Planned volume production of RV350 by end of month

OK, Joe--definitely sounds interesting--and likely, too. As I mentioned to Antlers, using RV350 to actually replace something in the line rather than make another addition makes the picture much more clear. I'm pretty interested in knowing what ATI's going to do late winter-early spring...(as is everybody else, I'm sure.)

 

speculation mode on.

RV350 - a 0.13 version of R300 that's stripped down but has a faster clockrate. Probably 60M transistors. 4 pipelines, 128-bit bus. perhaps a boost via DDRII, but that memory might be saved for higher end chips.


R350 - a 0.15 refresh of the R300. totally superior to R300. a new chip. improvements, tweaks, additions above an beyond R300. probably still a 8x1 configuration. but with 256-bit bus AND DDR-II to ensure superiority against GeForce FX. Perhaps R350 will support Pixel and Vertex Shader 3.0 specification? if not, it should at least match GFFX, which does not meet 3.0 but goes beyond 2.0


R400 - a 0.13 next generation architecture from the same team that brought us R200 (Radeon 8500) - this monster will crush anything in Nvidia NV3X line, including the NV35. The R400 will have 12-16 rendering pipelines (probably 16) will probably be TWICE the complexity of R300. R400 should also use GDDR3 without a doubt. All the new technology & algorithims that ATI has been working on over the last year or so will be in R400. will have Vertex Shader and Pixel Shader 3.0 if not 4.0 - will it be DX10? maybe, maybe not, but certainly DX9.1 and beyond. will be available by Sept most likely.


of course, for 2004 there is probably going to be a R450 and without a doubt, the R500 which would be the next chip from the team that made Flipper and R300. Perhaps with R500, we will finally see some on-die embedded memory? if it's not already in R400.

 

Hmm megadrive .. while i know it's speculation .. I would think you might be rather a little off.

First thing about the RV350. In the reactorcritical article I posted earlier it mentions the RV350 would have performance close to the Radeon 9700 .. and seeing as the Radeon has 110m transistors I can't imagine the RV350 having only 60m. I would think it would be close to 100m transistors. Remember it's a 0.13 process which would indicate that they can put more transistors on this little puppy for less(expense). As for the other items .. who will know .. guess we'll see soon(or later if the article is right).

On the R350. While the R350 is a refresh of the R300, I can't see the R350 being a totally new chip. Improvements and tweaks "yes" but not a new chip.

Soon puppies soon!! (if everything goes according to ATI plans )

US

 

my speculation is

RV350 will have 4x1 or 8x1(I think it's impossible, because its price target is only sub 100 bulks ) and better bandwidth saving techology base on R350,besides cut some useless feature , put the core into 6x million transistors count, i think ati will use 32bitx4 design to closer R9700 performance wise and put higher DDR-I memory( at that time i think DDR Price will become more acceptable for pepole to use it)

What I am curious is NV31 and NV34, although they are cut down version of NV30, but i think both of them will have two scenario,(1) NV31 8x1 and NV34 4x1 (2) NV31 4X1 and NV34 2x1

 

I would like to see the RV350 with 4x1 @ up to 550MHz and an 256bit memory interface.

This would make it easy to replace the R300-chips with an RV350-chip and use the same card-layout as with the R300 cards. This would drive down the cost of the R9500 toward $100 (using an slow RV350 @ < 350MHz); The R9500Pro toward $150 and the R9700 toward $200. Alternatively the R9700-replacement could use 500MHz DDRII memory with an 128bit memory interface, but IMHO this would be very expensive. The R9700Pro is replaced with the new R350 card.

I don't think we will see RV350 card only later this year. Why should ATi produce mainstream cards using the expensive R300 core, when they could use the cheaper RV350 core instead (even if the performance is slower as with the R200 <-> RV250 )?

 

I wouldn't put a great deal of stock in that RC article.

