G80 vs R600 Part X: The Blunt & The Rich Feature

[DemoCoder]

If the ALUs are waiting, you're either TEX or bandwidth limited. If you're TEX limited, then you'd be operating at the TEX rate. If you're bandwidth limited (which probably be the case if your "latency hiding resources" are exhausted), it doesn't matter how many TEX you have. That's the most important argument against scaling up TEX as fast as ALU. The bandwidth isn't growing as fast, so adding more TEX gives diminishing returns.

That seems to be backwards reasoning. If you're bandwidth or TEX limited, it means the workloads being thrown at you require you to find a way to scale up TEX/bandwidth. You're reasoning is that GPU vendors should instead burn resources on areas where you're not currently limited, blowing ALU resources all out of proportion to the problem requirements. It may be that scaling the ALUs are *easier*, and it may be that co-issuing ALU ops while waiting for TEX can be easily performed concurrently, but just because something is easier, doesn't mean it's neccessarily the right design decision.

Imagine you're requirement is better MPG, that's the workload being demanded, and it's like you're saying that scaling vehicle MPG is harder than scaling interior features, and therefore, since you get diminishing returns trying to optimize vehicle weight, internal combustion, et al, you may as well blow a hold bunch of R&D and costs on building super-internal DVD navigation, video, and hi-def vehicle LCDs.

I love monster ALU power as much as the next guy, but every single generation, people have been raving about more ALU power and prettier pixels, and every generation we still see games getting tex/fillrate/bandwidth limited. I'm willing to agree with you that scaling up the TEX : ALU ratio is useful, but I happen to think that disproportionate scaling will yield more benefit to GPGPU workloads, physics, audio processing, etc and less 'prettier pixels' because it appears to me that the biggest problems yet to be solved on GPUs with respect to lighting also happen to be those which require alot of scatter/gather/TEX functionality.

 

[flopper]

Nobody cares now that X1900XTX is faster than 7900GTX, because that was last generation, and a new generation comes out so often in the videocard industry that you're either performing great out of the gate, or you're dead meat.

I care about that.

I rather buy a card that has a technology superiority due to games having a hard time catching up to the hardware.

Crysis, well they make that game for cards that are old really old and support win xp, a over 5 year old OS?
Talk about back logging old hardware is one thing but the new games still support old stuff due to economical concerns.

Still using my x1950xtx after a year, not in the need to get a HD2900xt yet or a Nvidia g80, I still run my games just fine. I am waiting for the next generation since that makes more sense upgrading than this 1 gen dx10.

But my buying decision are based on knowing about how software takes long to update to a level where you cant play them with enough fps.
And a card using higher technologial hindsight have my vote.

I simply cant understand why the cardmakers and developers just doesnt say, this is the hardware we need to run the game at this resolution at this fps 2 years from now and this is how the game engine are working to do that.

For me not being a technical engineer it seems to be a huge difference in how the hardware is done and how the software is programmed.

I want hardware and software that can give me lifelike effects with all the physics at a decent framerate today.
That would also make people buy games due to having them being such experience playing them. Bring me back pac man and tetris

 

[Arun]

I love monster ALU power as much as the next guy, but every single generation, people have been raving about more ALU power and prettier pixels, and every generation we still see games getting tex/fillrate/bandwidth limited.

It's probably fair to dissociate 'fillrate' (aka ROPs) and texturing, IMO. The former is *always* going to sometimes be a limitation (unless you're triangle setup limited, for example) during z-passes, shadow generation passes, etc. - you just can't really get away from that, so it's a very different case.

And yes, with more ROP power, you can nearly always make some things look prettier fairly easily. You can have more AA, potentially larger shadowmaps (but that can get expensive for the rest of the GPU if you do advanced filtering) or at least you might have more time to work on the rest of the scene, thus potentially making it prettier. However, whether that makes as much of a difference (per mm2) as more TMUs and ALUs depends on a case-per-case basis.

There is one limitation to this, of course, and that's bandwidth - as G80 highlights, however, those limitations are probably not quite as strict as most would have imagined as long as your compression algorithm is good enough. But you probably couldn't go much beyond what G80 does, ROP-wise, without embedded memory or something similar. And G80 arguably already has too much ROP power (exception: stencil isn't so hot) for its bandwidth in many (but not all!) cases.

