The Fermi Thread - Part 3

[[XC] Oj101]

........

OMG I LAUGHED HARD!!! I''ve got tea all over myself

[Gorguls]

[Manicdan]

I guess that sort of answers my question, but it really doesn't.

In order for us to determine the true benefit of tessellation over standard polygon rendering is give me a scenario with similar complexity one done in tessellation the other using old school methods. Otherwise you aren't really making a point.

That's like saying "look how slow this car goes without turbo vs this other similar, but a 1000 pound lighter, car with a turbo!". The inverse of that analogy might make more sense but I think both apply the same meaning: not a fair comparison.

its alot more than just comparing how the GPU handles alot of polygons, vs tessellated polys, the designers apply a single map for the tessellation on top of the old model, which gives the game the ability to do unique things. theres a nice crab movie where it starts off basic, then they move a slider and it gets a whole bunch of spikes that just grow, all in real time. its not easy to do that with a simple model since the polys which are only part of the displacement map cannot be individually identified.

but in a nutshell, the polygons from tessellation are added later in the rendering process which can help reduce resource requirements. and a bunch of other bonuses only game developers really give a whoot about

[zerazax]

Yeah, I love how the GTX285 can run in DX11. And then you look at the lower resolution, and it is faster than the 5870


Talk about a bad review. Most likely, the Nvidia cards defaulted to DX9 hence the score

Look at techreport's (a much more reliable site anyways) own benches on the matter

http://techreport.com/forums/viewtop...&sd=a#p1009567

[SimBy]

Can someone with HD5870 confirm this. FPS looks kinda low here.

And since when is GTX285 DX11?

[Motiv]

Can someone with HD5870 confirm this. FPS looks kinda low here.

physx....

[zerazax]

2 (or 3?) hours to go!

[sideeffect]

ATI has a driver issue running 4 x AA in Metro 2033 which will be fixed soon i'm sure. The testing and optimisations for the game were done on Fermi cards so ATI needs some time to improve the performance and fix any bugs.

Advanced Physx on or off has little effect on the FPS in Metro 2033 because the game is GPU limited and also supports multicore CPU physx. On my system I have no performance penalty when enabling advanced Physx. Nvidia cards have also been shown to run faster using the CPU to run advanced physx and letting the GPU run the game.

[jaredpace]

anyone save the die shot from hexus review?

[japamd]

Based on Hexus' results, that's a quick peek at performance, all test at 1920x1200 AA 4x, except Dirt 2/AA 8x/DX 11 at 2560x1600

We need more data

[Chumbucket843]

I guess that sort of answers my question, but it really doesn't.

In order for us to determine the true benefit of tessellation over standard polygon rendering is give me a scenario with similar complexity one done in tessellation the other using old school methods. Otherwise you aren't really making a point.

That's like saying "look how slow this car goes without turbo vs this other similar, but a 1000 pound lighter, car with a turbo!". The inverse of that analogy might make more sense but I think both apply the same meaning: not a fair comparison.

a 1 million poly scene (average for games today)takes about 20gflops at 60fps for geometry. the bottleneck is memory capacity and in some cases bandwidth. tessellation increases polys while using much less memory, about 10x less. it also keeps data on chip so it uses less gddr bandwidth.

That depends on how tesselation is done in relation to the rest of the card.... which will be based on architecture

Yes yes DX11 was slapped on for Dirt2 and other DX11 games, but the Hexus benches showed the card falling down in DX11 so I'm not sure how to explain that at all

dx11 games were developed with 5xxx cards so they havnt really done much debugging or optimizations for nvidia. stalker should favor ATi with tessellation because it has used non-dx11 tess for a while. keep in mind none of those games have heavy tessellation yet, even metro 2033 doesnt. it will be nice to see who can get the first good dx11 title out (graphics wise).

[Motiv]

Based on Hexus' results, that's a quick peek at performance, all test at 1920x1200 AA 4x, except Dirt 2/AA 8x/DX 11 at 2560x1600

We need more data

did you catch the other site?

[hurleybird]

Don't forget that is system power consumption, so the 146.3% is deceptive. Power consumption of only the GFX is probably more like 190-200%

[zerazax]

Based on Hexus' results, that's a quick peek at performance, all test at 1920x1200 AA 4x, except Dirt 2/AA 8x/DX 11 at 2560x1600

We need more data

For a second I saw the power consumption (load) and thought... wow 48% more performance! What game was that!... then I facepalmed

Thanks for the chart though... its much faster in DX9 DirT2 and HAWX and Far Cry 2 (expected), but slower in Crysis Warhead by a little and just a bit faster in DiRT2 DX11 and BFBC2...

[WrigleyVillain]

I just bought a Sapphire 5850 from Newegg for $279 shipped.