 

so RV350 is even better? 8)

 

well, Unknown Soldier, R350 would have to be a "new" chip simply because there has to be a tape out, thus a new chip, but not a totally new architecture, which is saved for the R400 I do think that R350 will be as improved over R300 as NV25(GF4ti) was over NV20(GF3,ti200,ti500)

you might very well be right on RV350. it might indeed be all the performance of R300 and then some. same or similar transistor count, with faster clock, on 0.13 process. or it could be like RV250 was to R200, a slight drop in some areas, while more efficient in others.

no doubt that R400 will be a titan though. totally new chip with far greater complexity than either R300 or R350. should mop the floor with NV35

 

Nothing to see here, a moment of weakness - happens to everyone!

 

Alex - are you not heading for a fall here?

 

Sorry Dave, but it's just pathetic how uncostructive and offbase certain people's replies are around here!
You can delete my previous post if you'd like...

 

Unknown Soldier said:

Basically, it is my opinion that reactorcritical will turn out to be wrong. In other words, I have serious doubts that RV350 has performance that's comparable to the 9700.

With today's news that the RV280 is going to be produced at UMC at a cheaper price, this is a clear sign to me that ATI is paving the way move the Radeon 9000 down a notch to the sub-value market where the 7x000 currently is. (As my speculation yesterday indicated.) Which in turn, means that they are planning to bring some product into the $100 segment left vacated by the move of the RV-280 into a lower bracket.

So I would say at this point, there is no question that the RV-350 is indeed meant for the $100 price point.

Now, for the "close to 9700 performance" claim For that to be true, two things would have to be comparable on the RV-350 to the R-300: pixel rate and bandwidth.

I just don't see 8 Floating point pixel pipes on 0.13 for the $100 value segment. Would be awesome if true, but I have serious doubts about the viability of that.

So if we assume 4 pipes, then we're talking about needing to get to speeds of 550-600 Mhz of the 4 pipe part to meet the pixel power of the Radeon 9700. Again, for a value part, I don't expect to see clocks that high. (That would be an unprecedented jump in clock rate.)

The biggest barrier though, IMO, is bandwidth. Currently ATI won't implement a 256 bit bus on even their sub $200 cards, so I find it essentially impossible that we'll see 256 bit bus on $100 cards next year. So, to get near the 18-20 GB/sec bandwdith of the 9700, that means, again, using some exotic memory like 500-600 Mhz DDR-II. Again, not going to happen on a value part next year.

From a pure cost perspective, I just don't see RV-350 weighing in with more than a 128 bit bus, and with more than 4 floating point pixel pipes. Functionally, it will be pretty much equivalent to the Radeon 9500 non pro. It will be much cheaper though, and the RV-350 can be faster if we assume a faster clock rate.

By spring or summer, I can see the value parts topping out at 400 /400 Mhz core/memory clocks. Such a part would be clearly faster than the 9500 non-pro, and even faster than the 9500 Pro in many cases. (This 400/400 RV-350 would have a lower pixel rate than the 9500 Pro, but significantly more bandwidth. )

We may see a case of ATI phasing out BOTH the 9500 and 9500 pro in favor of the RV-350. That is, the RV-350 might serve both the $100 (replace 9000) , and $150-$200 (replace 9500s) segment, using clock speeds and memory to differentiate the products. Similar to what nVidia tried to do with the GeForce4 MX. (nVidia tried to use the GeForce4 MX architecture to replace both the GeForce2 MX and the GeForce3 Ti-200.)

Of course, nVidia wasn't too successful with that move, because the 3ti-200 was better in both performance and architecture, as was ATI's Radeon 8500, so the MX-460 flopped. However, the RV-350 would be architecturally on par with the 9500s, and very comparable in performance.

All speculation of course!

 

I'm sorry Alex but that is the funniest statement I've ever read, 'Mr. IMakeUpStories' about next gen products from Nvidia.
On Rage3D you stated Nv30 in Nov, 256 bit bus....blah blah yada yada..

I forgot nobody can make fast graphic cards besides Nvidia...DoH! Wait a minute..currently ATI has the peformance crown so that proves your quote below DOA.

