hardware questions

Feb 1, 2008 at 4:44pm

hardware questions

Hello list,
we are reconfiguring our WinXP PC at the moment to speed up the fps in
jitter. We use it basically to generate big matrixes (1920*1200) and
process these on the GPU.
One of the bottle necks in this system seems to be loading the matrix
on the GPU as a texture. We did a lot of tweaking in jitter to make
this as efficient as possible. So the next step is to upgrade the
hardware, to speed up the fps.
For this we have the following questions:
What is the best choice of CPU? Is it more important to have the
fastest CPU’s, the fastest CPU busspeed, or the size of the L2-cache?
Regarding the motherboard, is the busspeed critical? And what about
the ammount and the speed of the memory?

We currently use a 8800 GTX 768mb graphics card, with the Asus P5B
Deluxe wifi (busspeed 1066mhz, 4GB 667-DDR2), and a Core 2 Quad CPU
(Q6600 4 x 2400mhz, with a busspeed of 1066mhz and 8mb L2 cache)

Hope some of you can shed some light on these matters,

Best Lucas

#35618
Feb 1, 2008 at 5:51pm

your biggest speed increase would probably come from uploading uyvy
matrices as texture data, you would almost halve your bandwidth to the
GPU (and potentially CPU processing as well), you could then use the
cc.uyvy2rgba.jxs shader family to do YUV to RGBA conversion on the GPU.

Check the optimization threads about this technique. It works for
movies, so it should work for arbitrary matrix data being treated as
imagery on the GPU.

That machine is really fast, from what I can tell. I bet you would be
better off optimizing your patch. jit.change, and asymmetric bangs to
jit.gl.render vs metros to what your matrix/CPU land stuff is doing
would help, along with qmetro/qlim magic, etc.

its hard to tell w/o a patch.

On Feb 1, 2008, at 11:44 AM, subzero wrote:

> Hello list,
> we are reconfiguring our WinXP PC at the moment to speed up the fps
> in jitter. We use it basically to generate big matrixes (1920*1200)
> and process these on the GPU.
> One of the bottle necks in this system seems to be loading the
> matrix on the GPU as a texture. We did a lot of tweaking in jitter
> to make this as efficient as possible. So the next step is to
> upgrade the hardware, to speed up the fps.
> For this we have the following questions:
> What is the best choice of CPU? Is it more important to have the
> fastest CPU’s, the fastest CPU busspeed, or the size of the L2-cache?
> Regarding the motherboard, is the busspeed critical? And what about
> the ammount and the speed of the memory?
>
> We currently use a 8800 GTX 768mb graphics card, with the Asus P5B
> Deluxe wifi (busspeed 1066mhz, 4GB 667-DDR2), and a Core 2 Quad CPU
> (Q6600 4 x 2400mhz, with a busspeed of 1066mhz and 8mb L2 cache)
>
> Hope some of you can shed some light on these matters,
>
> Best Lucas
>
>

#121600
Feb 3, 2008 at 9:54pm

hello vade, thanks for your input, we’ve implemented most of your
suggestions in our system allready, and we have come to a point where
we want to understand the hardware we use better so we can work on
improving that side as well.
we worked a lot with mac’s until recently and the apple universe is
fairly simple, usually you can choose 1 out of 3 cpu speeds, and there
are 3 gpu’s available. so not a lot of tweaking on the hardware
possible.
but in pc land a lot more is possible, so thats why we would like to
understand the importance of the different components better in order
to get the maximum hardware configuration for uploading matrices to
the gpu..
best, lucas

On Feb 1, 2008, at 6:51 PM, vade wrote:

> your biggest speed increase would probably come from uploading uyvy
> matrices as texture data, you would almost halve your bandwidth to
> the GPU (and potentially CPU processing as well), you could then use
> the cc.uyvy2rgba.jxs shader family to do YUV to RGBA conversion on
> the GPU.
>
> Check the optimization threads about this technique. It works for
> movies, so it should work for arbitrary matrix data being treated as
> imagery on the GPU.
>
> That machine is really fast, from what I can tell. I bet you would
> be better off optimizing your patch. jit.change, and asymmetric
> bangs to jit.gl.render vs metros to what your matrix/CPU land stuff
> is doing would help, along with qmetro/qlim magic, etc.
>
> its hard to tell w/o a patch.
>
> On Feb 1, 2008, at 11:44 AM, subzero wrote:
>
>> Hello list,
>> we are reconfiguring our WinXP PC at the moment to speed up the fps
>> in jitter. We use it basically to generate big matrixes (1920*1200)
>> and process these on the GPU.
>> One of the bottle necks in this system seems to be loading the
>> matrix on the GPU as a texture. We did a lot of tweaking in jitter
>> to make this as efficient as possible. So the next step is to
>> upgrade the hardware, to speed up the fps.
>> For this we have the following questions:
>> What is the best choice of CPU? Is it more important to have the
>> fastest CPU’s, the fastest CPU busspeed, or the size of the L2-cache?
>> Regarding the motherboard, is the busspeed critical? And what about
>> the ammount and the speed of the memory?
>>
>> We currently use a 8800 GTX 768mb graphics card, with the Asus P5B
>> Deluxe wifi (busspeed 1066mhz, 4GB 667-DDR2), and a Core 2 Quad CPU
>> (Q6600 4 x 2400mhz, with a busspeed of 1066mhz and 8mb L2 cache)
>>
>> Hope some of you can shed some light on these matters,
>>
>> Best Lucas
>>
>>
>

Telcosystems
PO box 174
3000 AD
Rotterdam
http://www.telcosystems.net
info@telcosystems.net

#121601
Feb 12, 2008 at 10:17pm

hello list, we are still trying to figure out what would be the
fastest PC configuration for jitter texture/gpu usage.
Is Jitter taking advantage of multiple graphic cards in SLI?
Best, Lucas

On Feb 1, 2008, at 6:51 PM, vade wrote:

> your biggest speed increase would probably come from uploading uyvy
> matrices as texture data, you would almost halve your bandwidth to
> the GPU (and potentially CPU processing as well), you could then use
> the cc.uyvy2rgba.jxs shader family to do YUV to RGBA conversion on
> the GPU.
>
> Check the optimization threads about this technique. It works for
> movies, so it should work for arbitrary matrix data being treated as
> imagery on the GPU.
>
> That machine is really fast, from what I can tell. I bet you would
> be better off optimizing your patch. jit.change, and asymmetric
> bangs to jit.gl.render vs metros to what your matrix/CPU land stuff
> is doing would help, along with qmetro/qlim magic, etc.
>
> its hard to tell w/o a patch.
>
> On Feb 1, 2008, at 11:44 AM, subzero wrote:
>
>> Hello list,
>> we are reconfiguring our WinXP PC at the moment to speed up the fps
>> in jitter. We use it basically to generate big matrixes (1920*1200)
>> and process these on the GPU.
>> One of the bottle necks in this system seems to be loading the
>> matrix on the GPU as a texture. We did a lot of tweaking in jitter
>> to make this as efficient as possible. So the next step is to
>> upgrade the hardware, to speed up the fps.
>> For this we have the following questions:
>> What is the best choice of CPU? Is it more important to have the
>> fastest CPU’s, the fastest CPU busspeed, or the size of the L2-cache?
>> Regarding the motherboard, is the busspeed critical? And what about
>> the ammount and the speed of the memory?
>>
>> We currently use a 8800 GTX 768mb graphics card, with the Asus P5B
>> Deluxe wifi (busspeed 1066mhz, 4GB 667-DDR2), and a Core 2 Quad CPU
>> (Q6600 4 x 2400mhz, with a busspeed of 1066mhz and 8mb L2 cache)
>>
>> Hope some of you can shed some light on these matters,
>>
>> Best Lucas
>>
>>
>

Telcosystems
PO box 174
3000 AD
Rotterdam
http://www.telcosystems.net
info@telcosystems.net

#121602
Jul 17, 2011 at 1:53am

Bump!
I would be actually very interested in the previous questions and all possible issues with multiple graphics cards?
Thanks!!

#121603

You must be logged in to reply to this topic.