I'm willing to agree with you that scaling up the TEX : ALU ratio is useful

With TEX : ALU, the problem is different, because you always have the two working at the same time. So all that matters is the ratio, and it will vary substantially during a single frame (IMO, that is even more important than varying from game to game). It varies from pass to pass (shadow filtrering, lighting, post-process motion blur, post-process depth of field, tonemapping etc.) and within a single pass.

The typical case is that it varies from material to material, for example the rocks in 3DMark had a longer instruction count than anything else. This also applies to particles which tend to use fewer ALU ops (but they might be ROP/Bandwidth limited anyway). Another example is determining the average luminance of a scene for tonemapping - that's mostly TMU-limited, or perhaps ROP-limited if you do it naively.

However, it's interesting that you mention lighting because that's precisely the lighting model for specific material is arguably one thing that could benefit from a lot more ALU ops. The only good example of that I could find via Google is this thread on gamedev - look at the amount of ALU operations *per light* in that Cook-Torrance shader, along with the complete lack of TEX operations used to simulate that effect. The screenshot in the last post is pretty nice too!

Another interesting case of ALU-limitations is when you have a character with very complex animation that is relatively far away, and thus the triangles are only a few pixels big. If that character is not triangle setup-limited, then that extra VS work will affect the overall TEX : ALU ratio on an unified architecture. It could be thought of as a corner case, but it definitely could happen AFAICT...

but I happen to think that disproportionate scaling will yield more benefit to GPGPU workloads, physics, audio processing, etc and less 'prettier pixels' because it appears to me that the biggest problems yet to be solved on GPUs with respect to lighting also happen to be those which require alot of scatter/gather/TEX functionality.

Well, sure - GPGPU is the field that benefits the most from truckloads of ALUs. Certain subsets of it also benefit a lot from bandwidth though, interestingly. And physics/audio on the GPU would have a very low TEX : ALU ratio too (although still quite far from zero, I suspect!)...

But I disagree completely that TEX is the primary thing that is holding graphics quality back. It remains *very* important, but ALUs are also extremely important (and in terms of lighting equation, probably more so!) and I would be incredibly surprised if the ALU/TEX ratio didn't keep growing in the 'main pass'.

However, I also think that the other passes might have a much more variable ratio, so that increasing the amount of TEX you have will nearly always improve your performance too. The real question is how much it will increase it per-mm2, compared increasing TEX... And that depends on your existing ratio, obviously. If you take the G80, it seems kind of obvious to me that increasing the ratio is the best way to improve performance. On R600, that's probably not quite the case, however!

 

[DemoCoder]

I think your example just validates my point. For most people, the improvement Cook-Torrence brings to specular is marginal, whereas the improvements from image based rendering approaches (e.g. light probes) are staggering. I think subsurface scattering would have been a better example, then again, you can't compute SSS from a vertex and normal alone.

The "pure math f(normal,vertex, attributes)" approaches leave something to be desired IMHO. There are some good fakes and tricks that people do (my favor being the y-normal sun-hacks), but I think ultimately, you've got to take the environment into account, and there you have two options: either 1) you process either local or non-local data in the shader nearby the pixel of interest or 2) you use offline preprocessing to precompute alot of the work.

The pure ALU approaches like the CT lighting model example attempt a third technique, which is *analytically preprocessing* a hypothetical generalized model of a surface over a few parameters. But, as we all learned in basic analysis, not everything has a nice, neat analytic solution. In fact, the vast majority of problems cannot be solved analytically.

 

[Arun]

I think your example just validates my point. For most people, the improvement Cook-Torrence brings to specular is marginal, whereas the improvements from image based rendering approaches (e.g. light probes) are staggering.

The improvement is only marginal because it's not the biggest problem right now. If you had a hackish way to fake indirect lighting in a hackish way (which some games are starting to do rather nicely, but it's far from perfect yet) or even compute it quite accurately, then the lighting model's impact would grow in importance. And in fact, the lighting model already was a major problem in the past that annoyed everyone: Doom 3, anyone? It has just improved enough for it be less distracting today than back then. That doesn't mean it's a solved problem.

