The AMD Mantle Thread

[GaiaHunter]

I think you misunderstand the meaning of being CPU bottlenecked. Increasing resolution leads to being more GPU limited, and lowering it leads to being more CPU limited generally speaking.

The GTX 780 is CPU limited @ 1080p in the Civ 5 benchmark because it's MAXING IT OUT. Basically, the frame rate can't go any higher. Increasing the resolution to 1440p makes it GPU limited again.

I think I posted the wrong graph initially. I should have posted the 1080p graph from the get go..

You mean that CiV5 is capped at 138 fps?

http://www.hardware.fr/articles/887-18/benchmark-civilization-v.html



Might be close but not right there.

 

[ASM-coder]

Increasing resolution leads to being more GPU limited, and lowering it leads to being more CPU limited generally speaking.

I agree with the 1st part(in part,) but how do you figure on the 2nd part?

If you are already CPU limited, you probably can't become GPU limited by
increasing the resolution. The CPU is limiting that.
If you are not CPU limited at a higher resolution,
you would not become CPU limited by lowering the resolution.

 

[ASM-coder]

Again, don't misquote me. Don't try to make it sound like I said something that I didn't. If you want to argue facts, have at it. Don't try to twist and re-arrange my words into something else. Doing so makes it seem like your intentions are nefarious or you're trying to spread FUD. Again, i'd like to think that you aren't doing that. Hopefully, your mis-quotation was a mere accident. Perhaps double check before hitting the post button. Thank you.

Did you not say: "I understand that Mantle helps hide CPU bottlenecks..."?

 

[sushiwarrior]

I think you misunderstand the meaning of being CPU bottlenecked. Increasing resolution leads to being more GPU limited, and lowering it leads to being more CPU limited generally speaking.

The GTX 780 is CPU limited @ 1080p in the Civ 5 benchmark because it's MAXING IT OUT. Basically, the frame rate can't go any higher. Increasing the resolution to 1440p makes it GPU limited again.

I think I posted the wrong graph initially. I should have posted the 1080p graph from the get go..

No the GTX 780 is not maxing anything out, it is capable of higher FPS at lower resolutions and/or MSAA levels. That is a clear indication of being GPU limited, rather than CPU limited.

Also, this is a single game title, so whatever case you're trying to make about Nvidia making significantly better multithreaded drivers would need more than a single title to back it up...

 

[realibrad]

Most games are not CPU bound in single player. Multiplayer games load the cpu more, because it has to keep track and do more calculations. When a game is not CPU bound, it will become GPU bound. If you have a game where you are gpu bound at 60fps, and you get into multi player where you run out of cpu overhead compacity, you will see frames drop. This is because the cpu cant supply the gpu with enough data to bottleneck the gpu.

Mantle will help reduce some of the overhead and the cpu can do less work. Because of this, the gpu will get as much data as it can take. Lowering the resolution, will mean that the cpu is sending very little to the gpu to crunch. If you were not cpu bound at high res, you wont see a drop in fps by going down res.

What is meant when people say you are cpu bound in low res, is that your cpu will be the cap to fps at low resolutions. When the resolution is low enough, the data sent to the gpu is simpler and thus wont be the bottle neck.

If you have 2 people working on a task, and the first person can do all the work, the 2nd wont have anything to do. So until you give the people more work to do, the 1st person is the bottleneck.

 

[ASM-coder]

If you have 2 people working on a task, and the first person can do all the work, the 2nd wont have anything to do. So until you give the people more work to do, the 1st person is the bottleneck.

Unless you bypass the 1st person and give the work directly to the 2nd person,
which is what Mantle tries to do.

A decent enough analogy(I assume that was your intent.)

 

[realibrad]

It depends.

If you are playing a game at 480, then the main cap of fps will be the cpu. Thats because most of the work done is not rendering. You must go through the first person to get the work to the 2nd, and as far as I understand mantle does not bypass the first and offload the work onto the 2nd. Mantle can reduce the total work at a given constant of settings. By reducing the work the cpu has to do, it can then send more data to the gpu to process, and thus the gpu can work on outputting more frames. The CPU feeds the gpu, and if the gpu cant be fed lower fps. Mantle allows the cpu to do less work on other things, and thus can send more food to the gpu, higher fps.

 

[bystander36]

I agree with the 1st part(in part,) but how do you figure on the 2nd part?

If you are already CPU limited, you probably can't become GPU limited by
increasing the resolution. The CPU is limiting that.
If you are not CPU limited at a higher resolution,
you would not become CPU limited by lowering the resolution.

Actually, in both cases you can, but it is not a certainty either.

Increasing the resolution, puts more load on the GPU and lowers your FPS. If the load is enough, it can shift the bottleneck from the CPU to the GPU.

If you are GPU bound at a high resolution, lowering the resolution can increase FPS enough to shift the bottleneck from the GPU to the CPU.

