NV40/R420 Face-Off May 07, 2004

Round One

Page:: ( 1 / 6 )

Will the real XT please stand up?

Jakub: So I'm sitting here looking at the benchmarks in the X800 Pro and X800 XT Platinum Edition preview and recalling a conversation I had with Brandon when he arrived in Toronto and attended ATI's press event. He came away with the highly erroneous impression that the X800 would be faster. Unfortunately for him, his own article proves me correct. The X800 XT isn't faster than the 6800 Ultra. In fact, ATI didn't even have the minerals to bring the real XT to the party, instead they come to the table with a Platinum Edition card. Why would you bring a ringer, unless you're scared that your regular product can't measure up to the competition?

Whatever, I'll let Brandon have his cheat card. It still gets beaten in half the benchmarks, despite having the more mature drivers in Brandon's own words. Mature drivers are all great and wonderful, on release, but this is still the pre-season, time for testing and tuning, and I don't remember anybody winning the Super Bowl in August. Plus, if we look back on NV3X, we know damn well just how much NVIDIA can squeeze out of its cards with every driver release. They went from gimp when the NV30 launched, to being neck-and-neck with ATI now with the 5950 Ultra. NV40 is a wildly different architecture, and I'm willing to bet a Tommyburger and Tommy's chili cheese fries that that NVIDIA will improve the performance of its card a lot more over the coming year than ATI can. I mean seriously, what is R420? ATI abandoned the forward-looking R400 in favor of a card meant for "today" performance, an overglorified R300 with double the pipes, one new feature (3Dc) and not even *half* as many shader instructions as the 5950 Ultra has. The cards are so similar, ATI could just stick a new name into the driver file and be as optimized as they'll ever be. So Brandon, if you want to defend technology ATI dug up in archaeological excavations from 2002, you're more than welcome. At best you'll just have another Alamo on your hands, Texan.


Brandon: Jakub, looks like you got a whiff of something hallucinogenic up there in Canada…I know it’s a green country but wow!

Why don’t you take a moment to put down the pipe for a second and take a closer look at the benchmarks I posted. The RADEON X800 XT Platinum may not win all of the benchmarks, but it won the benches that matter: the games that deal with next-generation titles like Half-Life 2. The X800 XT Platinum Edition even won the half-baked next-gen benches like Far Cry, which uses 1.1 shaders for most everything. In my testing even the $400 X800 PRO gives NVIDIA’s flagship GeForce 6800 Ultra a run for its money at high-res with AA and AF. And did you happen to catch the Tomb Raider benches in the GeForce 6800 Ultra/GT preview? At 1600x1200 it’s a clean sweep for ATI with AA enabled!

Finally, don’t forget to check out the 3DMark 03 game test results. The test to really pay attention to is Mother Nature, the only game test in 3DMark 03 that uses 2.0 shaders. ATI wins there too. To borrow a term you can understand, ATI scores a hat trick in all of the next-gen benches.

And speaking of bringing in a ringer, what’s up with the overclocked GeForce 6800 Ultra showing up on my doorstep at the last minute? You were at NVIDIA’s Editor’s Day last month, do you remember any of NVIDIA’s employees mentioning adding an overclocked 6800 Ultra part to their lineup? Clearly NVIDIA saw the writing on the wall with X800 and tried to crash ATI’s party. Only problem is, even the overclocked 6800 Ultra isn’t able to outpace ATI’s RADEON X800 XT Platinum. Like the old saying goes, “never bring a knife to a gunfight” (especially if it’s a SIG).

Looks to me like NVIDIA’s best needs to go back to the drawing board. And besides, the Platinum Edition isn’t a ringer. It’s going into full production as we speak, and will be available directly from ATI and all of its board partners. Last time I checked, you can’t say that for NVIDIA’s overclocked 6800 Ultra part. I wish I could tell you the background story on how the Platinum Edition board came to be, it had nothing to do with the 6800 Ultra or NVIDIA, but trust me, this isn’t a special edition, for a limited time only type of part. Don’t worry, the truth will be revealed once the PCI Express boards hit retail.

Round Two

Page:: ( 2 / 6 )

Hype-Lie 2

Jakub: Brandon, listen to yourself. You're clearly being delusional here. A knife to a gunfight? It's more like I'm the one with the SIG, a laser sight, an upgraded trigger and all the goodies, and you came with some flint-lock smoothbore pistol from the 17th century. Just which pixel shader version does ATI support again? PS 0.3, or are they up to 0.4 now? I'm sorry, it's just so hard to tell from the lofty heights of Pixel Shader 3.0. You had best remember that half the reason NVIDIA got kicked in the nuts last time around was because they couldn't deliver the latest shader model features.