Furthermore, since what you seem to be most interested in is indirect lighting, may I point out that per-vertex spherical harmonics requires a *LOT* of math? And yet, for static objects, it's an incredibly good way to do things. If you want to do spherical harmonics per-scene or per-object rather than per-vertex, then there are ways to preprocess that with a texture, I suppose. The two are complementary, though.

I think subsurface scattering would have been a better example, then again, you can't compute SSS from a vertex and normal alone.

You can if you're a HackMaster(TM) The way Heavenly Sword does it is just scary - it makes sense, it looks kinda good, and it's laughably cheap. I did think of SSS but didn't take it as an example for this very reason.

The "pure math f(normal,vertex, attributes)" approaches leave something to be desired IMHO. There are some good fakes and tricks that people do (my favor being the y-normal sun-hacks), but I think ultimately, you've got to take the environment into account, and there you have two options: either 1) you process either local or non-local data in the shader nearby the pixel of interest or 2) you use offline preprocessing to precompute alot of the work.

Not that many things are pure math. You still need to fetch data and probably filter it most of the time. That doesn't mean they aren't sometimes 'mostly math', though.

The point Humus makes regarding bandwidth is noteworthy. Take G80, and imagine the same architecture with the same die size is used in a console in 45nm. Let's imagine power and cost are not an issue and it actually has 4-5x more of every kind of unit. How much bandwidth do you think that chip could potentially have without embedded memory? If G80 has ~100GB/s, does it seem reasonable to assume that memory technology will have scaled fast enough to get to 500GB/s in that timeframe? I don't think so...

The pure ALU approaches like the CT lighting model example attempt a third technique, which is *analytically preprocessing* a hypothetical generalized model of a surface over a few parameters. But, as we all learned in basic analysis, not everything has a nice, neat analytic solution. In fact, the vast majority of problems cannot be solved analytically.

Uhm, I would be very interested in any alternative you propose there to obtain a similar effect with textures. Isotropic Ward can be implemented with 19 instructions instead of 51 according to Advanced Lighting and Materials with Shaders (a decent book, but sadly far from perfect imo) when using a 3D texture. The catch? It's a 2MB texture per-material... And you still have a fair bit of ALU instructions around it!

STALKER does it indexing a 3D texture too (although in a slightly different) way and interpolating between 4 possible values, and that's not a too bad idea. It's much more limited though, and requires a lot more bandwidth. And as Humus points out, bandwidth is important. Today, G80 proves that you can have a design with truckloads of TMUs and not be too horribly bandwidth limited in most cases - but the number of transistors tends to grow faster than the memory bandwidth.

In the future, even if a constant TEX/ALU ratio would allow the best possible image quality (which I doubt) in a closed environment such as a console, if that environment had no bandwidth limitations, that doesn't really matter if you *do* have limited bandwidth. In that context, you have two choices: either using embedded memory to save on some framebuffer bandwidth, or lowering the TEX : ALU ratio anyway. And if everyone in the industry does the second one, then graphics programmers will find ways to benefit from it anyway, even if it's not optimal.

So, in conclusion, if you are right and that keeping high TEX/ALU ratios would be best (which I don't believe at all, although I'm not in favor of an extreme number of ALUs per TMU either), then the problem is roughly similar to what happened with single-threaded performance IMO. Obviously, everyone including Intel and AMD would *like* to just be able to scale single-threaded performance, but that just doesn't work. The same *might* be true for TMUs, because of bandwidth constraints. The exception is if all texture fetches had insanely good cache hit ratios, or if they were all on 4bpp DXT1 textures. I don't think that's very realistic in the long-term, though!

 

[Mintmaster]

I think your example just validates my point. For most people, the improvement Cook-Torrence brings to specular is marginal, whereas the improvements from image based rendering approaches (e.g. light probes) are staggering. I think subsurface scattering would have been a better example, then again, you can't compute SSS from a vertex and normal alone.

Excellent example. Forget trying to do lots of math on a few parameters like the normal, albedo, etc. Loading precomputed data is where it's at.

The improvement is only marginal because it's not the biggest problem right now. If you had a hackish way to fake indirect lighting in a hackish way (which some games are starting to do rather nicely, but it's far from perfect yet) or even compute it quite accurately, then the lighting model's impact would grow in importance.

