NV40: Surprise, disappointment, or just what you expected?

The more I look at benchmarks and think about this, the more I think that if ATI provides 8x MSAA or more, probably at a minimal hit (given that pretty much every test appears to be CPU bound at reasonable resolutions), and NV will lack anything comparable.

To me PS3.0 is an important feature, but I can see the vast majority of consumers finding better AA (or for that matter better shader performance) a bigger draw. It's going to be interesting to see how it plays out in a few weeks.
 
Chalnoth said:
I just don't think it's possible for the R420 to have dramatically higher shader throughput per pipe than the R3xx.

Here's another argument, looking at transistor counts:
Once again, the transistor count rumored for the R420 is 175 million, with a 16-pipe design.

The Radeon 9800 XT has 107 million transistors for an 8-pipe design.

How many extra transistors would it take to move from an 8-pipe R3xx design to a 16-pipe design? Well, let's look at how many transistors ATI saved when they made the Radeon 9600 XT.

The Radeon 9600 XT has 75 million transistors for a 4-pipe design.

So, if we do some simple math, and assume that if we can take the transistor difference in the 9800 XT and the 9600 XT to be indicative of a transistor difference if we added pipelines, the number of transistors a theoretical R3xx part would have is 107+(107-75)*2=171 transistors.

Given that it's a new core, I expect some improvements, but with barely enough transistors to support R3xx-style pipelines, I just don't buy that ATI could have dramatically increased efficiency with the R420.

This logic makes no sense..
The NV3x had 130 million transistors..4 pipelines and 3 VS units ( or was it 2)
NV4X has 222 million transistors..16 pipelines and 8 VS units
So with 90 Million transisters NV added 12 pipelines and 5 VS units.
yet it would be impossible for ATI to add only 8 pipelines and 4 VS units with 100 million more transistors?? :?
 
Bry said:
So with 90 Million transisters NV added 12 pipelines and 5 VS units. yet it would be impossible for ATI to add only 8 pipelines and 4 VS units with 100 million more transistors?? :?

I don't agree with Chalnoth but the problem with this is that NVidia didn't simply add pipelines for the NV4X. It's a different pipeline architecture.
And the rumours are that the R420 is a tweaked R300 architecture.
 
These discussions about what the R420 can and can't do are not worth it guys. Just weeks ago, ya'll were discussing whether or not NVidia could simultaneously made FP32 run at full speed, support VS3.0 and PS3.0, add RGMS, *AND* 16 pipelines, and have 222 million transistors. (remember all the "no way, it's gotta be 12! and 170 million transistors!" ) When the final chip was revealed, we found it it had all of that, plus 2 shaders per pipe and a separate programmable video processor.

It's more like astrology or numerology. Just wait a little bit longer.

Anyway, If the "in the know" rumors floating around are true (you know who you are), then the 6800U will beat the X800Pro by 15-25%, but the 16-pipe X800 will win with around 10-20%, and the X800XT will boost that more. In the end, with expected gains from the drivers, the two chips will perform alot closer together, but the 800XT will have the edge, unless that 600Mhz 6800 is for real. :)
 
Bry said:
This logic makes no sense..
The NV3x had 130 million transistors..4 pipelines and 3 VS units ( or was it 2)
NV4X has 222 million transistors..16 pipelines and 8 VS units
So with 90 Million transisters NV added 12 pipelines and 5 VS units.
yet it would be impossible for ATI to add only 8 pipelines and 4 VS units with 100 million more transistors?? :?

While I don't agree with Chalnoth reasoning, few minor corrections:
NV40 has 6 VS units, not 8. NV30 had multiple TMUs per pipe, NV40 does not. ATI would have (based on the rumors 50-70MT) more to play with, but only need to add 2 VS units to match NV40. The absence of PS3.0 and FP32, if true, will save R420 a fair number of transistors to be used elsewhere.

However, all of that speculation is meaningless. I think the only hard information we have available about R420 is that there will be 12 and 16 pipeline versions and that's about it. I don't know why anyone would put any stock in 160MT rumor after NV40 ended up with almost 50MT more then first rumored. Judging by how badly off base some of NV40 rumors were, R420 might be a 250MT chip for all we know.

Let's indulge Chalnot for a minute and presume that R420 is R350 with 16 pipelines. Even such chip clocked at 500MHz, would present a formidable challenge to NV40, being capable of performing 8x FSAA faster then R350 can currently performs 4x FSAA. Giving ATi well-deserved credit, we can say with fair degree of certainty that each R420 pipeline would improve upon its R350 counterpart.