 

While speculating on performance based on nothing but ATI chip codes is a risky business indeed, I'd say that your analysis is a bit pessimistic, Joe.

Consider that it is named RV-280, which implies it's a DX8 type part, with higher performance than todays RV-250. This gives us a higher baseline from which to speculate regarding the RV-350.

This is where I think you are exaggerating. Pixel rate would need to be comparable. Bandwidth would not. They could compromise on bandwidth, and the major hit would be on AA/AF.

Hmm. ATI is selling 8 FP pixel pipes right now at lower than $200. With finer geometry, I could well see them offering the same for $100, particularly if they were to reduce the memory controller/pinout/packaging cost to only support a 128-bit interface.

Agreed.

While I would take issue with the assertion that a 256-bit memory bus necessarily would be prohibitively expensive, I agree that it is likely to remain an "enthusiast" differentiator for most of 2003. The $100 and below cards are unlikely to have it. This will change though as it has in the past.

So my slightly more optimistic prediction would be that the RV-350 is likely to be an 8-pipe chip, with somewhat higher clockrate curtesy of the finer geometry (say 350-400MHz) but only supporting a 128-bit memory bus, at 300-350 MHz unless they go DDR-II.

This will produce a chip which will perform somewhat better than the current 9700 Pro without AA/AF, and worse with AA/AF enabled. Using a 4 pipe design would be cheaper of course. But I fail to see how such a chip would have much more than its DX-9 credentials to offer vs. the RV-280. But I'll grant that it may be enough of a differentiator.

The only thing that is really certain is that the current R9500 non-pro won't stay on the market.

Entropy

 

Thats a risky assumption.

like NV25->NV28 the name change will denote the change from AGP4X to AGP8X. I'd say its risky to assume they will alter this for performance purposes, in fact being a low end part I doubt they will.

Not a given - RV350 may have a lower geomtry rate.

 

Ah. Makes sense I guess. Still, if they have transfer the desing to the process of a new foundry, I'd wouldn't be surprised if they took the opportunity to improve it a bit. As I said, speculating on performance based on nothing but ATI chip codes is a risky business indeed.

The reason the R9500 will go is cost of production rather than performance per se. That chip is way to expensive, unless they are all 97/500 Pro rejects.

Entropy

 

UMC's .15um process didn't exactly do wonders for Parhelia's performance!

Oh, I can understand why they may be the want/need to drop it from the line-up once RV350 is here (and yeilds make it viable), I'm just saying that dependant on how RV350 is configured it may not be a like-for-like swap, unless they included all 4 vertex shaders and its 4x1 (and given this will be targetted mainly for mobile initially I;d be surpised if they did include all 4 VS's).

 

I was not aware that the RV-350 was targetted mainly at the mobile market. That makes quite a difference. A 4-pipe, 2 VS part would make a lot more sense in that case.

Entropy

 

Well, AFAIK M10 is to RV350 as M9 (Mobility Radeon 9000) is to RV250. Thats only based on the interviews / talk about M10 that we've heard so far.

 

As Dave said, M10 is (by all accounts) directly derived from the RV-350 core, just as the M9 is directly derived from the RV-250 core. ATI has a history (as does nvidia for that matter) of deriving their "performance mobile" part directly from their current "value desktop" part.

Another reason why I don't believe the RV-350 will be much more than 60 million or so transisors on 0.13u: need to keep power consumption under control.

I consider that a slight possibility. While it would be a good cost replacement for the 9500 series, I just don't think such a part is cheap and power friendly enough for the $100 and notebook markets.

There are two other likely key differentiators between a 4 pipe RV-350 and RV-280, other than DX9 support:

1) RV-350 will likely be clocked higher, and will come with faster memory compared to RV-280. So there will be a raw performance increase over the RV-280.

2) RV-350 will have multisampling AA and better aniso quality. AA performance on RV-350 will be significantly better than the supersampling AA on the RV-280.