The problem is that the rasterization paradigm doesn't allow you do compute indirect lighting using local data from the current pixel, whether in a hackish way or otherwise. Your only option is to have more data that tells you about the surroundings and incoming light. That data is stored in textures.

Direct lighting is very easy in terms of data because you only need a few pieces of information to get a 100% complete description of the problem at hand: Light position, eye position, and a few material parameters. The assumption is a planar surface perpendicular to the normal. Indirect lighting, however, theoretically needs the entire scene's information to get a correct answer for just one pixel. No matter what hack you come up with to reduce this load, you're going to need a lot of data to approximate it.

 

[Mintmaster]

Uhm, I would be very interested in any alternative you propose there to obtain a similar effect with textures. Isotropic Ward can be implemented with 19 instructions instead of 51 according to Advanced Lighting and Materials with Shaders (a decent book, but sadly far from perfect imo) when using a 3D texture. The catch? It's a 2MB texture per-material... And you still have a fair bit of ALU instructions around it!

People overestimate the importance of an accurate lighting model for direct reflections. You could use tweaked Phong lighting with GI in an offline renderer and it would look photorealistic. If you throw in a couple of 1D texture lookups using the various dot products in standard lighting, you get an immensely flexible lighting model. On the other hand, C-T lighting without GI looks terrible in comparison.

You can't get too caught up in academia for realtime graphics. People aren't good at judging the accuracy of a model to a true BRDF. They can, however, see the impact of GI immediately.

 

[banksie]

Nvidia designs cards to perform well on today's workloads, and they do. They don't care about next year's workloads, because that will be next year's design and next year's parts. ATI doesn't seem to understand this, and they have paid the price repeatedly for their inability to focus on what's going on right now.

It is a good point but not universaly true. I actually buy ATi cards because I know they will hold up to new workloads for longer and I want a graphics card that will hold me for around two years. I simply can't afford to upgrade on a yearly basis.

Sadly I suspect the number of users like me isn't a viable business model though, so I am just quibbling a bit.

 

[AlexV]

It is a good point but not universaly true. I actually buy ATi cards because I know they will hold up to new workloads for longer and I want a graphics card that will hold me for around two years. I simply can't afford to upgrade on a yearly basis.

Sadly I suspect the number of users like me isn't a viable business model though, so I am just quibbling a bit.

No graphics card holds on for around two years. None. Not at the top. It simply slumps lower and lower, and you disable feature after feature after feature. And you realise that if you paid the dough for top-end performance and you got a lukewarm thing then, banking on future OMG WTF THE SAUCE performance isn't a good idea at all.

That is if we're talking about top-end stuff. If we're talking about getting something functional that'll run things, then yeah, sure, but then why get an expensive GPU in the first place?The satisfaction of getting one/two of these overpriced things is related to enjoying great performance in your everyday gaming endeavour, not simply waiting for months/years for a bout of wing-stretching.

 

[Humus]

Nobody cares now that X1900XTX is faster than 7900GTX, because that was last generation, and a new generation comes out so often in the videocard industry that you're either performing great out of the gate, or you're dead meat.

If you have an X1900XTX in your system I bet you care. Winning benchmarks is great and all, but people also remember what happened in the last round. If the ATI card had longer lifespan in the previous generation, the consumer will likely take that into account when choosing videocard today. I understand lots of people on this forum upgrade every generation, but the average Joe doesn't.

That seems to be backwards reasoning. If you're bandwidth or TEX limited, it means the workloads being thrown at you require you to find a way to scale up TEX/bandwidth. You're reasoning is that GPU vendors should instead burn resources on areas where you're not currently limited, blowing ALU resources all out of proportion to the problem requirements. It may be that scaling the ALUs are *easier*, and it may be that co-issuing ALU ops while waiting for TEX can be easily performed concurrently, but just because something is easier, doesn't mean it's neccessarily the right design decision.

The problem is that ATI and Nvidia don't produce memory chips. We use what's available on the market, and that sets our bandwidth. We can add various forms of intelligent memory controllers to improve utilization, but ultimately we don't get more data down than what the memory chips can feed us. Adding more TEX units will not help if there's no bandwidth to feed it. If you on the other hand spend those transistors on ALU there's a good chance that the developer can find a good use for it. Lots of TEX power gives you awesome performance at mediocre quality levels. For high quality you need more bandwidth, and traditionally that's where ATI has beated Nvidia, because we usually have more bandwidth and better memory controller.

