another Dave Orton interview

[AlphaOmega]

i studied the reviews and noted these points
1. the 6800GT performed better then 6800Ultra or Ultra Extreme in some benchmarks
2. the 6800GT which is A02 board reached stable 425 mhz on overclock(firesquad review) with 1 molex connector(lower power consumption)
3.When only AA is enabled on NV40 there is less then 10% drop in performace(4x mode used)
4. Without AA or AF NV40 is on par, a little above, or a little below in performace even in FarCry.
5.When AA and AF(together) was used a 40% + drop in performance.

Seriously i think those who believe that the NV40 performance will increase are not being too optimistic. From an engineering point of view it seems that this board has some more potential. Another revision would lower power consumption more and give a higher clock most definately. But whats important is the drivers of the card at this point. Now NVDA as a company has made many mistakes, but i do have faith in the knowledge a programmer or a engineer has at NVDA. After all they have nothing to do with the outrageous stunts NVDA has pulled. Why i say nothing? well because they either do it or get fired, NV30 didnt perform upto par, they were asked the question could they make it perform up to par, they either said yes or they were fired. They tried and failed but you c they tried. Thats what is important because, it seems NVDA has given its engineers alot more resources and i very much doubt that they somehow can not perform up to par of other engineers. I also believe that given some time because i believe the engineering part of NV40 has been done correctly that the driver team will also make this card shine. There is no reason why to bring NVDA into an arguement about the NV40 if u understood correctly what went on behind closed doors in NVDA at the time of NV30. I do not wish for any company to win over another but i do hope that this design will shine for i see hope in it.

 

[Althornin]

I'll take that one step further, ATIs 2XAA looks as good as NVs 4XAA.

Obviously, FS doesn't agree with you.

Since the 6800s 4XAA is identical to its 8XAA

This is obviously not true at all. NV's 8xAA mode uses SS + MS.

What is this, a matter of national pride or something?

why do you keep going back to FS's opinion like it is gospel?

 

[Sabastian]

You have a sample size of one picture. Way to go.

Trust me, you are wrong. There is a difference between using MS vs using a mix of MS + SS. That difference isn't striking in every single picture at every single angle. You should talk with the reviewers and ask them to show you examples of where IQ will differ among the two modes.

Yeah. The major difference you'll see with using SS AA is that it applies AA to everything on the screen where MS AA only applies too primitives. Correct me if I am wrong here but isn't this SS + MS AA is a little redundant labeling? The problem with SS is that it eats up your memory bandwidth particularly when you apply it to alpha textures commonly used for things like leaves. The R200 suffered major performance hits because it used SS AA. I like SS but if it is not usable then... what good is it?

 

[AlphaWolf]

What facts are you using exactly? You are ignoring the "fact" that there is only a 1-4fps difference between 6800U and 9800XT in FarCry at high resolution with AA/AF, even though the 6800U surges past the 9800XT in virtually every other game. You are also ignoring the "fact" that FarCry is extremely buggy on the 6800 cards.

Fact, nv40 is rendering much more simple shaders than X800.
Fact, nv40 is not rendering some of the more difficult effects in far cry.

How does fixing these things lead to a performance improvement?

Just because the drivers are buggy doesn't mean they are slow, that's your interpretation, but it's not necessarily the truth.

9800Xt could be approaching the 6800U because it has better shader handling, because it isn't running into register issues. Hell maybe because it has a higher clock speed. I don't really know, and neither do you.

There are other games (mafia) where the nv40 finished behind a 9800XT.

Was this at 1600x1200, 4xAA/8xAF? And what makes you think this is in any way applicable to FarCry, or applicable to games in general?

I believe the test was run at 1280x1024 with AA/AF. Nordichardware ran the test. What makes you think it isn't applicable to games in general? I merely cited that far cry and tomb raider aren't the only games where nv40 is lagging behind badly in performance. You seem to think these are isolated abominations. I really don't know if that's the case.