In short, given the same transistor budget and lack of PS3.0 and FP32 in R420, I would expect ATi's part to perform better per clock then NV40.
 
Regarding the NV40 - I think you're all forgetting this is A1 silicon which has been available for sometime now. A2 silicon was available at reviewtime, but NVIDIA decided against it.
There's a reason the clockspeeds suddenly went from 475Mhz to 400Mhz - the final ones will be at or over 475Mhz, but NVIDIA's in their "let's make ATI more confident in themselves than they should be" mood.
If anything, this generation will probably be remembered by the gazillions of tricks & traps each company is preparing for its arch-nemesis... But I'd still be inclined to say the R420 will win this war for sveral reasons. We'll see though.

Uttar

EDIT: Seriously, aren't you wondering why Vegetto's sample, which IS for real, and which was in fact a A2, ran at 475Mhz no problem with no mention of a monster cooler? A1 is weeks, if not months old. There never was any last-minute clockspeed drop (AFAIK).
 
If both companies are going out of their way to deceive each other by making an ass out of everybody else, then I'd rather they just sent memoes to each other dictating their exact hardware specifications.

I honestly don't think there's a point in trying to get the other company to blink. It's really not easy to play chicken with a corporation. They aren't going to blink, and they aren't going to panic. If someone like that were in charge of either, they would have gone under long ago.

edit:
They know what their tech can do, and have doubtless laid plans to make sure that any real attempt at subterfuge would need to be needlessly complex and expensive.
 
I very much doubt they would have gone out to reviewers with reworked boards if they didn't have reasons for doing so.
 
Needless to say, this has been one fun Spring so far. I think that things are going to get even more fun once ATI joins the fray. A couple of things here:

Mintmaster, your whole "look at R200 vs RV350" argument is a bit flawed, mainly because we are talking about DX 8.1 hardware vs. DX 9.0 hardware, and the HUGE differences in architecture between the two chips. Yes, R200 was a 4x2 pipeline affair... but it was DX 8.1! It had no floating point pixel pipelines! Not to mention the Vertex Shaders on that were 1.1, while the VS on the RV350 is 2.0! Huge differences, so you can't really paint a big picture about transistor counts using these two chips. If there was ever an "apples and oranges" argument, this is it.

I think it is agreed on that the R420 will not support FP32 and PS 3.0. Both of those options do require more transistors. Now, ATI has said in the past that they could have done full FP32, but it would have required anywhere between 5% and 10% more transistors. Rendering FP32 at a decent speed would have required even more transistors, so the decision to go FP24 for the first few generations of chips was a very good plan, as it did save a lot on transistor counts, and those transistors can be used elsewhere for things like better visual quality (higher AA levels) and making a faster shading chip (1 full and 1 partial shader per pixel pipeline vs. only having 1 full shader). ATI has been working on the R420 chip for some time, and we can most likely take it that the pixel pipeline design has undergone some changes, most of which will help the overall speed and efficiency of their operation (as well as adding another 4/8 pipelines to the mix- giving chips with either 12 or 16 pixel pipelines). We can also guess that the vertex shader has also undergone a lot of work, and I wouldn't be surprised if it in fact was a VS 3.0 compliant piece of hardware. Current R3x0 chips have 4 VS units on board as is, and I can easily see them putting 6 VS units in there, which would give NV a run for its money in that area.

It is my opinion that the X800 will be slower overall than the 6800 Ultra, but that the X800XT will be faster than the 6800 Ultra in most operations. Of course, the X800XT will not support PS 3.0. This may or may not be a sticking point. One big point will be the power draw and heat given off by this chip. I am also betting that it will run cooler and consume less power than the NV40. I also believe that the X800XT will run at over 500 MHz. ATI does not sit on their hands, and the design groups that currently work at ATI are truly second to none.

So, I think ATI will continue to have the performance crown this summer after the dust settles, but it will not hold the overall feature crown. One other thing to consider is that the NV40 architecture still looks like it can run FP16 at significantly higher levels of performance over FP32 rendering, so in applications which utlize partial precision, the NV40 could see some hefty speed increases over pure FP32 (this really only stands to reason).

So, what is it going to be? I have no real clue, cause I haven't seen the new ATI hardware! All I know is that I believe consumers will have two very solid architectures to choose from this buying season, and I do not think that either will disappoint.

