HardOCP - 4.44GHz CPU -> VPU limited with a 425 MHz GPU

[g__day]

This is a clever job about folk using a rig with no less than 2 vapour chill units AND a water cooler for the 9700 GPU - with everything INSIDE the box!!! Mind you they take the PSU and stick it outside from the piccys.

They overclock the CPU wildy, and the GPU agressively but the real interesting part is they are say 3d tests are now once again GPU bound.

So its takes a 4.4 GHz CPU to keep feeding enough data to a o/c 9700 PRO to make it the bottle neck. Given the o/c GPU isn't expected to be that much much faster than the R350 due out in 2 months (I am guessing it will have a 400MHz core) I find this interesting. A modern 3d video card performs many functions. Many things limit it, fill rate and shader bandwidth, RAM to do really high level AA and AF, but to show it needs a CPU that is a 18 months away from what we have today to never be bottlenecked waiting for data is very interestng.

I wonder if videocard manufacturers will respond by moving ever more of a CPUs function onto the videocard itself?

http://www.hardocp.com/article.html?art=NDIy

We are holding off on the 3D benchmarks as our 9700Pro seems to be our current bottleneck. Yes, you read that correctly, the 9700 Pro seems to be keeping us "VPU limited" at 425MHz core speed. We still have a little ways to go till we get the wow factor tweaked, but we will of course revisit that at a later date.

 

[RussSchultz]

Of course, you up the AA, and the graphics card quickly becomes the bottle neck.

You up the anisotropic filtering, and the graphics card quickly becomes the bottleneck

Up the shader length, and the graphics card quickly becomes the bottleneck.

Don't believe for an instant that 9700, a GF-FX, a R350, an R400, NV35, NV36, 40, etc is too much power that a 2.0Ghz cpu couldn't make use of it in the more demanding scenes.

As one company that I've been voted spokesman for said: better pixels, not more.

 

[Doomtrooper]

I wonder if videocard manufacturers will respond by moving ever more of a CPUs function onto the videocard itself?

IMO this is a sign of things to come, some form of primitive processing onboard..

I personally think they should throw the AGP spec away as it is today..it is the weakest link in the PC platform for gaming..and I believe the video card IHV's should design and set the standard.
If you look at the current AGP spec, 8X does really nothing for performance..even if its faster transfering across the AGP bus it is dead slow on a high end card...vs onboard ram.
Underpowered, the two top dogs are forced to used external power (which I don't mind) yet why not develop a standard that is more Robust ??

Its like Intel took no input from the card manufacturers....the trend to avoid using the AGP bus (look at cards progression over the years--16,32,64,128,256 meg cards now) is a sure sign of it.
What makes up for the increasing prices of video cards (RAM)..so the AGP 3.0 spec is not helping keep the cards prices down either.

 

[Entropy]

The next step up would seem to be PCI-Express.
Architecturally speaking, it is more of the same.
Having the gfx-subsystem attached via an expansion bus is part of the PC tradition. It is neither the cheapest way, nor necessarily what offers the greatest potential overall system performance. It does however allow the modularity of system construction that has served the PC paradigm so well. Competing manufacturers drive costs down and performance up, and the modularity insulates manufacturers from some supply contstraint problems.
It works.

This architectural model is currently only challenged from the cost angle.

The time may come when we regard motherboard/CPU/Video to be a single upgradeable unit for performance reasons rather than cost. That could offer some interesting possibilities. But I seriously doubt it will happen.

The PC architecture of today is the result of a 20 year old technical/economic model. Small wonder that the PS3 looks to be much more interesting.

Entropy

 

[Ailuros]

As one company that I've been voted spokesman for said: better pixels, not more.

SiS?

*runs*

 

[Tahir2]

Didn't NVIDIA say better pixels, not more?

 

[Sharkfood]

Russ is absolutely correct. The R300 and (by all reports) the NV30 both become the bottleneck once AA and/or AF are enabled. These are the features we buy these cards for.

Even a 1.5ghz CPU is effectively VPU strapped by an OC'd 9700 Pro when using higher resolution + AA and/or AF.

It's amazing how low they have gotten the hit in this day and age, but it still remains the case and it has come at the price of reduction in quality (i.e. MS vs SS, more adaptive algorithms ,etc.etc.).