But it is still a significant advantage in fillrate, 12%, is it not? The 6800U also has nearly double the fillrate of the 9800XT too, correct? And yet that is not in any way reflected in the FarCry results.

12% isn't 31%. I was merely pointing out than there is more to performance differences than drivers.

Then you are just being ignorant. I showed you a FS article where the 4xAA between ATI and NV cards was very comparable. I also told you about how some reviewers have found NV's AF method to be clearer/sharper than ATI's. I just don't see how anyone in their right mind can say that image quality between the ATI and NV cards will not be very similar this round.

I am not being ignorant, I am being accurate, you seem to want to compare only 4xAA. If nv40 didn't offer AA at all would you say it's not fair to compare it and limit comparisons only to non AA? For people playing older games 6xAA is very usable. I wouldn't say that the IQ advantage for ATi is huge, but the advantage exists and right now it's a lot more usable than ps3.0 is isnt it?

When did JHH claim that ATI counts transistors in the same way as NV? The NV40 should have more transistors than the R420 when all is said and done, so what's your point? Since when do people purchase cards based on transistor count?

JHH stated that nv40 had 40% more transistors than r420. Do the math, that's about 60 million. I never said that people base their purchase of video cards on transistor counts. I said it was a tradeoff for ATi and Nvidia. More transistors equals higher cost, equals a higher power requirement, limits core speed, leads to lower yields. That's what I was saying.

You seem to be really grasping for straws here to justify one company's design decisions vs the other company's design decisions.

I am not grasping at anything, merely presenting the facts. And that's sorta what this thread is about isn't it? One IHV's decisions vs anothers.

 

[jimmyjames123]

Fact, nv40 is rendering much more simple shaders than X800.
Fact, nv40 is not rendering some of the more difficult effects in far cry.

Fact, the Forceware 6x.xx drivers are extremely buggy in FarCry, making proper apples to apples comparison impossible at this time.

Fact, there have been tests done showing that the NV40 is at times faster using PS 2.0 vs using PS 1.1 shaders.

Fact, the 6800 Ultra is generally faster in shader performance than the X800 Pro, looking at shadermark tests. The X800 XT is often faster than the 6800 Ultra in Shadermark, but at times the 6800 Ultra is faster too.

How does fixing these things lead to a performance improvement?

Perhaps NV could answer this for you?

9800Xt could be approaching the 6800U because it has better shader handling

Have you actually seen Shadermark pixel shader tests between the 9800XT and 6800U? Obviously not.

I believe the test was run at 1280x1024 with AA/AF. Nordichardware ran the test. What makes you think it isn't applicable to games in general?

You think that the results in this game are representative of anything at all? The 9800XT was less than 1 fps behind the X800 Pro! Really grasping for straws here.

I merely cited that far cry and tomb raider aren't the only games where nv40 is lagging behind badly in performance. You seem to think these are isolated abominations. I really don't know if that's the case.

FarCry is hardly worth debating over anymore. It is impossible to get an apples to apples comparison at this point because of buggy drivers. Let's wait and see what happens, and revisit this later.

12% isn't 31%. I was merely pointing out than there is more to performance differences than drivers.

12% isn't 31%, but then again, 31% isn't 90% right? 90% (?) would be the fillrate advantage for the 6800U over the 9800XT.

I am not being ignorant, I am being accurate, you seem to want to compare only 4xAA. If nv40 didn't offer AA at all would you say it's not fair to compare it and limit comparisons only to non AA? For people playing older games 6xAA is very usable. I wouldn't say that the IQ advantage for ATi is huge, but the advantage exists and right now it's a lot more usable than ps3.0 is isnt it?

How many times do we have to say that ATI has the slight advantage in AA? You constantly keep ignoring the fact that IQ is not solely determined by max AA levels! My goodness. There is also AF quality to consider when talking about image quality.

JHH stated that nv40 had 40% more transistors than r420. Do the math, that's about 60 million.

