AMD(ATI) vs Nvidia 2007

[guest.r]

Originally Posted by ChrisRay

Seriously.. what the heck did you just say guest?

If you enable multisampling on both 2900XT and 8800GTX the second simply produces more samples and is faster. Games mostly use multisampling AA theese days and most reviewers let the app. decide the tweaks and settings. So the 8800GTX has some sort of advantage (like 20%).

Recent gfx. adapter generations had a preety linear pixel pipeline, the first part did texturing and rastering, the second some shading calculations, FYI.

[ChrisRay]

The HD2900XT and Geforce 8800GTX support the same multisample capabilities as far as maximum goes.

2x, 4x, 8x multisampling.

Nvidia has 8x CSAA ((4x Multisample 4x CSAA))

16x CSAA (( 4x multisample 12x CSAA))

16xQ CSAA ((8x multisample 8x CSAA))

And ATI merges its 2x/4x/8x with nearby pixels to do its filter AA and edge detect modes.

[guest.r]

2900XT should have a theorethical advantage with filter AA (compared with if-then filter AA by nvidia) cause it can read the data faster and write at the same time (512bit read/512bit write). Those feats simply matter with shader ops..I guess the optimal AA for this card is therefore 2xMSAA & FilterAA or simply just FilterAA.

[fivefeet8]

Bioshock benchmarked with newest drivers from both camps. It includes the x2900xt 1gig for those that care.

BioShock DirectX 10 vs DirectX 9 Performance and Image Quality

[Cid Highwind]

Well for once the results are surprising in many ways:

1. The ATi HD2900XT performs much worse than expected
2. The game runs fine on an X1950 Pro
3. Enabling DX10 mode finally manages not to halve your fps

Still, I'm more interested in seeing how the midrange cards perform in this game, mainly the HD 2600XT and Pro, and the 8600 series.

Btw, I'm glad I went with ATi last gen, this AGP card definitely doesn't disappoint me in this game, and perhaps the 512mb might even help as well when I get a bigger monitor.

[ian209]

Hmmm, i hope future drivers like 9.0 gives at least 200% performance boosts; cause according to the specs, the HD 2900XT Should pound the 8800GTX.

[fivefeet8]

Originally Posted by ian209

Hmmm, i hope future drivers like 9.0 gives at least 200% performance boosts; cause according to the specs, the HD 2900XT Should pound the 8800GTX.

Why should a $400 card be faster than a $500 card? There's no mystery or question why the x2900xt was priced the way it is. The specs list shown by both IHV's are more marketing than anything else. AMD's tech specs are especially prone to confusion because of the way they count their shaders and some of their architectural design choices.

[fivefeet8]

Some ET: Quake Wars Demo performance comparisons from various cards with the absolute newest drivers from both camps.

Enemy Territory: Quake Wars Demo Performance

Interesting OpenGL benches.

[guest.r]

Originally Posted by fivefeet8

AMD's tech specs are especially prone to confusion because of the way they count their shaders and some of their architectural design choices.

If i get things right the Ati gpu is capable of processing 320/5 = 64 fragments at the same time. 5 shader processors are active for a single fragment and the HW/compiler logic tries to reduce idlation to practic zero.

Dunno how much calculating power is acctualy lost cause the task-finished 5-cluster doesen't help the neighbour. It fetches a new thread (fragment task) instead. Sound fair and reasonable.

[tuanming]

The currently DX10 gpus suck in DX10 games... we need something much more powerful (2x and up) Speaking of DX10, I don't see anything impressive about it, bring out dx11 already!

[ChankastRules]

The currently DX10 gpus suck in DX10 games... we need something much more powerful (2x and up) Speaking of DX10, I don't see anything impressive about it, bring out dx11 already!

You don't think geometry shader, bigger textures, new virtual memory model, better soft shadows and other lighting improvements, better performance, and more possibilities with light sources is not impressive at all? hehe you are quite high-demanding!

[fivefeet8]

Originally Posted by tuanming

The currently DX10 gpus suck in DX10 games... we need something much more powerful (2x and up) Speaking of DX10, I don't see anything impressive about it, bring out dx11 already!

My 8800gtx runs Bioshock in Dx10 mode quite well.

Originally Posted by guest.r

If i get things right the Ati gpu is capable of processing 320/5 = 64 fragments at the same time. 5 shader processors are active for a single fragment and the HW/compiler logic tries to reduce idlation to practic zero.

The problem arises from dependant math operations that can stall these units from actually using all 5 ALu's. The compiler is suppose to schedule and rearrange these operations to make each unit process 5 operations, but this is very complicated without a lot of compiler optimizations and sometimes even shaders coded to optimize for it.

Contrast this with the G8x architecture which flies through dependant math operations and the fact that most newer games have been developed using G8x hardware and it becomes apparant why there is so much variancy in the R600's performance with new games.