As for arguing that Far Cry and Tomb Raider are representative of next generation technology, what's up with that? That's like going to the Detroit Auto Show, picking up the latest and greatest Korean-built econobox and telling people that's the car of the future. No, wait, in Tomb Raider's case, it's more like spit-shining a Ford Pinto. And please, save yourself the embarrassment of arguing anything about Half-Life 2. Between Gabe "I sell Half-Life 2 every time I pitch someone a Radeon" Newell and that overhyped Source engine of Valve's, I think it's safe to discount that game as a worthwhile benchmark. Just what does it exactly do that's special? Ooooh… facial animation. Yeah, that's really going to stress the GPU. Physics? Not the graphics card's department. The lighting is straight out of DirectX Ancient History - lightmaps… pssh. The shader tricks that Valve showed are now appearing in every game. A man made of water was impressive last summer and might have been even when September 30th came around, but if you've played Painkiller, you'd have seen something way more impressive than a mere man made of water. Face it, the benchmark standard of the future will be from id Software, as it has always been and ever shall be (amen).

I see you've failed to counter any of my remarks regarding driver maturity. No wonder the Radeon X800 does so well with anti-aliasing and anisotropic filtering, it's not so different from its predecessors. You can't say anything like that about the GeForce, on the other hand. This card is brand-spanking new. NVIDIA went from an overstrung, small chip (like that RSX Type S K20 engine of yours) to a big, fat, beefy 16 pipe monster (think: 426 Hemi). So please, counter if you will my point that NVIDIA likely has a lot more optimization - and thus performance headroom - left than ATI. On top of that, you should know better than anyone else the logo you see whenever you start your benchmark games - The Way It's Meant to Be Played. So don't doubt for a minute that NVIDIA's horde of engineers and programmers aren't tweaking the living crap out of next-generation games. Where's ATI? Oh right, working on Half-Life 2.

Shaders galore

Page:: ( 3 / 6 )

Brandon: Jakub, don’t get me started on the shader model two versus shader model three debate. You’re the games EIC, you should know better than anyone else how few SM 2.0 titles there are on the market right now. While there are millions of SM 2.0-capable cards out on the market, there are barely a handful of titles that utilize 2.0 shaders and Painkiller isn't one of them!

Besides, half the features in shader model 3.0 are already supported in RADEON X800. In fact, they’ve been offered by ATI since the original RADEON 9700 PRO launched nearly two years ago. In some ways you could say that Microsoft and NVIDIA are just now catching up to ATI.

And sure, you’ve got support for more pixel shaders in your precious GeForce 6800, but even the best 2.0 titles that are coming out this year are only using 40-60 shaders. NVIDIA put way too much effort into optimizing for very long shader programs, a situation which won’t exist for years.

When developing their 2.0 titles, game developers are always going to program for the lowest common denominator. And the first generation “3.0” titles like Far Cry and S.T.A.L.K.E.R. will run just fine on 2.0 hardware. At best they may paste on a special 3.0 mode for GeForce 6 users that runs one or two things faster than the same effect using 2.0 shaders, but that’s about it.

The other new addition, 32-bit precision everywhere, is great for offline rendering, but this is not ILM or Pixar. Games must deliver high performance along with excellent visuals in order to immerse the user. Gaming at 10-15 frames per second isn’t an enjoyable experience, no matter how good the final product looks. NVIDIA went too far offering pie-in-the-sky paper specs that no game developer is going to push the limits with right now.

As far as Half-Life 2 is concerned, Valve’s Source engine is coming the closest to pushing the state-of-the-art. DOOM 3 was originally intended for DX7 hardware like the original GeForce 256. Shaders were only recently added, and we still don’t know how frequently they’re used.

Source, on the other hand, has shaders everywhere, not to mention high dynamic range lighting and normal maps. The Source engine is also being licensed by a number of game developers, so the results in Half-Life 2 could largely mirror the performance results in a licensees’ title. We saw that in the GeForce 6800 preview where the trends in Quake 3 also held up in Call of Duty, despite the fact that CoD uses a heavily modified version of the engine with 1.x shaders for water.

So when Gabe Newell says one architecture runs faster than another in their performance testing with Half-Life 2, you could speculate the same will occur in Vampire: The Masquerade Bloodlines or Counter-Strike 2.