Let's also put things in perspective. A 4x ALU:TEX ratio is not massive. Finding 4 useful ALU instructions to put on the data from a single texture fetch is hardly a huge problem. Take a simple per-pixel lighting model and you have 2 TEX and something like 15 ALU.

Imagine you're requirement is better MPG, that's the workload being demanded, and it's like you're saying that scaling vehicle MPG is harder than scaling interior features, and therefore, since you get diminishing returns trying to optimize vehicle weight, internal combustion, et al, you may as well blow a hold bunch of R&D and costs on building super-internal DVD navigation, video, and hi-def vehicle LCDs.

Works for the car industry doesn't it?
I'd liken it to trying to run your car as fast as possible. You can run your average car at 70km/h and be very efficient. Speeding it up to 100km/h and you're still doing good. But at 150km/h you're probably consuming several times as much gas, and the last squeeze to say 170km/h when you push it to the floor gives a helluva lot more gas consumption for nearly no gain. It's diminishing returns. Instead of pushing into a direction that's not working very well you should look for alternatives. Higher ALU:TEX ratio may require you to rethink your algorithms to utilize the available power better, but it's at least a way forward.

I love monster ALU power as much as the next guy, but every single generation, people have been raving about more ALU power and prettier pixels, and every generation we still see games getting tex/fillrate/bandwidth limited.

ATI has traditionally been ahead of the curve and games are always behind the curve. So perhaps 4:1 isn't the optimal ratio today, perhaps it's more like 2.5:1? I'd say the R580 is a pretty good balance today. Once all the UE3 games starts rolling out on the market I bet people that went with a G70 will not be as happy.

No graphics card holds on for around two years. None. Not at the top.

Of course it doesn't stay at the top, but cards can definitely be useful for more than two years.

 

[Blazkowicz]

probably most people that talk of graphics cards on webforums either spend a significant part of their income in their PC or have a high income. a high end every two years is very decent, most people buy midrange or low end every two years. hell FEAR was a commercial failure, not too many people could play it as it required a PC both recent and high end.

 

[tabs]

It didn't require that at all. It has a DX8 mode that runs very quickly indeed on older hardware (even for 2004), just like COD2's DX7 mode catered for that need. If FEAR was a commercial failiure then why did the publisher think it was worth releasing not one, but two expansions and the developers have got a sequel in the works?

I do agree with the point of your comment though. New high end kit every 6-12 months is far from the norm. After much reading and research I opted to buy my first ATI card, an X1900XT. One of the reasons I got it was because it seemed apparent at the time that it would have 'legs'. I'm pleased to say it does and am still getting great performance out of it now. Just played the Bioshock demo there at 1280x1024 and was very pleasantly surprised by how smooth it was.

Longevity is important to a lot of people, including those that buy high end graphics cards.

 

[ninelven]

games are always behind the curve.

Games are the curve.

 

[Bjorn]

I'd liken it to trying to run your car as fast as possible. You can run your average car at 70km/h and be very efficient. Speeding it up to 100km/h and you're still doing good. But at 150km/h you're probably consuming several times as much gas, and the last squeeze to say 170km/h when you push it to the floor gives a helluva lot more gas consumption for nearly no gain. It's diminishing returns. Instead of pushing into a direction that's not working very well you should look for alternatives. Higher ALU:TEX ratio may require you to rethink your algorithms to utilize the available power better, but it's at least a way forward.

But that's implying that there is another goal besides getting your car to run faster

And what of increasing the ALU:TEX rate will give you diminishing returns when it comes to prettier pixels ? I'm sure that developers will use the ALU power if it's there, but using it is one thing, making good use of it is another.

 

[Xmas]

If the ALUs are waiting, you're either TEX or bandwidth limited. If you're TEX limited, then you'd be operating at the TEX rate. If you're bandwidth limited (which probably be the case if your "latency hiding resources" are exhausted), it doesn't matter how many TEX you have. That's the most important argument against scaling up TEX as fast as ALU. The bandwidth isn't growing as fast, so adding more TEX gives diminishing returns.

Yet from R420 to R600 bandwidth has grown twice as fast as bilinear texturing rate.

