NV30 vs R300?

[BoddoZerg]

Now that everything is 256 bit, can it be taken as a given that the NV30 will have a 256bit-wide memory bus?

It looks like NV30 is going to take a long time to get released... too long for Nvidia's sake. In the time between now and whenever NV30 is out, R300 will reign unchallenged, for the first time ever the fastest GeForce card will be called "slow". It's kinda ironic that the NV30 is based on Rampage, because if nVidia repeats 3dfx's mistake in taking too long to release it, ATi will overrun them with R300 based cards. ;p

 

[Ailuros]

Pardon me, do you have any substantial documentation that prooves that NV30 is "based on Rampage"?

Other than that it looks like it is going to be a tough period for NV until NV30 release. What will happen after that I don't dare making any predictions without knowing the full NV30 specifications.

 

[Joe DeFuria]

The big question, IMO, is whether or not nVidia will actually go with a 256 bit memory interface.

Past speculation (based on nVidia comments) said "no." Now, "just because everyone else" is doing it, people are starting to assume that nVidia will do it. I'm not so sure.

nVidia may be relying on "super fast" DDR or DDR II memory at 128 bits (400-500 Mhz), for their bandwidth increase.

IF that's the case, I think NV30 might have a difficult time showing any significant advantage over R-300.

Most reviewers are stating how "on paper" NV30 beats the R-300. There's no reason to doubt that's not the case. Say it's an 8 pipeline board running at 400+ Mhz. That certainly beats the R-300 on paper. But what if it's paried with 128 bit DDRII memory at even 500 Mhz?

In such a cse, you can easily forsee situations where NV30 would be faster (fillrate limited, but not bandwidth limited), and R-300 would be faster (bandwidth limited). And most would agree that being faster in bandwidth limited situations (high quality modes), is more important.

In any case, I'm eagerly awaiting the NV30, because whether or not it is superior, we can expect a mini "price war", because certainly, the parts will be competitive with one another.

 

[Ailuros]

Actually if you overgeneralize an assumption it can turn out hideous at times.

Although irrelevant to NV30, a 256bit bus should not be faster in EVERY case then a 128bit bus. Depends on many things, architecture first and foremost.

In a hypothetical case scenario what would make you think that on equal specifications a 128bit bus TBR architecture would be disadvantaged by a 256bit bus on an IMR?

As said this has nothing to do with NV30 as I really don't know.

 

[LittlePenny]

To be honest I think these NV30 vs R300 conversations are close to pointless. We didn't discuss GF4 vs R300 when the GF4 was first announced, and if we did we shouldn't have. If we don't know anything(aside from rumors) about a product that isn't out yet then how can we compare technology?

 

[Joe DeFuria]

Although irrelevant to NV30, a 256bit bus should not be faster in EVERY case then a 128bit bus. Depends on many things, architecture first and foremost.

Exactly.

And this is what concerns me with the web reviewers already claiming NV30 "superiority" over the R-300 based on specs. That's a very dangerous game to play for two reasons:

1) Specs can and do change before the product is released...(and for nVidia in particular.)
2) Specs must be combined with "architecture" to determine actual perfomance (as you said.) And if NV30 is a "new architecture" then real-world performance guesses are really a shot in the dark. Are the web reviewers taking PR estimates of the benefit of the new architecture when reaching their conclusions?

I would really prefer that the web reviewers just keep their mouths shut about performance of "future parts" until the time when they can actually legally disclose the specs.

 

[Ailuros]

I usually don't defend most web site reviewers much for obvious reasons. But at this stage I can't judge their predictions either since I haven't the vaguest idea what they know and don't know or what they are looking at under NDA or not.

In all fairness if I think about Anand he did "predict" that R300 would produce some impressive scores in UT2003. If any possible "assumptions" are based on a similar basis then I don't have much to object about it other that I'd really prefer to see those hints missing concerning all hardware that is still under NDA.

It gets more an issue of ethics to state "hey that's fine but I know of something better coming out soon". Where "better" is subjective and "soon" rather vague.

It would be far more simple if all of them would stick to JUST what they are allowed officialy to state, wether they know or not is irrelevant. That's the stage where many reviewers before analyzing hardware to the best possible degree are endangered to shoot themselves in the foot and turn out with minor or major at times mistakes.

No one is perfect and we all make mistakes but a major reviewer has quite a responsibility against his readers and I wish they would consider it more often.

 

[Doomtrooper]

This is incorrect anyhow, it will be Nv30 vs the Radeon 9700 @ .13 Micron...

 

[hughJ]

assuming it's just a 'typical'(whatever that means) die shrink to .13, what % clock increase do you think would be made possible by it?

 

[Dave Baumann]

Another point is that if ATI have 'hand tweaked' R300 to get the speeds up to what they say then to get similar a relative hike in performance going to .13um they would have to do the same, which will take more time. So (if we were to believe they are actually going to go .13um with R300 reasonably soon) have they given themselves this time, or will they have to rely on the tools to do the layout hence losse some of the advantage?

