current/next bottleneck

[epicstruggle]

I am trying to learn more and more about 3d hardware, and had a question for you. What is/are the current bottleneck(s) in 3d hardware (ie agp, memory,...). And what do you expect will be the future bottlenecks in the future (after the release of ati 9700, nv30, p10,...).

Thanks for any and all constructive replies .

--
epicstruggle
"I'm drowning here, and you're describing the water!"

 

[Tahir2]

AGP bandwidth, CPU speed performance.

 

[demalion]

Hmm...bottleneck for what and by what measure?

For high resolution and graphic detail by the measure of fps for a 128 MB card (of which I have one) wouldn't it be videocard bandwidth? For a 64 MB card I think the AGP bus would have more of an impact, but I still think videocard bandwidth would be more important.

For lots of textures and/or shaders on top of the above, it seems to bear out that it is fill rate influenced fairly directly by GPU/VPU clock speed and architecture, but I don't think we have a game where this is the primary bottleneck yet...do we? I don't count demos and unreleased benches.

I think the overall "bottleneck" besides these is the mainstream common denominator, where we have things like slow CPUs and low memory video cards limited by the AGP bus, but for the currently AVAILABLE (even if not widespread) hardware, I think the top 2 are the real bottlenecks.

Of course, all the corrections that are bound to come in are welcome.

 

[MuFu]

AGP bandwidth, CPU speed performance.

Yes, those two spring immediately to mind as well. Another issue that has been discussed here recently is the inability of the AGP spec to provide enough power for the latest graphics cards (even in the 3.0 incarnation). Had ATi not used an external power supply for the R300, it would be limited to a core frequency 300MHz on all but reference mobos (*read Intel, VIA*). A process change will probably bring things back within spec once more, but it won't be long until boards need a 50W source one more, even with 013u fab. Still... CPU performance is the biggest bottleneck, IMO (except at very high resolutions or when using higher degrees of FSAA). Especially since the fastest cards have framebuffers large enough so that they don't have to resort to painfully slow AGP texturing.

MuFu.

 

[Clashman]

Isn't really the biggest bottleneck right now the software? I mean, there really aren't any games out there that really push current hardware. Even my vanilla GF 256 still gets along more than fine in most of the games I've come across, as long as I keep the resolution to 800 by 600. Doom III and Unreal II are probably going to change that, but realistically, the performances people are demanding out of cards right now at 1600 by 1200 with 4x FSAA, 16x Anisotropic, and 32 bit color far exceed what was considered smooth gameplay at 512X384 during the time of Rendition, Voodoo I, etc.

I think that might be a big reason why people are seeing a decline in many sectors of the computer industry. The software isn't keeping up with the hardware, thus we're not seeing people upgrading their hardware like they had to in the early to mid 90's in order to run the latest apps. And in that sense it is likely even more true outside of the gaming world, as most people are satisfied with their basic Word Processing/Email/MP3 or Video capabilities that should be fine on anything around a Pentium II or higher.

Perhaps I'm wrong in this, but as I see it most of the hardware bottlenecks being faced by the industry in general are insignificant in comparison to software.

 

[Doomtrooper]

For the AGP bandwidth issue I see the Solo AMD chipset helping that issue, just look at the traces leaving the 8151 for AGP transfers

 

[CMcK]

I have to say I agree with Clashman.

The reason I have yet to upgrade from my PIII and Kyro 2 is that there is no software which would cause me to do so.

It also seems to me on the hardware side that 3d cards are still reliant on a fast CPU and system memory bandwidth. Wouldn't it be possible to offload common tasks within a game engine to a fast maths co-processor on board the graphics cards of the future. This co-pro would be designed to run common tasks such as in game physics. Could DX and/or OGL be made to support such a capability?

 

[3dcgi]

I'll look at the question from a different perspective, similar to calling software the bottleneck. There is no single bottleneck. It depends on the test parameters. Actually in some cases even the monitor can be considered a bottleneck. Older games can run fast at 1600x1200 and maybe even higher, but most monitors can't display a higher resolution and maintain a fast refresh rate.

 

[pascal]

I agree with Clashman that software is not up to the hardware development.