Let's also put things in perspective. A 4x ALU:TEX ratio is not massive. Finding 4 useful ALU instructions to put on the data from a single texture fetch is hardly a huge problem. Take a simple per-pixel lighting model and you have 2 TEX and something like 15 ALU.

Given the (super)scalar design of the latest GPUs we should probably drop the habit of counting vector ALUs and say it's a 20:1 scalar ALU to bilinear TEX ratio. In reality ALU utilization is not 100% of course, but then many texture fetches are also not single cycle.

 

[chavvdarrr]

If you have an X1900XTX in your system I bet you care. Winning benchmarks is great and all, but people also remember what happened in the last round. If the ATI card had longer lifespan in the previous generation, the consumer will likely take that into account when choosing videocard today. I understand lots of people on this forum upgrade every generation, but the average Joe doesn't.

average Joe remembers that X1800 was slower and X1900 was "late, but not faster" . No way he'll know that 1 year later , now, X1900 is "better value"
Where is the reasoning?
"I buy card now, but I won't buy the fastest card atm, but instead will buy the slower one, hoping that after 1y it'll be faster."
There is no reason is such thinking

Higher ALU:TEX ratio may require you to rethink your algorithms to utilize the available power better, but it's at least a way forward.

come on, think as a game developer.
You have Vendor1 who has high Alu/Tex ratio and rethinking my algorithms is needed in order to speed up from 100 to 170, and Vendor2 who has moderate Alu/Tex ratio - without rethinking speed up will be 100 to 150 for V2 and 100 to 140 for V1. Notice that previous generation of cards from both Vendors have had moderate Alu/Tex, so without spending time and efforts old gen cards will happily move at 90, while with "new" algorithms you'll get them down to 70.
EA is pressing on you to release "in time" while cutting your budget with 20%.
Ah, alse there are these rumours that Vendor2 has way better game-dev relations in terms of help&support, add to that that V1 is late with new gen cards...


How many game-devs will "rethink" algorithms?

FF to next gen ~ 1y later, V2 release new gen cards, with higher Alu/Tex, although still not quite at the lvl set by V1 (which card was late!)
V1 promises to release "another revolutionary card with breakthrough in alu/tex" ... without specifying how late... will you rethink algorithms even more? Installed base is still largely with moderate ratios, because V1 was late and it took months before launching cards with good price/performance, not to mind the lost mindshare

ATI has traditionally been ahead of the curve and games are always behind the curve. So perhaps 4:1 isn't the optimal ratio today, perhaps it's more like 2.5:1? I'd say the R580 is a pretty good balance today. Once all the UE3 games starts rolling out on the market I bet people that went with a G70 will not be as happy.

Ha! beg your pardon, but ANYONE who had money and will to buy anything like 7300GT or better an year ago, will DEFINTELY buy NEW card when all the UE3 games start rolling out on the market.

Of course it doesn't stay at the top, but cards can definitely be useful for more than two years.[/quote]

 

[_xxx_]

Games are the curve.

LMAO! But actually you're right

 

[Silent_Buddha]

Ha! beg your pardon, but ANYONE who had money and will to buy anything like 7300GT or better an year ago, will DEFINTELY buy NEW card when all the UE3 games start rolling out on the market.

You seem to have a pretty high opinion of your own opinion. A very small portion of the OVERALL computer gaming market will upgrade their card when UE3 games start rolling out. Mostly because, well your average gamer no matter how knowledgeable they may or may not be are still on some form of a budget.

Everytime a game is released on a (at the time) next gen engine people run around forums like chickens with their heads cut off proclaiming it's time to upgrade video cards as if the sky was falling.

And everytime I see the same thing. The majority of gamers continue using what they have or upgrade to a mid-range solution. A VERY small amount of gamers (the enthusiats if you will although I don't particularly like that title) will upgrade their video card to the latest and greatest uber expensive balls to the walls video card.

And yes, quite a few gamers even try to plan out their video card purchases such that they will last at the very least 2 years. I have one such friend that just upgraded his 9700 Pro (from Spring 2003) to a 8800 GTX (Fall 2006). He doesn't plan on even THINKING about a new video card until at the very earliest Fall 2008, although he's probably going to wait until 2009.

It doesn't matter a single bit to him and to quite a few others in my local LAN gaming group what games are released between their upgrade cycles. At worst, they'll adjust graphics settings and make do.