 

[Nexiss]

If NV30 is significantly faster (and at the moment I doubt it will be significantly faster), I wonder if ATI will counter with a dual chip configuration. At the very least I would expect a switch to .13 micron and a higher clock, and maybe DDR2.

 

[Geeforcer]

With a single-chip card going for $400, I don't think a dual-chip one would make a effective counter to anything in the near future.

 

[pascal]

Another point is that if ATI have 'hand tweaked' R300 to get the speeds up to what they say then to get similar a relative hike in performance going to .13um they would have to do the same, which will take more time. So (if we were to believe they are actually going to go .13um with R300 reasonably soon) have they given themselves this time, or will they have to rely on the tools to do the layout hence losse some of the advantage?

Probably they will be able to scale down with some automated tools because the basic hand design optimization was already done.

NV30 will have twice the GF4Ti transistors but the manufacturing process only give a 20% area reduction and 30% speed boost for the same heat.
http://www.tsmc.com/english/technology/t0101.htm

The hand made design make things smaller and consequentlly faster. ATI with some hand made design may have a clear advantage.

This is why I say I will not be surprised if R9700 have the same die size and Mhz as NV30. One possible advantage NV30 may have is the use of faster DDR-II, but by than a simple ATI design refresh with DDR-II could be available.

 

[RussSchultz]

Another point is that if ATI have 'hand tweaked' R300 to get the speeds up to what they say then to get similar a relative hike in performance going to .13um they would have to do the same, which will take more time. So (if we were to believe they are actually going to go .13um with R300 reasonably soon) have they given themselves this time, or will they have to rely on the tools to do the layout hence losse some of the advantage?

Probably they will be able to scale down with some automated tools because the basic hand design optimization was already done.

NV30 will have twice the GF4Ti transistors but the manufacturing process only give a 20% area reduction and 30% speed boost for the same heat.
http://www.tsmc.com/english/technology/t0101.htm

The hand made design make things smaller and consequentlly faster. ATI with some hand made design may have a clear advantage.

This is why I say I will not be surprised if R9700 have the same die size and Mhz as NV30. One possible advantage NV30 may have is the use of faster DDR-II, but by than a simple ATI design refresh with DDR-II could be available.

Perhaps incorrect on a couple of points:

ATI could go to .13 using standard cells in an automated manner, but it would not inherit any improvements of the hand tweaking from the hand tweaked .15 design.

.13 from .15 will reduce power by reducing the gate voltage. A standard .15u device uses 1.5V for Vcc; a Standard .13u device uses 1.2V. In reducing the voltage by 25%, the power will actually go down about 45%. (squared relationship)

 

[MuFu]

If they do go the DDR-II/QDR-II route there are additional power consumption savings to me be had (at the moment the memory subsystem consumes 12-15W; about 1W per chip and 0.5W for each I/O branch). QDR-II/DDR-II should operate at lower voltages (1.8V?) and feature on-die termination, which will free up precious PCB space and possibly allow for cleaner/easier routing.

MuFu.

 

[Bogotron]

What is to say Nvidia hasn't had to tweak the layout manually as well? Considering all the "problems" going to .13 micron and the reports that automated tools do not do a good enough job of the layout, maybe ATI is not alone in the hand-tweaking business.... Just a thought.

 

[Xmas]

This is incorrect anyhow, it will be Nv30 vs the Radeon 9700 @ .13 Micron...

This is also incorrect, it will be whatever card you can buy for a certain price at a certain time vs. its likes.

I expect NV30 to come earlier than R300 @ 0.13µ

 

[jvd]

well in that lodgic it would be the geforce 4 ti 4600 vs the r300 since you can buy them at the same price range at about the same time . Besides i predict that the nv30 on .13 will come out much later than the r300 on .15 micron

 

[pascal]

Perhaps incorrect on a couple of points:

ATI could go to .13 using standard cells in an automated manner, but it would not inherit any improvements of the hand tweaking from the hand tweaked .15 design.

.13 from .15 will reduce power by reducing the gate voltage. A standard .15u device uses 1.5V for Vcc; a Standard .13u device uses 1.2V. In reducing the voltage by 25%, the power will actually go down about 45%. (squared relationship)

Russ,
Could they use some kind of geometrical design rule based tool to scale down the original design? After that some points here and there could be tweaked again. Just guess.

About the power, this scalling down did not happened with thoroughbred (.18 micron to .13 micron). There are other things that may not go as well as we want.

I have to go now. See you tomorrow

 

[pascal]

Few minutes before I go to a long meeting.

Russ,
After a second thought my guess you are right about the power scalling, but it means that NV30 will probably be in the AGP specs. The problem is that ATI went beyond the AGP spec.