Neither are certainties, in either case, you may be a long ways away from a bottleneck shift, but it certainly happens. You'll notice when review sites want to review a CPU, they pick the fastest GPU and use the lowest resolutions possible, in order to put the bottleneck on the CPU if it is possible. When reviewing GPU's, they like to use higher resolutions and as fast of a CPU as possible to compare, in order to put the bottleneck on the GPU.

Generally speaking, the CPU becomes bottlenecked by hitting specific FPS. The CPU's primary hold up is feeding the GPU, and the resolution does not effect the CPU's job. The GPU's is bottlenecked by giving it more work to do and that can be done with IQ settings, resolution and AA.

 

[ASM-coder]

It depends.
You must go through the first person to get the work to the 2nd, and as far as I understand mantle does not bypass the first and offload the work onto the 2nd.

Ah, Ok. So perhaps the 1st person gets more detailed instructions so that he has less to figure out, and passes it more quickly to the 2nd.
I was thinking that Mantle allowed you to leverage more of the processing
power of the GPU by bypassing the CPU for some work.

 

[realibrad]

Just to make it a little clearer, raising resolution increases the total work to be done. Look at it like this:

at 480, there are 10 units of work to be done by the cpu and 5 units of work for the gpu.
at 720, there are 12 units of work for the cpu, and 10 units for the gpu.
at 1080, 15 units for cpu, 20 for gpu.

If you cpu can handle 15 units of work, then the bottleneck becomes the gpu, regardless of how many potential units of work the gpu can do. If you follow the scale backward, then at lower resolutions, the gpu has to do so little work, that the fps is capped by the cpu.

in this example, dont confuse fps with units of work, because each frame at a different resolution is very different amounts of work. A single frame at 480 is only 2 units of work, where as a 720 would be 4 units of work. So as you scale up the resolution, your cpu has to do a little bit more work, but the gpu must do a lot more work.

What amd did was to make the units of work the cpu must do lower, but the total units of work to be done for each frame for the gpu remains the same. Its just that now the cpu has less work to do, it can work on sending more work to the gpu. Or, to keep with the working people, the 1st person no longer has to make coffee and can focus on its work and complete it faster.

 

[Gloomy]

theres also more than one kind of gpu bottleneck

 

[bystander36]

Just to make it a little clearer, raising resolution increases the total work to be done. Look at it like this:

at 480, there are 10 units of work to be done by the cpu and 5 units of work for the gpu.
at 720, there are 12 units of work for the cpu, and 10 units for the gpu.
at 1080, 15 units for cpu, 20 for gpu.

If you cpu can handle 15 units of work, then the bottleneck becomes the gpu, regardless of how many potential units of work the gpu can do. If you follow the scale backward, then at lower resolutions, the gpu has to do so little work, that the fps is capped by the cpu.

in this example, dont confuse fps with units of work, because each frame at a different resolution is very different amounts of work. A single frame at 480 is only 2 units of work, where as a 720 would be 4 units of work. So as you scale up the resolution, your cpu has to do a little bit more work, but the gpu must do a lot more work.

What amd did was to make the units of work the cpu must do lower, but the total units of work to be done for each frame for the gpu remains the same. Its just that now the cpu has less work to do, it can work on sending more work to the gpu. Or, to keep with the working people, the 1st person no longer has to make coffee and can focus on its work and complete it faster.

Increasing the resolution does not add work to the CPU. Only the GPU's work increases when increasing the resolution.

 

[realibrad]

I was trying to illustrate a point, but I guess I could have just left the cpu unit work constant, as resolution was the only thing I was talking about increasing.

 

[ICDP]

Increasing the resolution does not add work to the CPU. Only the GPU's work increases when increasing the resolution.

Surely the CPU load does increase by a marginal amount? After all the load on the HD, SATA, RAM, etc will increase due to much higher requirements on data throughput. The CPU controls these components and I assume has to work harder as a result?

 

[Haserath]

Increasing the resolution does not add work to the CPU. Only the GPU's work increases when increasing the resolution.

Sure it does. The CPU has to draw the frame and the GPU has to fill it in. It doesn't affect the CPU much but it's there.

 

[bystander36]

If there is a difference, it is miniscule.

You see it all the time when people are severely bottlenecked. They can play at low, medium and high with absolutely no difference in FPS.

I also am not sure the CPU really draws a frame, but rather calculates all the objects, but the number of data points remains constant. Then the GPU takes over and fills it in, which does increase the work for the GPU.

Example on a review of Neverwinter:

 

[ICDP]

It may be small but it is still there. While it might be nitpicking to say your statement was wrong it is important to differentiate between right and almost right.

 

[bystander36]

It may be small but it is still there. While it might be nitpicking to say your statement was wrong it is important to differentiate between right and almost right.

Again, I'm not even sure there is a difference in most cases. Though I have not coded for game related task since before the 3D accelerator, so I guess I may be wrong, so someone with current info, feel free to correct me...