About 10-20% of our local LAN group will upgrade yearly or sooner. Granted all of us are older (25+ with quite a few 30+) gamers with full time jobs.

Hell, quite a good portion of our group still LAN games on AGP cards in the latest games.

Regards,
SB

PS - I agree with Humus that games are generally behind the curve. Developers can only program for what's out there or what they know is coming. (UE3 is an exception as it targetted what they HOPED would be on the market when it released, and unfortunately for them CPU's have progressed slower than they predicted). Look at the release of any video card compared to the games that actually take advantage of them. DX8, DX9, DX10. They all lagged video cards, probably because it's hard to get your card optimized for a certain card until it's out, eh?

 

[AlexV]

You seem to have a pretty high opinion of your own opinion. A very small portion of the OVERALL computer gaming market will upgrade their card when UE3 games start rolling out. Mostly because, well your average gamer no matter how knowledgeable they may or may not be are still on some form of a budget.

Everytime a game is released on a (at the time) next gen engine people run around forums like chickens with their heads cut off proclaiming it's time to upgrade video cards as if the sky was falling.

And everytime I see the same thing. The majority of gamers continue using what they have or upgrade to a mid-range solution. A VERY small amount of gamers (the enthusiats if you will although I don't particularly like that title) will upgrade their video card to the latest and greatest uber expensive balls to the walls video card.

And yes, quite a few gamers even try to plan out their video card purchases such that they will last at the very least 2 years. I have one such friend that just upgraded his 9700 Pro (from Spring 2003) to a 8800 GTX (Fall 2006). He doesn't plan on even THINKING about a new video card until at the very earliest Fall 2008, although he's probably going to wait until 2009.

It doesn't matter a single bit to him and to quite a few others in my local LAN gaming group what games are released between their upgrade cycles. At worst, they'll adjust graphics settings and make do.

About 10-20% of our local LAN group will upgrade yearly or sooner. Granted all of us are older (25+ with quite a few 30+) gamers with full time jobs.

Hell, quite a good portion of our group still LAN games on AGP cards in the latest games.

Regards,
SB

PS - I agree with Humus that games are generally behind the curve. Developers can only program for what's out there or what they know is coming. (UE3 is an exception as it targetted what they HOPED would be on the market when it released, and unfortunately for them CPU's have progressed slower than they predicted). Look at the release of any video card compared to the games that actually take advantage of them. DX8, DX9, DX10. They all lagged video cards, probably because it's hard to get your card optimized for a certain card until it's out, eh?

Whilst everything you said is quite beautiful, it also doesn't really contradict what's been said prior to that. IMHO, we're not clear WRT what we're talking about here:are we talking about the high-end buyers that may actually care about high-end features/seek performance/be interested in the added math OOMPH that enables them to have everything on max/actually read anything but Amazon reviews of stuff?

Or are we talking about the budget conscious dude that is not an early adopter, but waits for a high-end part to drop a tad in price before grabbing it and holding on to it until he has to disable half of everything to enjoy decent playability?Because the second category doesn't really care that he has a high ALU ratio or whatever, because he won't mind disabling stuff all that much. And if you account for the second category, the 7900GTX is quite fine for him in competition with the 1900XT today, because he's not going to be anal-retentive like the Compubase or whatever guys and disable every possible optimization just for shits and giggles, nor will he enable max AA+max AF. I've yet to see the 1900XT owning the corresponding nV part in a scenario that doesn't have the above characteristics.

Based on what we know about UE3, it won't change that fact. Sweeney was quite outspoken WRT the fact that they've not placed an exxagerated focus on sheer math and that they make considerable use of texturing. It'll also probably be/is optimized enough for it not to matter. Nor is there any indication of focus on Dynamic-branching. And long-term usability is pish-posh anyway as a major advantage to either IHV, unless one of them produces something that sucks as badly as the NV3x line did. Which is doubtful. Since that fiasco, both have been quite safe choices for the category that upgrades at about two years and doesn't mind adjusting settings in that interval in order to maintain performance.

Of course, everything is IMHO.

 

[_xxx_]

The argumentation is bull, since the cheap bunch will also wait a year till the games land in the bargain bin, thus they'll also be ok within their own timeline.