And ATI stated that they count transistors differently than NV, so do the math! You are just arguing semantics here.

More transistors equals higher cost, equals a higher power requirement, limits core speed, leads to lower yields. That's what I was saying.

That's not exactly true either. The 6800 GT has almost 90 million more transistors than the 5950 Ultra, and yet has similar power requirements and higher yields! Process improvement is what leads to more efficient use of transistors, high yields, and higher clock speeds.

I am not grasping at anything, merely presenting the facts. And that's sorta what this thread is about isn't it? One IHV's decisions vs anothers.

No, I think you are just so bullheaded and fixed in your ways that you just can't imagine that maybe NV really can achieve near parity to ATI this go around with respect to general IQ and performance. Time will tell, I assure you.

 

[Heathen]

Obviously, FS doesn't agree with you.

hint, read more than FS.

 

[AlphaWolf]

Perhaps NV could answer this for you?

And perhaps they can't. You still have only your hopes.

Have you actually seen Shadermark pixel shader tests between the 9800XT and 6800U? Obviously not.

I meant specifially in that title. There are some tests that the 6800U was also unable to run at all.

You think that the results in this game are representative of anything at all? The 9800XT was less than 1 fps behind the X800 Pro! Really grasping for straws here.

You're the one grasping at straws. Looks to me like the radeons are cpu limited. I really don't know what's wrong with the nvidia cards, but they sure are slow in comparison.

FarCry is hardly worth debating over anymore. It is impossible to get an apples to apples comparison at this point because of buggy drivers. Let's wait and see what happens, and revisit this later.

Sure lets just forget about all the tests where nvidia loses badly. Perhaps we can talk about the open gl games where ATI is suffering through at 175 frames per second.

12% isn't 31%, but then again, 31% isn't 90% right? 90% (?) would be the fillrate advantage for the 6800U over the 9800XT.

Do you have the memory retention of a gerbil? Read my posts or stop bugging me. I HAVE NEVER STATED THAT ANY ONE THING WAS THE DIFFERENCE, I MERELY SUGGESTED THAT THERE ARE OTHER FACTORS AT PLAY THEN DRIVERS.

How many times do we have to say that ATI has the slight advantage in AA? You constantly keep ignoring the fact that IQ is not solely determined by max AA levels! My goodness. There is also AF quality to consider when talking about image quality.

AF is roughly equal. Certainly no difference decernable in any screenshot of actual gameplay I have seen. Please don't quote FS again, it's getting old. 6XAA is an advantage, temporal AA is an advantage. I know it's not fair because the IHV you worship doesn't provide them, but who said life was fair. It was nvidia's decision as to what they would provide.

And ATI stated that they count transistors differently than NV, so do the math! You are just arguing semantics here.

Go back and read what I posted that in relation to, go back and understand the point of this thread. And after you do that feel free to look up the word semantics. But for gods sake please don't waste anyones time posting it here, we actually know what it means.

That's not exactly true either. The 6800 GT has almost 90 million more transistors than the 5950 Ultra, and yet has similar power requirements and higher yields! Process improvement is what leads to more efficient use of transistors, high yields, and higher clock speeds.

The 6800GT is also running 125 mhz lower clock speed and running with much lower power gddr3. As for higher yields, that remains to be seen. JHH claiming it will be true by the end of this year doesn't make it so. Not that the nv3x yields were anything to brag about.

No, I think you are just so bullheaded and fixed in your ways that you just can't imagine that maybe NV really can achieve near parity to ATI this go around with respect to general IQ and performance. Time will tell, I assure you.

Well unless they add a new AA mode I don't really see how than can match ATi's IQ. Time will most definitely tell. And it's hard to not be bullheaded when you are arguing with someone who has no facts to back up his arguments.