 

[Ailuros]

edit: brainfart

 

[Tahir2]

edit: brainfart

I read your original post and thought it was alright... didnt smell to me :?
To re-iterate with no AA and AF it takes a 4.4GHz CPU to bottleneck the R300.. thats like "wow man" But then whoever has a Radeon 9700 Pro and doesnt use AA or AF in some form in their games is a bit crazy (IMHO of course)... and I will not 'wink'. Oh yea, where can I get a 4.4GHz CPU that wont cost me a whole months wage?

 

[Doomtrooper]

IMO video cards have exceeded motherboard technology, if you could take that CPU and put it on a very high speed bus right on the card, that same video card would respond.

I much rather see 'platform' vs 'cpu' when talking limitations.

 

[Tahir2]

I believe an AMD processor at high clockspeeds would bottleneck the VPU even sooner than 4.4GHz. Increasing memory bandwidth between the CPU and system memory wont help too much in this case either I believe.
It has been mentioned before and just recently that the AGP bus stinks. But even with PCI Express it can't compete with 19GB+/sec of onboard memory. And I am sure the ATI Radeon 9700 Pro at 1024*768 no AA and no AF is not really utilising the AGP bus except for nominal system/application data.
It definitely is not using it for anything like for textures which I guess would have a large effect on performance.

Maybe this is a brainfart.

 

[Ailuros]

IMO video cards have exceeded motherboard technology, if you could take that CPU and put it on a very high speed bus right on the card, that same video card would respond.

I much rather see 'platform' vs 'cpu' when talking limitations.

Not necessarily a bad thing either. I hope I'm finally through next week to kill my SDRAM setup in favour of a DDR setup with the same VGA. There the difference will be huge in noAA/aniso situations, while still considerable with AA/aniso scenarios.

Additionally any current "bottlenecks" detected, can be taken for reliable if future past 4GHz CPU's will carry identical platforms as today too. Nothing remains idle.

 

[Nagorak]

I don't really feel current graphics cards are powerful enough. What is this nonsense about nothing being VPU limited? Back in the day of the Voodoo2 we ran at 640*480. If you look at that resolution everything became CPU limited years ago.

But that's not how it goes. Now we're running at 1024 minimum and with the R9700 and GFFX 1280+ with full AF and 4-6X AA on. In modern games like UT2003 we're ALREADY VPU limited in that you can't run the game at 1600*1200*32*6X*16X. I'm always worried that my R9700 doesn't have enough power (of course maybe I'm just paranoid due to having all my settings so high).

 

[jvd]

Hey i run in 640x480 with 6x fsaa and 16x aniso on my radeon and i love it .... of course thats so i can play earth and beyond on my tv while laying down.... but still....

 

[Saem]

I believe an AMD processor at high clockspeeds would bottleneck the VPU even sooner than 4.4GHz.

First of all, this is irrelevant, besides which even with all that exotic cooling, the K7 is hitting architectural limits. I don't think it'll scale in clock all that much more. In any case, it's irrelevant because if you know when the P4 will begin to limit current day VPUs, you can compose a fairly good guess as to when this time will come around that P4's and K8's will shift the bottleneck onto the VPU.

 

[Entropy]

I much rather see 'platform' vs 'cpu' when talking limitations.

Although it dates me something terrible, I often use the term 'host' to designate CPU(s)+memory+expansion bus. Everything up to the gfx-card, basically.

Migrating the gfx functions to the motherboard, (or even a MCM together with CPUs and fast local memory (acting as cache/gfx mem)) makes sense from several angles, but it breaks the paradigm.

Intel doesn't seem prepared to try to satisfy all major uses with their on-board graphics. It may be that they will offer integrated gfx in the future that spans a wider gamut of performance, but it would lead to product multiplication for their motherboard partners. An expansion bus gives the required flexibility to the platform. Drawbacks are cost, bus performance and that it cuts Intel out of part of the system.

The current state of affairs is likely to continue though, with an increased percentage of integrated gfx. The overall benefits are too great for at least another 5-10 years for gfx-cards to become an obsolete solution.


Oh, and obviously what is performance limiting, host or gfx-card, depends on what you are doing. Generalizing, playability (can I run the game? Can I attain competitive framerates?) is determined by the host and the visual quality is determined by the gfx card.