Sometimes a poor software will be transformed in CPU limited games.
CPUs and GPUs today are really good, we need developers to use it elegantlly.

 

[demalion]

Well, I see the software problem being due to the "mainstream common denominator" I mentioned. For the available, but not mainstream, hardware, isn't bandwidth and fill rate bottleneck?

 

[Entropy]

I agree with Clashman that software is not up to the hardware development.

Sometimes a poor software will be transformed in CPU limited games.
CPUs and GPUs today are really good, we need developers to use it elegantlly.

The same sentiments came up in an earlier thread that was more provocatively named. I have a feeling that the populace of the B3D forum has a somewhat unrealistic view of the hardware that the general gamebuying public has access to. Valve made a survey, and while you can argue about the statistical selection, at least they have reasonably comprehensive data from over half a million people who are online gamers and were interested enough to answer.

The results are mandatory knowledge for any discussion such as this:
http://valve.speakeasy.net/

Since Half-Life is far and away the largest online shooter, it is the customer base that everyone else wants to steal. The results are somewhat confirmed by the fact that typically many of the top 10 bestselling titles are very light on the system requirements.

My point? Only in a place like this forum is it considered a benefit if the game demands the absolute latest hardware to run well. We're not representative.
For the forseeable future, (which admittedly is very short) we can expect the usual - fillrate demands are user adjustable by way of resolution, geometry and similar are adjustable to some smaller degree, by differing mechanisms.

Developers will be very reluctant to make their games require more advanced features in order to run, since that would risk cutting off too many potential customers. Which means that we might see some nice effect requiring higher level hardware, but the fundamentals of the engine will not.

I can't see this changing in any direction but for the worse since overall the rate of system upgrading seems to be slowing down. Developers shouldn't be blamed for taking the realities of their target market into account.

My candidates for top bottlenecks? CPU/memory currently, with AGP joining eventually. However, developers design around the available hardware - what is percieved as a bottleneck is to a large degree subjective and probably has mostly to do with changes in the status quo.

Entropy

 

[pascal]

Good link Entropy.
I was talking about the quality of the software.
Lets see if I can explain it because my english is terrible.

From the statistics:
50.1% have a HW T&L card (18.5% are DDR cards)
50% have a 1GHz or faster CPU
62% have 256MB or more of ram

IMHO a 1GHzCPU/256MB/HWT&L-DDRcard is very good.
My question is: are the current games exploring the maximum of this kind of configuration?

I am not talking about 2GHz P4 with GF4.

 

[TheMightyPuck]

Entropy: the technology of 3d and the business of 3d are of course forever linked. That said, arguments about margins. average selling price, oem channels, etc. wouldn't be IMHO very interesting in the long run when compared to arguments about 3d architectures, what is the best card, and what cool stuff is going to happen in the near future. As far as bottlenecks? Only our imaginations--and memory bandwidth.
(since apologies to those allergic to cheese w/r/t the imagination remark)

cph

 

[Entropy]

Entropy: the technology of 3d and the business of 3d are of course forever linked. That said, arguments about margins. average selling price, oem channels, etc. wouldn't be IMHO very interesting in the long run

Damn Right!
Particularly since it offers little hope atm.
I just felt compelled to speak up for the poor developers, who would otherwise weep from the unfairness of it all as they drive home in their Ferraris.

I did mean what I said about the bottlenecks being percieved from where we are towards where we think we are heading. For instance, I percieve the host as a big problem going forward, but that may be because of limited knowledge and ability to extrapolate. Even if I'm right, it is self adjusting and simply means that eye-candy will be added in some ways but not in others (since the load on the host would be excessive). I think consumers will appreciate the advances they do get and not percieve that there are aspects that develop more slowly.

From the statistics:
50.1% have a HW T&L card (18.5% are DDR cards)
50% have a 1GHz or faster CPU
62% have 256MB or more of ram

IMHO a 1GHzCPU/256MB/HWT&L-DDRcard is very good.
My question is: are the current games exploring the maximum of this kind of configuration?