 

[webmedic]

There were some games where the X800 XT finished ahead and where the 6800 Ultra was still around 100 fps. The 6800 Ultra also was arguably as fast or faster in games like Call of Duty, RTCW, Halo, Jedi Academy, NWN, Prince of Persia, Serious Sam. And the differences in most other games was relatively small. The X800 XT PE generally had a slght edge at 1600x1200 with 4xAA/16xAF over the 6800 Ultra, but then again it depends entirely on the game. I think there is little question that the X800 and 6800 cards are peers in terms of overall performance. One card is not inherently superior to the other with respect to performance this go around.

thats not a view that the majority hold with you. There are threads both here and at nvnews were the polls from users at those forums choose the r420 based cards. I tried to find something like this at rage3d also but was unable to do so. This to me is pretty telling when even the majority of the users chose 4r20 based cards over the nv40 solutions. Admittedly it wasn't by much. Here on beyond though the r420 got like 60 something percent and another 19 % or so said they tied.

Not saying which is which but thats the current public opinion on which one is best.

 

[DemoCoder]

Polls online are meaningless. Self selected samples.

In any case, it should be expected that a 400Mhz 16-pipe 6800 will lose to a 500+Mhz 16-pipe part, with faster RAM to boot. If the XT didn't win, it would be pretty bad.

The problem with FarCry is two fold: Currently, it isn't running the same path as ATI cards because of application bugs. Secondly, the drivers are still buggy and immature. It's NVidia's fault, they should have delayed the launch until the drivers were in a better state. They also should have gotten a FarCry version into reviewer's hands that wasn't detecting their flagship card as a NV3x.

 

[martrox]

I'm personally of the opinion that, unlike the last generation where ATI spanked nVidias collective ass, things are much closer now. I can see where someone might have interests in buying one card over the other..... IF I were interested in programming advanced shaders, I'd really have to consider the 6800 as the card to have. But, facts are I'm just a gamer. To me, IQ is the alter that I worship on. The overall take, on everything I've read, is that the X800 is the winner in what I want in a video card. To me the size & heat of the 6800 are a non factor (I watercool) and I have the power supply, too.

Notice I didn't mention speed.... because at their max IQ all the cards are more than fast enough for everything I play. It's just that the ATI will be running better FSAA at those resolutions and speed, in most everything. How much better? To me it's a noticeable difference. Here's an example. I'm playing City of Heroes ATM. On my 9800Pro I can play at 1280X1024, 4Xfsaa 8X AF with everything turned on. Or, I can turn down the textures and run at 6Xfsaa........ I prefer the 6X fsaa.... BUT I miss the higher textures. If I had a 6800, I'd have no choice. BTW, I'm running medium textures and 6X...and I can't wait to get a X800.... Because with either X800 I'll be able to play at 1280X1024, 6XFSAA, 16AF and high textures. You will never be able to do that with the 6800... no matter what the "final†core & mamory speeds are. But, remember.... I'm just a gamer! BTW, CoH is an OpenGL game.

Now, mind if I mention drivers now? ATM, ATI is (supposable) hard at work to upgrade their OpenGL drivers, and it's pretty universally accepted that ATI a)needs to, and b) there's plenty of power to do so. So, IF true, you will see major upgrades in openGL with ATI - this is the only area where the 6800 really have a lead driverwise. I keep hearing from the faithful about the wonders of nV's drivers. Facts are that ATI, with the exception of the previously mentioned OpenGL drivers, are much better that nVidia's, and have been for the last year+. Ever since the FX series stumbled onto the scene I've heard the faithful mantra "just wait till nVidia does their driver magic and finds that "huge" speed increase". Now, lets examine exactly how many times the nVidia driver "pixie dust" has been used..... Um.......how about once! I believe it was the GeForce 2's, just before the 3dfx V5-5500 came out the original Detonator drivers came out with a huge performance spike. Let's face it, nVidia was laying in wait with these drivers... they set 3dfx up - and there was nothing wrong with doing that. I propose that nV, who had just started to dominate the market, actually produced the original slower drivers in order to take the wind (and market share!) out of 3dfx's sales (pun intended).

