CPU Bottlenecks - worth knowing !

[exAndrzej]

Hi All,

First up - let me point out that I am not a technical guru

That said, I work around graphics and benchmarking 24x7, so a little 'something' seeps through my non-porous skull sometimes

A HEXUS reader asked a question earlier about why a 6800GT can get close to a much faster Radeon card in 3DMark05

The answer is to do with the CPU becoming a bottleneck

Graphic cards can only process the code/geometry that they are sent by the system running the game

Bear in mind that graphic cards are AWESOMELY FAST and the demand on the CPU is IMMENSE

If the CPU falls behind - then your graphic card sits there 'twiddling its thumbs' - waiting for the next block of data

How can you tell if you have a bottleneck ?

I have to thank the almighty 'Wavey Dave' for this knowledge (Baumann @ Beyond)

He uses a system of graphs for testing that don't appear to make sense to the uneducated eye

Dave runs custom benchmarks and records the results on a graph... nothing new there

However, he starts from 640x480 - runs every resolution mode - ending as high as he can go - then looks at the angle of the graph

The line stays flat - representing the maximum possible throughput (in frames per second) for a few resolutions...

...then you see a drop off in the score

The point where that drop off happens is the point at which the demand on the graphic card finally outweighs the demand on the CPU

i.e. The bottleneck passes from the CPU to the graphic card

This rather crude and totally made up diagram explains the principal:-






You can therefore easily work out if your game/benchmark is limited by the CPU or the graphic card

Set the resolution to 640x480 with IQ forced off (i.e. no AA or AF) and the overall quality of the game set way down below what you would normally play...

...run your test and record the results in a spreadsheet

Increase the resolution and wait for the line to drop off

Once you have done it for 0xA/0xAF - you might want to engage these IQ settings and repeat

Finally, you can increase the in-game quality to maximum and repeat one final time

With 15 results per game on a chart - you will have a very intimate understanding of your system & card's ability to engage that game

Lecture over - hope you found this useful - if you are a boffin and spot any mistakes with this 'presentation' then please post and let me know

[Dougal]

Found this very useful.

But I thought that benchmarks (such as 3DMark 05) were designed to test the GPU wholey on the first 3 and the CPU on the last 2?

Is there any benchmark which does this?

[kempez]

Found this very useful.

But I thought that benchmarks (such as 3DMark 05) were designed to test the GPU wholey on the first 3 and the CPU on the last 2?

Is there any benchmark which does this?

Aquamark 3 is pretty good for this.

I do tha bove for testing the CPU bottleneck. Got 80FPS on all settings on lowest so I'm fairly happy (HL2)

[Dougal]

Aquamark 3 is pretty good for this.

Tried that.

I got 54000 on X800XT.

Tuffin got 75000+ on his 6800GT.

So if mine is superior why did he smash my score?

[kempez]

What CPU you got?

I got 70k with my AGP 6800GT + 754 Athlon 64 3200+ (2.5GHz)

I now get 95k+ on my 939 Venice @ 2.8GHz and 7800GTX

Aquamark measures both CPU and GFX

EDIT: And I meant for CPU bottleneck I do what Andrzej posted

[exAndrzej]

...maybe we need the HEXUS Hierarchy to do an article on benchmarks that completely fail to predict performance in the most popular games

The issue is a tricky one though

Think about the question - 'Which is the fastest car ?' - with the choices being (a) Ferrari F1, (b) Dragster or (c) BMW X5

If we are talking straight line speed (i.e. fast/fixed synthetic benchmark that does not worry about the real world) - then the dragster is champ

Synthetic benchmarks that try to mimic the twists and turns of a real word - albeit in perfect/predictable conditions would be an F1 track and the Ferrari would come in first by a heafty margin

However, in the 'down and dirty' world of real programming - the off-road challenge is maybe a better analogy - and it is the test of all-round performance

To say that the dragster and F1 car would not feature is an understatement

What is more relevant is that the X5 could also put in 'respectable' times on the F1 and dragster track

At ATI, we work to have polished all-round performance - no matter what the terrain ends up looking like

There will always be cases where we don't win...

...but we aim to give our customers 'great performance in all environments' PLUS 'romping victories when the conditions are just right'

Side Note: Just right = FarCry, Counter Strike, HL2 etc

[kalniel]

CPU bottleneck can be good I get AA and AF for 'free' on my system

But more seriously, it's often the case that new games increase their demands on the gfx card faster than the increase in demand on CPU. That means that while your CPU may be the bottle neck for todays game, your GFX may become a bottle neck in the future - or quite likely, they both are. That means you'll see and increase in performance by upping the card AND also see an increase by upping the CPU.