Entropy

 

[Nagorak]

There's nothing really wrong with the AGP bus anyway. Just think of it as being a nice fall back, not as something you actually want to use. It's better to get 30 fps due to texturing across the AGP bus, than it is to just have the 3D application crash on you. No matter what using memory that isn't onboard the card just isn't going to be efficient.

Even if they put a VPU socket in motherboards, it still wouldn't work out as well as in a completely proprietary card. We'd have to buy expensive memory and stuff, and then we'd just have motherboard bottleneck problems. Then we'd also have to buy exotic ram to run in the cards, and we'd probably just end up paying more overall due to the separate packaging and lack of buying power (in comparison to graphics companies that buy in volume).

I don't think anyone is interested in moving away from the AGP specification, graphics developers least of all. It would just be a total disaster. And I go back to what I said earlier: there's no such thing as CPU limited.

 

[Ailuros]

there's no such thing as CPU limited.

Of course there is. Try to pair the R300 with a 1.0Ghz CPU. Unless you want to interpret then the noAA/noAniso results compared to AA/Aniso results as "AA and aniso for free" you're going to be CPU bound since the results will be too damn close if not identical.

One notice I had in the former post I deleted is that Multisampling as found on NV20/5 and R300 has minor fillrate impacts allowing not only times higher resolutions than Supersampling, but it scales quite well with added CPU frequency too, which in certain conditions moves some limitations away from the VGA alone and splits them between the VGA and the CPU when it comes to AA.

Each accelerator requires a minimum of CPU power and will stop scaling with certain CPU frequency too. You can't say that a R300 will work optimally with a 300MHz celeron or continue scaling with a 6GHz CPU/platform of the future.

Nonetheless take a CPU with a 1500+ PRrating and a R300 in UT2003 with AA/aniso compared to a CPU with a 3000+ PRrating same settings. Getting approximately twice the performance (outside of cases where you hurt the card's bandwidth) if not more, would fall under what kind of interpretation?

edit: V2 was more like 8*6, for dual V2 setups 1*7 back then.

 

[K.I.L.E.R]

Didn't NVIDIA say better pixels, not more?

I thought JC said that?

Of course there is. Try to pair the R300 with a 1.0Ghz CPU. Unless you want to interpret then the noAA/noAniso results compared to AA/Aniso results as "AA and aniso for free" you're going to be CPU bound since the results will be too damn close if not identical.

Just make sure you are running at 1280x1024 too.

 

[Entropy]

Nagorak, you keep thinking about the system in current terms with lots of little interchangeable parts. It doesn't have to be that way. You could instead choose to think about it as a unit to be exchanged as a whole. Just as, incidentally, most institutions and many individuals do when purchasing computers. And it makes sense too, since the system after all is a coherent snapshot in time, replacing parts will typically cause them to be bottlenecked by other parts in the system. Plus software is aimed at such 'typical' systems anyway.

Going across the AGP port unnecessarily kills performance, so developers avoid it, and video card technology and gfx-engine development take directions that avoid AGP traffic. That is not an indication that the AGP is 'fine'. It would make more sense to say that the port is such a bottleneck that it dictates the evolution of graphics hardware and code on the platform.

The situation now is that the CPU processes data in its own local memory, then sends as little as possible over to the gfx card which processes it on local gfx memory. Imagine instead that the CPU and the gfx chip accessed the same memory directly, and you can see that this change in architecture could affect the way things are done. (Not to mention that 128MB of fast L3 could help some non-gfx code immensely.)

To what extent such a system would be better at this point in time is debateable. But the current paradigm doesn't allow it, and generally does not handle either integration or parallellization particularly well.

I don't think anyone is interested in moving away from the AGP specification, graphics developers least of all. It would just be a total disaster. And I go back to what I said earlier: there's no such thing as CPU limited.

Why would it be a disaster to move away from the AGP spec? We will you know, and pretty soon too. PCI-Express is just around the corner. And already half of all desktops are shipped with integrated graphics, not to mention laptops/notebooks where the industry growth is to be found.

Re: CPU limits - if you compete, and want to run UT2003 at, say, 125 fps constant in order to sync with your monitor, you need at least 250 average fps, preferably higher. Now dial down the graphics and try to get those frame rates.
Now do the same with DOOM3.
No CPU limits? Yeah, right.

Entropy