Again, don’t get me wrong, this is no putdown of nVidia’s driver team. The only increases I have seen with regard to nV’s driver have been the expected incramentle driver increases expected of newer technology. The GF3’s got better over time, the GF4’s did too. But none of these were of the “magic pixie dust†original detonator increase. Now…the FX’s on the other hand, showed the problems of nV’s pixie dust…. There was not a single, large performance increase without a subsequent IQ drop. The magic of the pixie dust was, well, never really existed.

So, this begs the question, will the 6800 drivers get faster? Sure they will – BUT anyone that thinks there is extreme (read original Detonator) speed to be found is living in the land of the….um….Pixies! Facts are that the 6800 will get better – they kind of have to, don’t they? Probably they will gain some speed over ATI due to the newness of the architecture vs. ATI’s more mature drivers. But to think there’s a hidden 20-50% speed increase is absolutely absurd.

One last thing. During the next 12 month, it’s going to be ATI that has the better drivers – many think they have had the lead here for the last year+. Again, not a knock on nV, just the fact that it’s a new architecture. (see Farcry as an example)

 

[Bjorn]

Probably they will gain some speed over ATI due to the newness of the architecture vs. ATI’s more mature drivers. But to think there’s a hidden 20-50% speed increase is absolutely absurd.

I doubt that we'll see 20-50% speed increase overall. But i definitely think they can gain 20% or more in shader limited applications.

 

[DemoCoder]

Subsequent NV3x drivers did show performance gains in shader throughput. Of course, it's hidden by the fact that they did other things which you conveniently subscribe 100% of the gains to, but in PS2.0, throughput was increased.

The NV3x was hampered by a severe penalty for using > 2 registers. Thus, if the driver compiler was able to, for example, reduce the number of registers used by 2 on a given input shader, it yielded large speedups in PS2.0 throughput.

You can see the difference today by writing a shader which you think takes N registers, and you will see that FX Composer/NVShaderPerf can in a majority of cases, reduce it to 2 registers.

The fact is, the current 6800 drivers are immature. Not just from a performance standpoint, but from rendering problems, crashes, and missing features (DXVA acceleration). There is a 20% boost left in PS2.0 shader throughput from what I know. The current shader compiler was written for NV3x. Alot has changed since then, in terms of registers, dual issue, some instructions getting alot cheaper, others getting more expensive.

(the dll name for the unified compiler from the forceware drivers, nv_perf3x.dll should tip you off)

 

[Dave Baumann]

ATI's newer compiler is also said to gain up to about 20% in long shader cases. It'll be interesting to see how this translates into real world performances as you aren't going to squeeze anything out of PS1.x length shaders and most PS2.0 shaders in games are still relatively short.

 

[DemoCoder]

They might be able to squeeze more performance in short shaders by taking more advantage of the mini-ALUs (e.g. recognizing more constructs where they can rearrange expressions to fit the mini-ALU ops)

 

[Dave Baumann]

Given the ratio of textures to instructions that are likely in PS1.x its generally the texture accesses that will be the biggest inhibiting factor in these cases (I'm told that while the DEC guys knew their compilers, it took them a while to get to grips with the apparent unpredicability of texture accesses).

 

[tEd]

Given the ratio of textures to instructions that are likely in PS1.x its generally the texture accesses that will be the biggest inhibiting factor in these cases (I'm told that while the DEC guys knew their compilers, it took them a while to get to grips with the apparent unpredicability of texture accesses).

Could you tell a little bit more about those DEC guys. What kind of company is DEC , did ati licenced a compiler from them or did they wrote one from scratch for ati or how does this cooperation work?

 

[DemoCoder]

Well, on PS1.x, if you can eliminate an instruction or two, you can eliminate a stage/loopback, which gains you back anywhere from 12-25% throughput on a small 8-12 instruction shader. Potential gains on short shaders are higher than longer shaders.