[zerazax]

I just bought a Sapphire 5850 from Newegg for $279 shipped.

Did prices just drop? I thoght they were still running at $300ish

[Olivon]

Supersonic Sled Demo is up on nVidia site.

Run only on GT4X0 Series

[Manicdan]

Did prices just drop? I thoght they were still running at $300ish

they are, most are 310-330, that one for some reason is 30$ cheaper. but has only a few reviews and they are not very high.

[WrigleyVillain]

Yeah would have liked to see more reviews but it's a Sapphire. I'm confident it will be fine. And prices did just drop somewhat this morning but I don't expect that to last now.

Mine dropped $20 from when I first looked at like 10AM CST too. Then the cheapest was $305 w/ free shipping.

[tajoh111]

Power consumption is piss poor. Its pretty hard to ignore, I don't know how they can consume so much power with a single chip. Powerwise this is as much of a disaster as the pentiumr 4 D editions. These things look like they are going to bench like barncats under LN2, if they are still clocking 15% very warm conditions.

This generation the 5850 is the card to get for price to performance and watt.

The delta between the the 5850 and the 5870 is smaller than the delta between the 5870 and Fermi. Considering pricing, the 5850 has a 120 price difference compared to fermi and the 5870 100 dollar price difference. If you don't mind Cf solutions, the 5850 is a combo pretty hard to ignore if you want to spend about 550 dollars.

I can't believe people are dismissing driver potential here, atleast team Red fan. This card is brand spanking new architecture and it shows with the up and down performance. The current drivers are so specialized for the gtx 28x - g80 its not even funny. The very same people that were saying 5870 didn't perform that well at launch because of drivers are the same ones dismissing fermi driver potential.

The fact is fermi biggest improvements are in games that they performed poorly in prior generations in the past which shows how new the architecture is. Considering that these chips have only had 4 months for driver development, I think it unreasonable to think fermi won't pick up big driver gains in the future.

I agree with you annihilator that future driver improvements should never be part of the purchasing equation. None the less, I think its impossible for fermi not to pick up huge jumps in the future.

[annihilat0r]

2 hrs 45 minutes remaining until DH review.

[Manicdan]

Yeah would have liked to see more reviews but it's a Sapphire. I'm confident it will be fine. And prices did just drop somewhat this morning but I don't expect that to last now.

Mine dropped $20 from when I first looked at like 10AM CST too. Then the cheapest was $305 w/ free shipping.

i figured it out, the other saphire card is 25$ higher, but come with CoD, so both cards are a good 40$ cheaper than the competition, which is a nice bonus. but in my clicking around i saw a 514$ 5870 with a water block, which i think is the way i would go. OC the crap out of it until it consumes as much power as the 480, and see who can win then.

[zerazax]

I can't believe people are dismissing driver potential here, atleast team Red fan. This card is brand spanking new architecture and it shows with the up and down performance. The current drivers are so specialized for the gtx 28x - g80 its not even funny. The very same people that were saying 5870 didn't perform that well at launch because of drivers are the same ones dismissing fermi driver potential.

The fact is fermi biggest improvements are in games that they performed poorly in prior generations in the past which shows how new the architecture is. Considering that these chips have only had 4 months for driver development, I think it unreasonable to think fermi won't pick up big driver gains in the future.

I agree with you annihilator that future driver improvements should never be part of the purchasing equation. None the less, I think its impossible for fermi not to pick up huge jumps in the future.

The question actually is how much is it that the architecture is influencing performance in those said games?

As B3D users have pointed out, Fermi might have changed how the card does setup and tri/clock, which is corroborated in its performance in HAWX (a triangle intensive game). In traditional shader/texture games, like BFBC2 and Crysis, it doesn't seem to do all that much better than previous generations - then again, with lower ROPs and same TMU's as a GTX 285, that might be why it has limited gains.

So my point is that "up and down" performance isn't indicative of drivers when the other variable, a different architecture, is in play - if Fermi improves triangle rate, then it's not drivers at play - its architecture taking advantage of a game's engine for example.

We'll need the full review of the architecture to see whats at play here, but that's probably the biggest reason there IS a performance delta that big between games

[Loque]

furby or 5990...

hmm besides 5990 costing more I guess the only other con for it is being dual gpu, all rest is better than furby.. *undecided* x_x

[jaredpace]

they are, most are 310-330, that one for some reason is 30$ cheaper. but has only a few reviews and they are not very high.

It's crazy that 6 months ago that sapphire card for $329 with 220 reviews was just $259 on 10/2/2009. Half a year goes by, and the price rises $70.
Go figure. Compare the 5850 to the GTX280. Debuts at $659 and 6 months later, the price has dropped by $300 to $359