[kempez]

I'm sure nvidia try to get good all-round performance as well though.

I'm no Fanboy (looking forward to seeing what your - ATI's next gen can do ): but wasn't HL2 optimised for ATI cards as a partner release with Valve? Thereby giving you benchmark/speed advantages in oneof/the biggest releases of the summer! Same with nvidia and Doom 3 (tho IMO thats cr@p).

Anyway: point is that benchmarks don't measure performance (so I agree) and its you favourite games you should bench on to check how they will work when it all gets down to the nitty gritty

[Zak33]

Think about the question - 'Which is the fastest car ?' - with the choices being (a) Ferrari F1, (b) Dragster or (c) BMW X5

If we are talking straight line speed (i.e. fast/fixed synthetic benchmark that does not worry about the real world) - then the dragster is champ

Synthetic benchmarks that try to mimic the twists and turns of a real word - albeit in perfect/predictable conditions would be an F1 track and the Ferrari would come in first by a heafty margin

However, in the 'down and dirty' world of real programming - the off-road challenge is maybe a better analogy - and it is the test of all-round performance

To say that the dragster and F1 car would not feature is an understatement

What is more relevant is that the X5 could also put in 'respectable' times on the F1 and dragster track

At ATI, we work to have polished all-round performance - no matter what the terrain ends up looking like
Side Note: Just right = FarCry, Counter Strike, HL2 etc

Analogy of the year. 10/10 goes to the ATI dude for the vid card -> car comparison

One of the critical things to also consider, but which benchmarks often fail to help anyone with is MINIMUM, AVERAGE and FLUCTUATION on frame rate.

CPU'S tend not to fluctuate too much in mid game. GPU's do.

I don't care if it peaks at 100 fps, because if it drops to 5 fps straight afterwards, it's no use to me! I'd rather have the card that averages 40fps all day long, thanks very much, and scores lower too

Best example was always the 9800 Pro. It fluctuated very little on my games

[exAndrzej]

...analogy of the year...

Digs out old electric pump - sets it to suck - inserts in ear - closes nose and eyes - reduces head to get through door

[Dark Horse]

Something I always thought would be useful for graphics cards reviews would be to plot a graph of fps through a timedemo through hl2 or farcry so you could see how the card deals with intense moments.

There must be a lot of cards these days that can do the easy bits (all the eye candy etc) fine but maybe not when there are a lot of little men with guns and bullets flying over the screen...

[Zak33]

Something I always thought would be useful for graphics cards reviews would be to plot a graph of fps through a timedemo through hl2 or farcry so you could see how the card deals with intense moments.

There must be a lot of cards these days that can do the easy bits (all the eye candy etc) fine but maybe not when there are a lot of little men with guns and bullets flying over the screen...

peaks and troughs....bloody good idea.... a long graph looking like a rollr coaster.

You could overlay one to the other!

[Dark Horse]

Looks like I've just found the perfect review for this post:

http://www.extremetech.com/article2/...1854735,00.asp

[mike_w]

One article I liked was this one: http://www.firingsquad.com/hardware/x800_pro_athlon_xp/

It showed the 9800XT and X800 Pro across five processors in a multitude of games.

[oralpain]

I like to look at min FPS and to a lesser extent average FPS. Max FPS is totaly worthless.

To test between two different brands of cards, you can also run the cards slower to make sure there is not a bottleneck on the cpu or system memory.

Like if you have a 6800ultra and an X800XT, try droping the ultra down to 300/775 and the XT to 375/750, as the ratios to the original clock speeds would be the same (75% in this case). Then test at reasonably high graphics settings.

[Butcher]

CPU bottleneck can be good I get AA and AF for 'free' on my system

But more seriously, it's often the case that new games increase their demands on the gfx card faster than the increase in demand on CPU. That means that while your CPU may be the bottle neck for todays game, your GFX may become a bottle neck in the future - or quite likely, they both are. That means you'll see and increase in performance by upping the card AND also see an increase by upping the CPU.

Also more is being done to shift load off the CPU since GPUs are (currently) increasing in performance at a much greater rate. Geometry instancing is one example of this, as is doing more work on the GPU just generally (physics etc). Also things like physics processors and such if they catch on (jury still out on that one).