256 bit bus card is placed in the "low-end performance" market...not the mainstream market.
I'm talking about price range here. You might be positioning the RV380 as a "mainstream chip", but it will be sold (speculation) in the "performance" market when paired with a 256 bit bus.
Perhaps I should apply some prices (MSRP) to my 1H 04 speculative line-up....it may make things clearer.
1) <$50: Low-end Value: RS400 (2 pipe DX9)
2) <$75: High-end Value: RV370 on 64 bit bus. (4 pipe DX9)
3) <$100: Low Mainstream: RV370 on 128 bit bus (4 pipe DX9)
4) <$200: High Mainstream: RV380 on 128 bit bus (8 pipe DX9)
5) <$300: Low Performance: RV380 on 256 bit bus (8 pipe DX9)
6) <$400: High Performance: R420 on 256 bit bus (12 pipe DX9, PS 3.0)
7) <$500: Enthusiast: Same as number 6, but higher clocks.
Joe DeFuria
Legend
Tim said:
I hope that's the case. A "true" RV350/360 being sold in the Value discrete market would be great.
I'm gussing that ATI will definitely want a $75 MSRP SKU, though they'll ignore the $50 MSRP discrete market and go after that with the RS400. So yes, the question is, would a RV370, paired with 64 bit DDR, be cheap enough for $75 MSRP?
Makes sense.