Driver

Page:: ( 4 / 6 )

Brandon: And when it comes to drivers, you’re making it too easy for me. I didn’t feel the need to address your comments on driver maturity, considering all the “optimizations” that have gone into NVIDIA’s drivers recently.

What’s up with NVIDIA’s guideline: “An optimization must not contain a pre-computed state.” Then they turn around and dump shader replacement code in their newer ForceWare drivers. So I take it a hand-optimized shader isn’t considered “pre-computed state” by NVIDIA? Looks to me like they’re not even following their own stated guidelines.

Also, don’t forget NVIDIA’s use of partial precision. NVIDIA loves to promote their 32-bit precision capability, but they fall back to FP16 in many cases.

And equating NV3x to my car’s K20 engine is like comparing Godzilla to a gecko. There’s a night and day difference between NV3x’s 125 million-130 million transistors (depending on FX 5800 or FX 5900) and ATI’s 110 million transistor RADEON 9800. ATI was able to squeeze more performance per transistor out of their high-end offerings than NVIDIA, so ATI’s always been the small efficient one, just as the RSX-S’s 100 horsepower/liter is one of the highest figures in the industry for a naturally aspirated engine. If anything, that margin has only increased with this latest generation.

ATI’s got 16 pipes, just like GeForce 6800 Ultra, only they’ve been able to squeeze theirs into a smaller 160 million transistor core. In addition, their core is running at 525MHz on the X800 XT Platinum Edition. This pales in comparison to GeForce 6800 Ultra’s paltry 400MHz core.

Sure, ATI’s engine may be a little smaller, but just like the automotive industry, the size of the engine tells you nothing about the overall performance of the entire product. The real question is what is NVIDIA doing with all those extra transistors?

Besides, it’s not like ATI’s driver team hasn’t delivered performance enhancements in the past. Just read my CATALYST 4.3 report from mid-March to see the DX9 performance increases ATI’s driver team just delivered for its users. I saw double-digit performance increases in Halo. Albeit, it’s more of a synthetic test than a real-world benchmark, but I still saw some nice improvements in Tomb Raider, which is benched with a demo based on actual game play.

The fact of the matter is performance is only going to improve on both cards. To speculate on which card will improve the most is a little premature. NVIDIA has obviously had a few weeks to polish their ForceWare 60 driver since GeForce 6800 Ultra was initially launched and while there were some improvements in Far Cry I saw the exact opposite in Tomb Raider. The new driver was also far more buggy than 60.72.

You also mention NVIDIA’s The way it’s meant to be played campaign, but it’s not like these titles offer anything special. Okay, Splinter Cell got better shadows, I’ll give you that, but NVIDIA’s distance fog in Call of Duty was downright silly. It wouldn’t make any sense for a game developer to offer proprietary features for one set of users – they’d risk pissing off their customers. Those days died with Glide and the gaming world is better off because of it.

The way it’s meant to be played is marketing speak for the average consumer who knows nothing about hardware, nothing more, nothing less. Developers have had their hands on DX9 hardware from ATI longer than NVIDIA anyway, so if anything, NVIDIA’s going to need those engineers more than ATI does.

Jakub's response

Page:: ( 5 / 6 )

Jakub: Trying to drown me out in volume when you can't with quality? Alright, let's get this over with.

Oh there's few enough Shader Model 2.0 games out there, but that's changing. That'll also change for SM 3.0 even sooner, I think. The leap between the two isn't nearly as drastic, plus shaders are becoming ever easier to implement, and their importance is increasing exponentially. You saw how easy it is to combine multiple shaders in the UnrealEngine 3.0 demo. I bet those instructions pile up fast when you start stacking them together. Now while I'm sure developers will feel sorry for ATI owners and include lower-precision and lower-instruction modes, don't be surprised to see games released in the next 12-18 months busting through 24-bit precision and 1536 instruction limits.

I'm sure you're tempted to point out that 12-18 months is a long time, but computer hardware has outpaced even the most bleeding edge games by a nautical mile. Up until very recently, the GeForce 4 was still an intermediate performer and the GF3 an entry-level card. The 9700 Pro has a good 6-12 months before it becomes a handicap, and that's only because of features (the 96 shader instruction limit), not performance. Think about it, Doom 3 is still aimed at the GF3 as the minimum card. Game hardware lasts longer, and if you think people won't find themselves swapping out R420s because of 24-bit precision or 1536 instruction limits, you're dead wrong. This isn't the age of Voodoo, TNT and GeForce 1 where skipping a generation was painful. Now you can skip two comfortably.