Well, some are. Morrowind, Dungeon Siege and Neverwinter Nights are fairly recent and successful games that have that kind of configuration on its knees, even taking into account the very low demands on frame rate these games have for playability. The host on its knees, mind you - apart from the water effects in Morrowind, a GF256-DDR handles these games nicely, if you don't insist on very high resolutions. JKII, MOHAA and a bit older RtCW demand higher framerates, but can be adjusted to provide good gameplay on such a system. Not too much lower though. Again, a HWT&L-DDR card can handle these without problem at decent resolutions.

Since the topic was about bottlenecks - well if the future is to be judged by the trends and values of the past and present, then the snail-like pace of the host development will have us gnashing our teeth in frustration. It already does! One reason gfx-processing have advanced swiftly is parallellisation. Generalizing broadly, those parts of the game engines that do not lend themselves well to parallellisation on comparatively simple hardware are fine candidates to become bottlenecks.

Entropy

 

[g__day]

I would have said the bottlenecks / constraints are all of:

Complexity and lack of economic reason to develop 3D programs to run on latest generation 3D cards
Games not generally using hardware shaders - software not written
GPU speeds for full screen AA
Raw fill rates GPU <-> Video RAM bandwidth
Video memory available to hold enough textures to avoid AGP transfers

 

[Saem]

Sometimes a poor software will be transformed in CPU limited games.
CPUs and GPUs today are really good, we need developers to use it elegantlly.

I think one of the biggest problems in the industry today is the general reluctance to lisence engines. Don't get me wrong, there are quite a few games with thrid-party engines, but we need more of such engines and more people using them. With the current time constraints games have to be developed in, such things would allow for more time spent on making games that take advantage of new features. That can only be done by increasing the time it takes to make a game, or outsource it via getting third party engines. On the other side of things, third party engines need to be more accomadating in terms of what they can do and what they allow.

 

[Typedef Enum]

If you were to exclude 3D hardware-specific IQ features, such as advanced filtering/Antialiasing AND we're talking about currently shipping software, it seems to me that the dominant bottleneck has nothing to do with the 3D pipeline per-se...

There appear to be many more CPU-limited games out these days than I seem to recall "back in the day."

With that said, it's also fairly obvious that the software-underutilizing-the-hardware trend is still with us today. That's one of the reasons why, to me, evaluating an R300/NV30 @ this point in time is so very difficult...What "real" software can push such chips? In all honesty, none. Until Doom III comes out, I don't think we're going to see any shipping game which will push either...Hell, I might even argue the same holds true for the previous generation (GF4/8500).

What game(s), if any, pose performance issues with either an 8500 or GF4? I know I don't have one installed here.

 

[pascal]

...
Well, some are. Morrowind, Dungeon Siege and Neverwinter Nights are fairly recent and successful games that have that kind of configuration on its knees, even taking into account the very low demands on frame rate these games have for playability. The host on its knees, mind you - apart from the water effects in Morrowind, a GF256-DDR handles these games nicely, if you don't insist on very high resolutions. JKII, MOHAA and a bit older RtCW demand higher framerates, but can be adjusted to provide good gameplay on such a system. Not too much lower though. Again, a HWT&L-DDR card can handle these without problem at decent resolutions.
...
Entropy

Well, I mean I am expecting someone with a very good wow factor and good sustained gameplay (around 60fps) using the spec above (1GHzCPU/256MB/DDR3Dcard). Not too much, not too little and with optimized game engine. Quality with performance.

Saem,
I agree with you. Part of the solution could be better developed game engines, well documented, with tools and relativaly low cost license. This is a layman/gamer opinion.

 

[pascal]

Will the UT 2003 be the first to have a good wow factor and good performance with a mainstream computer?

Anandtech UT2003 GF2 scalling analysis: http://www.anandtech.com/video/showdoc.html?i=1650&p=5

 

[Typedef Enum]

Has anybody honestly 'tried' UT2003? Quite frankly, I'm really shocked @ how good the performance was, after all is (nearly) said and done.

To put it into perspective, compare Unreal Tournament using the D3D driver, and I swear to God, UT2003 yields better performance than UT...Of course, we all know just how _good_ the UT D3D driver was, with respect to performance.

Anyhow, in some ways, I was a little dissapointed in the fact that UT2003 doesn't really seriously stress out my system. I have no doubt this will definitely not be the case when Doom III arrives.