One area that could truly define each company is their midrange offerings. With NVIDIA set to release NV4x variants at the midrange and low end, what will ATI do? The FX5200 proved to be a great seller for NVIDIA because it did support DX9 fully (but of course without the speed to actually run it well... at all!). However, major OEM's and many mainstream consumers bought up the FX 5200 in droves, while ATI's 9200 did not sell nearly as well. These feature checkboxes are very important to OEM's trying to sell a computer to the "unwashed masses" who don't particularly know a whole lot about computer hardware and many of the terms associated with it. The 4 pipeline and 8 pipeline NV4x variants could very well compete with the current crop of R350 and RV360 chips on the market, and support PS/VS 3.0 to boot. What do you think the average buyer will choose?

Big questions, and we won't know the answer until at least Christmas!
 
I want to know if Shader model 3.0 can be implemented in any realistic sense in this generation, or if early adopters are going to find out they should have waited for NV50.
 
DemoCoder said:
These discussions about what the R420 can and can't do are not worth it guys. Just weeks ago, ya'll were discussing whether or not NVidia could simultaneously made FP32 run at full speed, support VS3.0 and PS3.0, add RGMS, *AND* 16 pipelines, and have 222 million transistors. (remember all the "no way, it's gotta be 12! and 170 million transistors!" ) When the final chip was revealed, we found it it had all of that, plus 2 shaders per pipe and a separate programmable video processor.
It's more like astrology or numerology. Just wait a little bit longer.

Exactly. I cannot help but comment on certain posters sayings such as Hellbinder for example (this is not an attack in an way) who even one week before the NV40 launch was saying "you all bought NVIDIA's FUD, NV40 is a 16 pipeline card like NV30 was an 8 pipe card" and other similar underestimating comments which all appeared to be false. I prefer to wait 10 days than fall into the assumptions - fud spreading mill.

Anyway, If the "in the know" rumors floating around are true (you know who you are), then the 6800U will beat the X800Pro by 15-25%, but the 16-pipe X800 will win with around 10-20%, and the X800XT will boost that more. In the end, with expected gains from the drivers, the two chips will perform alot closer together, but the 800XT will have the edge, unless that 600Mhz 6800 is for real. :)
Democoder, you got me a bit confused with the above statement. I thought the 16 pipeline part was going to be the XT part ?? You seem to imply that there will be the Pro part (12 pipes), one 16 pipe part and the XT part :?
 
Uttar said:
Regarding the NV40 - I think you're all forgetting this is A1 silicon which has been available for sometime now. A2 silicon was available at reviewtime, but NVIDIA decided against it.
There's a reason the clockspeeds suddenly went from 475Mhz to 400Mhz - the final ones will be at or over 475Mhz, but NVIDIA's in their "let's make ATI more confident in themselves than they should be" mood.
If anything, this generation will probably be remembered by the gazillions of tricks & traps each company is preparing for its arch-nemesis... But I'd still be inclined to say the R420 will win this war for sveral reasons. We'll see though.

Uttar

EDIT: Seriously, aren't you wondering why Vegetto's sample, which IS for real, and which was in fact a A2, ran at 475Mhz no problem with no mention of a monster cooler? A1 is weeks, if not months old. There never was any last-minute clockspeed drop (AFAIK).

YOU are the FIRESTARTER man :devilish:
 
Uttar said:
EDIT: Seriously, aren't you wondering why Vegetto's sample, which IS for real, and which was in fact a A2, ran at 475Mhz no problem with no mention of a monster cooler? A1 is weeks, if not months old. There never was any last-minute clockspeed drop (AFAIK).

I'm not sure what you're referring with this Vegetto-sample but if to the card which run UnrealEngine3, wasn't it reported that people saw artifacts on the screen, similiar to the ones you see when you overclock too much?

And umh, if the final clocks aren't 400/1100, when we're going to find those out? After the X800XT-release? For some reason I have this feeling that, at least from NV's side, the boards which are previewed will vary greatly from cards which are sold.
 
I've already made my choice for this generation. I'm going with NV40 all the way... Not really becauseI don't think the X800XT will be a bit faster, but because I'm only really going to play Doom 3, Call of Duty (And the new one too :devilish: ), and a few other games that are obviously going to be where nVidia simply takes the cake without much of a problem. I don't think anyone is going to argue that nVidia will have ATi beat in many of the upcomming OpenGL games, and that's primarily what I'm going to play.
 
Back
Top