I think you also exaggerate the 32-bit penalty, and you harp unnecessarily on mixed-mode precision. You don't see gamers complaining that a doorknob in a game doesn't have a 2048x2048 texture, and soon you won't see them complaining if an insignificant shader is done in 16-bit, as long as the important ones look right. Giving the developers a choice between what should look great and what should look mediocre is a lot better than saying everything should be average.

Don't get so easily confused about Doom 3's development. I know you hardware guys have trouble understanding, but software takes as long if not longer to develop than hardware. When Doom 3 started development, the best platform was indeed the GeForce 256 in all its DX7 glory, and Doom 3 was going to take advantage of every feature the card offered - John Carmack knowing full well that by the time development was done, cards would exist that could run all those features. Along the way, features have been added, like fragment programs (pixel shaders in DirectX speak), which "the artists have gone and put […] all over the place" in John's own words. Even if it was fast enough, a GeForce 256 doesn't have the features to run Doom 3. Half-Life 2 started development even earlier, right after the original was released, but I don't see anyone accusing it of being a DirectX 5 game.

If you actually look at it, Half-Life 2 is really the DirectX 7 title with DirectX 9 pixel shaders. Doom 3 is the game boasting per-pixel, real-time dynamic lighting, with every surface bump- and normal-mapped beyond recognition. Vertex lighting just doesn't compare, whether or not you use soft shadows on a hundred different men made of water.

What is NVIDIA doing with extra transistors? Providing the features the consumer will want - not need, I won't lie. But really, wants are so much more powerful than needs, aren't they? I need a car, but I want a sports car, even if it costs me on my insurance. So when you start getting emails from your readers who are disappointed that you recommended a card they "had to" (really, wanted to) get rid of in order to get full features in future games, don't go blaming me.

Yeah, ATI is using only 160 million transistors. Send them my congratulations, they made a great drag racing card. It'll go fast, but it won't turn. Video cards nowadays need to be flexible. As I've pointed out earlier, we have the speed. Both cards deliver that aplenty. Maybe NVIDIA will be faster upon release, maybe ATI. But NVIDIA's 6800 Ultra is the 12 second 1/4 mile Ferrari that can turn in corners and turn the ladies' heads. The X800 is the muscle car - it'll do the 1/4 in 12s, where it's pointed at already, but it won't turn to face the future. As for "480W" power supplies and power consumption, anybody whining about that is like a Ferrari owner complaining that his insurance is high and the car gets bad gas mileage.

Anyway, we all know where this war will be won. As long as NVIDIA stays even moderately close to ATI up top, the 6800 GT will kill the X800 Pro. Nobody would take a three-legged dog against one with four, no matter how fast those three legs move. The only danger NVIDIA faces in the midrange is that the 6800 GT is good enough to cannibalize Ultra sales, almost as bad as it will eat up ATI's offerings.

So peace, out. Go relieve your frustrations cow-tipping, shooting dirt with your guns, or whatever it is that you Texans do whenever you're not being obstinately wrong about computer hardware.

Brandon concludes

Page:: ( 6 / 6 )

Brandon: Drowning you out? If anyone’s drowning the other out, it’s you with that last response. I didn’t see one credible counter point to anything I just said!

Sure, Epic showed off their Unreal Engine 3.0 demo at NVIDIA’s Editor’s Day, but guess what, at GDC they showed off the same demo on ATI DX9 hardware as well. In fact, rumor is that it ran faster on ATI’s hardware than NVIDIA’s.

There’s also a night and day difference between showing off some eye candy in a tech demo and building a full game around it. By the time Epic’s next generation engine debuts in a shipping title ATI will be ready with their next generation hardware.

I also don’t doubt that SM3.0 will be picked up sooner than SM2.0, NVIDIA’s done a very good job of working with developers to get SM 3.0 implemented in their upcoming titles. But I don’t think there wilsl be a huge difference between SM 3.0 and SM2.0, especially with the first batch of SM3.0-capable titles that NVIDIA has mentioned. We’ll see what happens though, Painkiller and Far Cry were on NVIDIA’s list of SM3.0 titles, so conceivably we may have some early answers in a matter of a few months, but clearly those same titles are running just fine on X800 today.