The CPU doesn't actually draw on the screen, but rather calculates a fix number of points for polygons and what not. Whether it is at 720p or 4k, the number of points to make up the polygons does not change. They just calculate larger numbers and hand it off to the GPU.

 

[Paul98]

Surely the CPU load does increase by a marginal amount? After all the load on the HD, SATA, RAM, etc will increase due to much higher requirements on data throughput. The CPU controls these components and I assume has to work harder as a result?

Unless the engine changes LOD based on resolution both resolutions will have the same vertices sent to the GPU.

 

[itsmydamnation]

Windows 7 once again.

It doesn't look like they used Windows 8.1..

The PClab.pl numbers mirror those of some of the other reviews, particularly that Mantle gives greater benefits for lower performing CPUs, and that NVidia can catch up and even exceed Mantle's performance provided certain conditions are met:

1) The use of Windows 8/8.1

2) A powerful Intel Core series processor is used.

so?????? look at the it from a perspective of relativity. lots of data aligns except for the highend mantle data. so please explain how thats "windows 7 once again"?

the point is on PCLAB's mantle stop scaling on the high end and on everything else it keep scaling. PClab is then being used to say on the highend CPU mantle benifit is limited, but that part of PClabs data doesn't align to anyone else...........

 

[Carfax83]

You mean that CiV5 is capped at 138 fps?

http://www.hardware.fr/articles/887-18/benchmark-civilization-v.html



Might be close but not right there.

I translated the section with Google, and this is what it said:

This second use of DirectX 11 does not concern us here, however, since we use the integrated card game benchmark we get zoomed slightly to reduce the CPU limitation that is very strong in this game

So it appears they zoomed in to reduce the CPU limitation in the benchmark. Basically, zooming in means less of the map is shown, which would reduce pressure on the CPU as there are less objects.

Using the standard view though, it appears there is a CPU limitation (I thought it was a frame rate limit), as a Titan, GTX 690 and the GTX 780 hit a frame rate wall.

 

[Carfax83]

I agree with the 1st part(in part,) but how do you figure on the 2nd part?

If you are already CPU limited, you probably can't become GPU limited by
increasing the resolution. The CPU is limiting that.
If you are not CPU limited at a higher resolution,
you would not become CPU limited by lowering the resolution.

From what I understand, the lower the resolution and graphical settings, the faster the GPU is able to draw frames. So the faster the GPU can work, the faster the CPU has to work as well to provide it with the data to draw those frames.

That's why performance is more reliant on the CPU at lower settings, as that's where the bottleneck is.

But as the settings go up and more of the burden is shifted to the GPU, the CPU doesn't have to work as fast to provide data for the GPU as the GPU is taking longer to draw each frame.

So yes, you can become CPU limited at lower settings, and then become completely GPU limited at higher settings. Civ 5 shows that in fact..

 

[Carfax83]

No the GTX 780 is not maxing anything out, it is capable of higher FPS at lower resolutions and/or MSAA levels. That is a clear indication of being GPU limited, rather than CPU limited.

Also, this is a single game title, so whatever case you're trying to make about Nvidia making significantly better multithreaded drivers would need more than a single title to back it up...

Then how do you explain this:

The GTX 780, Titan and the GTX 680 SLI all have very similar performance, which implies they've run into a wall.

And the hardware.fr article suggested they had to "zoom in" to reduce the CPU limitation to get higher frame rates.

 

[Carfax83]

the point is on PCLAB's mantle stop scaling on the high end and on everything else it keep scaling. PClab is then being used to say on the highend CPU mantle benifit is limited, but that part of PClabs data doesn't align to anyone else...........

Are you serious? Other reviews have definitely shown that Mantle's benefits are far lesser on high end CPUs than on low end CPUs; including Anandtech.

I think the only time that changes is with CF or lower settings which makes the game more CPU bound. But anyone that's going to buy a powerful gaming CPU like a Core i7 likely isn't going to be playing at low or medium settings..

Did any reviews use CF?

 

[el etro]

Then how do you explain this:

The GTX 780, Titan and the GTX 680 SLI all have very similar performance, which implies they've run into a wall.

And the hardware.fr article suggested they had to "zoom in" to reduce the CPU limitation to get higher frame rates.

GTX 690 and 680 have superior theoretical performance. I can't understand this test.

and again, about the multithreading question(All frostbite games):



Battlefield Bad Company 2 singleplayer, two years after:

GTX 680 30% faster than HD 7970.



Battlefield 3 Multiplayer, game launched:

GTX 580 10% faster than HD 6970. Tests here suggests drivers don't favors anyone.



Battlefield 3 Multplayer, one year later:

GTX 680 21% faster than vanilla 7970 and 14% faster than 7970 GHZ at mutiplayer.




Battlefield 4 Multiplayer, three months later:

(attention: its High settings plus field of view 70 and without AA)

GTX 780Ti 22% faster than R9 290x@1000Mhz.