 

[Dave H]

Could you tell a little bit more about those DEC guys. What kind of company is DEC , did ati licenced a compiler from them or did they wrote one from scratch for ati or how does this cooperation work?

Would be fun to see Mr. Spink answer this one, but in the meantime...

DEC--the Digital Equipment Corporation--was one of the leading computer systems companies in the 70s and 80s, comparable to an IBM today, or even an Intel. DEC first gained prominence in the late 60s with their PDP series, which introduced the concept of the minicomputer at a time when the computer industry consisted almost entirely of the mainframes of IBM and others. The C programming language and the UNIX operating system were both originally developed (by AT&T's Bell Labs) for the PDP-11.

In the 70s, DEC introduced the VAX architecture. VAX eliminated most all of the arbitrary limitations of the PDP series and is universally considered the epitome of the fully orthogonal CISC ISA. VAXen were, I believe, the most successful computer line of the late 70s and early-to-mid 80s.

By the late 80s, however, RISC-based computers running UNIX had clearly taken the price-performance title from VAX, so DEC responded with their own RISC architecture, the Alpha (incidentally the first 64-bit ISA). The Alpha held the single-threaded performance crown (at least as measured by SPEC) for most of the 90s, despite often being fabbed in processes a half to a whole generation behind some of their competition, particularly Intel. This was first of all testament to the Alpha design team, generally regarded as the best in the business and their designs the most elegant. And secondly--and pay attention here, this is important--to the compiler team, again regarded as either the very best or, in the late 90s and early 00s, second only to Intel, particularly in the sort of linear algebra heavy code that makes up much of SPECfp--and nearly all of 3d graphics workloads.

As a business, however, DEC was not that well run, especially into the mid 90s. DEC got sold to Compaq in 1998 as the complicated result of a settlement in a patent infringement suit DEC brought against Intel over aspects of the P6 architecture. Despite its superior performance, Alpha never really caught up to Sun, HP and IBM in terms of software support. And the marketing of the Alpha was incompetent at DEC and almost nonexistent at Compaq, which after all was not really used to being in the high-end systems market.

As part of Compaq's merger with HP, the Alpha team was sold off to Intel (where most of them are designing Tukwila, the first radically new IA64 core since McKinley) and Alpha development was ceased. Both then and during the DEC->Compaq transfer, though, a good number of ex-DECers took the opportunity to work elsewhere. For example, Dirk Meyer--lead designer of the AMD K7--and much of the K7 team were ex-Alpha engineers.

And now, apparently, a number of engineers from the Alpha compiler team have turned up at ATI working on shader compilers. I don't know the details, but I would guess this happened about the time of the Compaq-HP merger or soon afterwards. In any case, given their track record and the so-far rather unimpressive state of shader compilers, it's reasonable to expect impressive performance gains from them.

(Although IMO NV40 will see much greater gains just from having so much more low-hanging fruit still left on the tree.)

edit: One final comparison that might help. The way the Alpha team was viewed by many in the MPU enthusiast community is a lot like the way PowerVR is viewed here in the GPU enthusiast community: as something between a potential top-tier competitor and a clearly superior architecture which was perennially held back in the market due to annoying business decisions and lack of financial resources. The main difference is that where PowerVR gets its mystique from utilizing one extremely elegant approach, the Alpha designers were playing the same game as everyone else, just with more overall elegance and excellence of design.

 

[3dcgi]

Could you tell a little bit more about those DEC guys. What kind of company is DEC , did ati licenced a compiler from them or did they wrote one from scratch for ati or how does this cooperation work?

I'm no expert in DEC history, but I believe it stands for Digital Equipment Corporation. They made the Alpha chip and were bought out by Compaq a few years back. Many of their engineers went to Intel and other tech companies.

Edit. I see Dave H beat me to the punch. Just consider my post the cliffs notes. I.e. summary if you aren't familiar with cliffs notes.

 

[tEd]

thanks a lot for this history lesson