I don’t have all the next generation titles memorized, but I don’t think you’ll see any games coming out in the next 12 months that will make the R300 family obsolete, much less X800, especially with all the even less capable GeForce FX 5200 cards floating around out there. And how can being comfortable skipping two generations support your argument that SM 3.0 is a huge oversight on ATI’s part for R420?

And I don’t feel I’m over exaggerating the performance penalties of FP32. It’s a proven fact that they’re relying on FP16 in many situations. I hate to bring up Half-Life 2 again, but don’t forget that at one point Valve was considering treating NVIDIA’s high-end GeForce FX parts as DX8 hardware because their performance was so low.

I totally fail to see your points with DOOM 3 and Half-Life 2. Yes, Carmack has raised the specs on DOOM 3, great! Valve is doing the same with Half-Life 2; they plan on implementing ATI’s 3Dc to by the way.

I also disagree with you on your point about the extra transistors in GeForce 6800 Ultra. Again, what’s the point of providing all these extra features if your hardware doesn’t have the brawn to render it all in real-time with good performance? And Jakub, since when has “good enough” performance been enough for you? You sound like some of our readers who were complaining when we said you need 100 frames per second sustained in Quake 3 for the optimal game play experience. I think you’re getting soft in your old age! That’s just a sorry excuse for finishing in second place. And power is a big issue if you’d like to build a high-end small form factor system. Say for instance, for use in your dorm room.

The GeForce 6800 GT is definitely a strong card. Of course, it was another product NVIDIA added to its lineup at the last minute to compete with ATI, but that’s another story. There are a couple of question marks surrounding the GeForce 6800 GT however. Number one is it still falls behind the X800 PRO in next-gen benchmarks, just like the 6800 Ultra (although it is a much closer battle).

Another problem with the GeForce 6800 GT is availability. Boards won’t hit shelves until mid-June at the earliest. ATI’s X800 PRO is shipping to retailers as we speak. With all of the board partners ATI has picked up in the last nine months, street prices on X800 PRO boards could be lower than the early GeForce 6800 GTs. And I’ll take your so-called three-legged dog (not like the X800 PRO is one anyway) if it’s faster and cheaper.

Finally, the GeForce 6800 GT sports the same massive NV40 die as NVIDIA’s high-end, $500 GeForce 6800 Ultra. Even though the X800 PRO is based on the same core as the XT, at 160 million transistors RADEON X800 PRO should be cheaper to produce assuming equal yields. The bottom line is ATI’s profit margins at the high-end should be higher than NVIDIA’s.

In closing, I’d just like to add that ATI’s RADEON X800 should be more than capable of handling everything the next generation of games throws at it. Sure, it may not have SM 3.0, but those first generation shader model 3 titles should run just fine on X800 in 2.0 mode. No developer is dumb enough to just abandon SM 2.0 overnight, remember that besides ATI, there are also plenty of GeForce FX cards in user's hands that developers must support. ATI learned this lesson with pixel shader 1.4 and NVIDIA will probably learn it now with SM 3.0.

I also don’t think the fact that it’s based largely on the two-year old RADEON 9700 is a detriment. Obviously the fact that it’s still around and running strong in the latest games proves that ATI had a solid architecture. With X800 ATI is building on that foundation and making it even better with new features such as temporal anti-aliasing and 3Dc, which could be just as big a deal as S3 texture compression was a few years ago when it was first introduced.

The biggest question I have is the mainstream market where the GeForce FX 5900 XT and RADEON 9600 XT are currently competing. What is ATI going to do to counter say, an 8-pipe NV40 board with a 256-bit memory interface? Assuming NVIDIA will continue its 5900 XT philosophy at the $200 price point, ATI’s really going to have to step up to the plate this time in the mainstream sector. They were too conservative with their last refresh, but at the same time I really can’t see them releasing a 4-pipe RADEON X800 with a 256-bit memory interface at the $200 price point.

NVIDIA could give ATI a real challenge here.

(Jakub's final note: Brandon gets the last word, but I'd just like to say I disagree with his assertion that the mainstream market is as significant as he states. ATI attacked the mainstream for half a decade before David Orton steered them for NVIDIA's high end segment, just as NVIDIA was losing focus by exploring professional-level cards and mainstream electronics. We all see how that turned out - the high-end rules the market, it's that simple. The two companies are too evenly matched, hardcore gamers are too ruthless in their performance and feature demands, and too many people rely on the word of mouth spread by the hardcore gamers for either company to relinquish the high end and hope to lead in the mainstream.)

Have questions or comments on our article? Post them in the forums and let us know what you think!