[ian209]

ATI's HD 2900Pro is soon to release. It's about $350USD and from what i've seen, it beats the 8800GTS 320MB in most cases and comes close to the 8800GTS 640MB

Originally Posted by tuanming

The currently DX10 gpus suck in DX10 games... we need something much more powerful (2x and up) Speaking of DX10, I don't see anything impressive about it, bring out dx11 already!

Hmmm, my 2900XT can run Bioshock w/ full settings at 1280*1024 with over 30FPS, which isn't too bad..

[Godwin]

Newegg.com - SAPPHIRE 100212SR Radeon HD 2900PRO 512MB 512-bit GDDR3 PCI Express x16 HDCP Ready CrossFire Supported Video Card - Retail

Newegg.com - SAPPHIRE 100214SR Radeon HD 2900PRO 1GB 512-bit GDDR4 PCI Express x16 HDCP Ready CrossFire Supported Video Card - Retail

Those don't look like $350

[tuanming]

Originally Posted by ChankastRules

You don't think geometry shader, bigger textures, new virtual memory model, better soft shadows and other lighting improvements, better performance, and more possibilities with light sources is not impressive at all? hehe you are quite high-demanding!

What you listed, I have yet to see any major changes/differences in dx9 vs. dx10. Demanding in the tech world is a very good thing

See it for yourself DriverHeaven.net: Bioshock Performance

HotHardware - DX9 vs. DX10 with Lost Planet

As of right now dx10 is still in it infancy...I mean the visual part is 98/99 percents no differ from dx9... Not to mention, it's slower than dx9.

[ChankastRules]

It took one year before we started to see changes between DX8 and DX9, suddenly people think we'll see the effect of DX10 within a few months... have some patience :P it is still going faster than the previous transitions.

[ian209]

Originally Posted by Godwin

Newegg.com - SAPPHIRE 100212SR Radeon HD 2900PRO 512MB 512-bit GDDR3 PCI Express x16 HDCP Ready CrossFire Supported Video Card - Retail

Newegg.com - SAPPHIRE 100214SR Radeon HD 2900PRO 1GB 512-bit GDDR4 PCI Express x16 HDCP Ready CrossFire Supported Video Card - Retail

Those don't look like $350

lol holy ****, thats insane!!!! im gonna buy two and go crossfire

[guest.r]

Originally Posted by fivefeet8

The problem arises from dependant math operations that can stall these units from actually using all 5 ALu's. The compiler is suppose to schedule and rearrange these operations to make each unit process 5 operations, but this is very complicated without a lot of compiler optimizations and sometimes even shaders coded to optimize for it.

Contrast this with the G8x architecture which flies through dependant math operations and the fact that most newer games have been developed using G8x hardware and it becomes apparant why there is so much variancy in the R600's performance with new games.

I also imagine 4 units doing some vec4 ops in a row while the 5th does sine/log/div/sqrt...
The G8x shaders are cool since they cost the gpu less transistors. Rescheduling isn't for free and it cost's cycles (more then 0 for sure) then stalls can occour simultaneously since most of the fragments "carry" the same code and come out of pipes at the same time. Can't go around that.

Originally Posted by various authors

DX10 is slow.

DX 10 is a nice API collection but the game authors want to try out different new candies and this slows things down (the number of calculations the gpu must make increases, bottlenecking occours more often etc.).
The dx and OGL drivers were built and optimized over a decade long period so they are sort of as fast as they can get.

[fivefeet8]

Originally Posted by tuanming

The currently DX10 gpus suck in DX10 games... we need something much more powerful (2x and up) Speaking of DX10, I don't see anything impressive about it, bring out dx11 already!

What you listed, I have yet to see any major changes/differences in dx9 vs. dx10. Demanding in the tech world is a very good thing
As of right now dx10 is still in it infancy...I mean the visual part is 98/99 percents no differ from dx9... Not to mention, it's slower than dx9.

Some of the differences in current Dx10 rendering modes on current games are not that significant, but are still rather nice and welcomed. I'd much rather see small changes than none at all. Developers will optimize and enhance such effects in later revisions and games as they become more accustomed to the new API and because Dx10 allows them more freedom to do certain things in more different ways.

Here's a great article on Dx10 effects and how they were done in Lost Planet with the latest patch. It also illustrates that shadow rendering performance has actually increased with the Dx10 path while making the rendering a little better. It also shows some rendering comparisons from different effects between the 2 modes. The differences between the motion blur is quite remarkable.
Beyond3D - CEDEC 2007: Capcom on Lost Planet Part II

Performance is arguable since Bioshock and Lost Planet(latest patch) perform very similary between their Dx10/Dx9 rendering paths but with better effects in Dx10. Basing assumptions about Dx10 on current games is a bit premature. Kind of like basing performance of a video card purely on tech specs alone.