Program to test fillrate and mem band

[grisnak]

Well, as the topic says, is there a program for testing the fillrate and mem bandwidth on a graphics card? Id like to see how much my OC increased it just for fun :)

[djbbenn]

When I was testing my overclock I used 3DMARK03. There was a fillrate test, but no mem bandwidth. I tested default then did it at overclock. After your finished running it you can look at how many FPS you got and compair it to default. :D

-Dan

[grisnak]

Yeah, but I wanna know gb/s and such ;)

Maybe everest pro?

EDIT: everest pros membandwidth was strange... According to it I had 36000mb/s on my graphics ram, compared to the like 14 gig the gpu database says. Are there different ways to count?
Is there any test that I can compare with the gpu-database?

[djbbenn]

Wow...ha that means you got a 21gig increase lol. I don't know any goods for testing graphic mem though. Google it up I guess. :D

-Dan

[zAAm]

You can work it out. Memory bandwidth is measured like this:

(Path width (64/128/256bits) / 8) x RAM MHz = MB/s

example:

9550 128bit with 400MHz (DDR) memory

(128 / 8) x 400 = 6400MB/s

9550 128bit with 500MHz (DDR) memory (overclocked) then

(128 / 8) x 500 = 8000MB/s

Although the GPU database calculates the bandwidth with non-DDR speeds for some or other reason not known to me?? :confused:

[grisnak]

Is there a way to calculate fillrate too?
Thx for the answer!

[zAAm]

Yeah, but it's more tricky to calculate. An easy way to do this is do a Quake3 timedemo at 640x480 at 32bit and check your framerate. Multiply the resolution (640 x 480 = 307200) and multiply it with your framerate. Then multiply it with the bit depth / 8.

Example:

Quake 3 at 640x480 @ 400fps on 32bit

640 x 480 x 400 x (32/8) = 491520000 pixels/second

devide by a million to get Mega pixels/sec. ie. 491 Mega Pixels/s

I hope this is accurate enough... (Use different resolutions to check. CPU and memory bandwidth limitations all influence the fill rate :D )

[grisnak]

thx. wanted to compare to the gpu database so ill try the core*bus/8
cheers!

[zAAm]

Although it's purely theoretical o'course... ;) But I can see the need to compare it to other GPU's.

[djbbenn]

Nice, those calculations could be very usefull. ;)

-Dan

[grisnak]

Think you should use the max possible resolution to try and keep the CPU out of the equation. Mine differs between 350 and 870 MegaPixels/s.

If you want fillrate in GB/s like the GPU database:

Core clock x (bus width / 8)

ex. 250MHz 9550 with 128bit

250 x (128 / 8) = 1000MB/s = 1GB/s approx.

Uhm.. wait.. theres something wrong here...
128/8 = 16
250*16 = 4000
?

[zAAm]

Oops. My bad. Second time this has happened. I don't pay attention when doing these formulas... ;)

It's:

Core clock x (bus width / 32)

So that would be 250 x (128 / 32) = 1000MB/s = 1GB/s approx.

Somehow 4 stuck in my mind...

[grisnak]

This formula (475*(256/32) should predict the fillrate if my x800pro. But 475*(256/32) is 3800. And according to the gpu database its 5400..
Does it have nothing to do with pipelines or something? Or is the gpu db maybe wrong? +_+

[zAAm]

Yup. Just tried to adjust my theory. You're right. When I tried to get it, I tested it only on a couple of cards with all of them 4 pipes. (I though 4 stuck in my mind! ;) ).

So new formula (the right one this time):

Core clock x pixel pipes = fill rate :p

That is on the 9550:

250 x 4 = 1GB/s

And on you X800 Pro:

475 x 12 = 5.7GB/s

Sorry about this... I should learn to test these before I give them to people! :o

[zAAm]

So, to sum up the correct formulas:


Memory bandwidth

(Memory path width (64/128/256bits) / 8) x RAM clock in MHz = MB/s


Fill Rate (MPixels/s)

Resolution x Framerate x (Colour depth / 8) = MPixels/s


Fill Rate (GB/s)

Core clock in MHz x amount of pixel pipelines = GB/s


Sorry again for all the other mishaps... :o