Skip to main content

⚠️ Please note that this topic or post has been archived. The information contained here may no longer be accurate or up-to-date. ⚠️

OpenCl benchmark / OpenCL-Luxmark vs. Direct Compute?

BeO
Top Commenter
In a former thread (C1 v7), Christian Gruner from PO linked to a Direct Compute (DirectX) benchmark (Passmark software)



Quote from this site:
"Note: OpenCL and Direct compute are different programming interfaces for compiling and running code on video cards. Once the code is running on the card, the performance should be roughly comparable with either interface. So you should be able to use the Direct compute chart above as a reasonable indication of OpenCL performance. "

I have found a link with both benchmarks, Direct Compute (ComputeMark software) and OpenCL (Luxmark software) are shown.


Now, if you compare NVIDIA GeForce GTX 780 Ti with AMD Radeon R9 290X you'll find the following:

- both Direct Compute benchmarks indicate that the NVIDIA is somewhat faster.
- the Luxmark results indicate that the AMD is TWICE as fast

So, it is the opposite!

Luxmark tests OpenCl, and C1 uses the OpenCl interface, but Christian linked to the DIrect Compute benchmark (and they (Passmark) claim it is suitable for OpenCl performance as well), so,

which benchmark is better suited to simulate C1 performance?

If Phase One could state which standard benchmark we should look at, this would be really really great!
It might turn out that one or the other is better suited depending on which C1 task, but this information would also be super super great! ❗️

Thanks and regards
BeO
0

Comments

14 comments

Date Votes
  • Christian Gruner
    Try this link: https://compubench.com/result.jsp?bench ... ase=device

    Also note that per money spent, AMD/ATI cards seems to be faster in Capture One.
    0
  • BeO
    Top Commenter
    Thanks Christian, very fast reply, btw 😄

    Which of the tests to choose from is meaningful for C1? What do you mean by "try", is this benchmark used at Phase One?

    Thanks and regards
    Beo
    0
  • Christian Gruner
    [quote="BeO" wrote:
    Thanks Christian, very fast reply, btw 😄

    Which of the tests to choose from is meaningful for C1? What do you mean by "try", is this benchmark used at Phase One?

    Thanks and regards
    Beo


    The link should lead to the "Video composition" part of the benchmark. This seems to match what we are seeing internally.

    We don't use benchmark pages as such, as Capture One provides its own benchmark number in the logs.
    0
  • BeO
    Top Commenter
    Thanks Christian.

    So I read you like "the ranking of that certain benchmark is highly correlated to our internal experience / our internal benchmarks".

    AMD/ATI cards give more C1 bang for the buck, that's understood. To your internal (and support call) experience, are they equally stable to NVIDIA cards with C1? (In terms of "please set hardware acceleration to Never")

    Thanks again,
    BeO
    0
  • Christian Gruner
    [quote="BeO" wrote:
    Thanks Christian.

    So I read you like "the ranking of that certain benchmark is highly correlated to our internal experience / our internal benchmarks".

    AMD/ATI cards give more C1 bang for the buck, that's understood. To your internal (and support call) experience, are they equally stable to NVIDIA cards with C1? (In terms of "please set hardware acceleration to Never")

    Thanks again,
    BeO


    I didn't say "highly correlated", I said "seems to match" 😉

    Stability-wise there is no obvious difference. Usually it is also a good idea to keep up to date with the drivers.
    0
  • BeO
    Top Commenter
    Thanks Christian.

    Please take it as an idea to publish Capture One's OpenCl benchmark figures in the knowledge base for example.

    Btw, are these figures linearly? E.g. 0.2 is twice as fast as 0.4?

    Cheers,
    BeO
    0
  • Alain Decamps
    Hi

    I don't get the effect op GPU's. When I do test with 9.0.1 while creating local masks (simple brush with zero hardness), the CPU is clearly the bottleneck at 95-100%, my GPU -AMD 7870- is running at 10-25% (and is not getting hotter, so this measurement seems rather correct).

    My CPU is an older i5 2500K,but the newest consumer i7 6700K is only between 50% and 100% faster, enthusiast CPU's like the 5820K will be about 100% faster (if the 6 cores are used fully). This gives a big margin on the GPU before it becomes the bottleneck.


    kind regards,

    Alain
    0
  • BeO
    Top Commenter
    hmm, I assume the OpenCL utilization balances computing power of the CPU and GPU. Maybe it does excatly that, more load on the CPU as this might be the faster device?

    The AMD 7870 has 2 GB memory so it should be used by C1 if you have the preference settings right and if the performance of the card is sufficient. Do you know what the C1 benchmark test figure is for your card (and where to find it)?

    Edit:
    There might possibly also be a dependency on which files (camera) you are using... (see Christians answer)


    cheers
    0
  • Christian Gruner
    [quote="Alain" wrote:
    Hi

    I don't get the effect op GPU's. When I do test with 9.0.1 while creating local masks (simple brush with zero hardness), the CPU is clearly the bottleneck at 95-100%, my GPU -AMD 7870- is running at 10-25% (and is not getting hotter, so this measurement seems rather correct).

    My CPU is an older i5 2500K,but the newest consumer i7 6700K is only between 50% and 100% faster, enthusiast CPU's like the 5820K will be about 100% faster (if the 6 cores are used fully). This gives a big margin on the GPU before it becomes the bottleneck.


    kind regards,

    Alain

    Getting a quicker CPU will help you a good bit (if your disk is fast enough to handle the increase throughput from the GPU, as the CPU will be able to feed it faster)
    0
  • Christian Gruner
    [quote="BeO" wrote:
    hmm, I assume the OpenCL utilization balances computing power of the CPU and GPU. Maybe it does excatly that, more load on the CPU as this might be the faster device?

    The AMD 7870 has 2 GB memory so it should be used by C1 if you have the preference settings right and if the performance of the card is sufficient. Do you know what the C1 benchmark test figure is for your card (and where to find it)?

    Edit:
    There might possibly also be a dependency on which files (camera) you are using... (see Christians answer)


    cheers


    Nop, CO does not balance it, as the GPU and CPU are good at different things. We utilize that when processing. So if one of the components are not fast enough, it will became a bottleneck.
    0
  • Alain Decamps
    [quote="Christian Gruner" wrote:
    [quote="Alain" wrote:
    Hi

    I don't get the effect op GPU's. When I do test with 9.0.1 while creating local masks (simple brush with zero hardness), the CPU is clearly the bottleneck at 95-100%, my GPU -AMD 7870- is running at 10-25% (and is not getting hotter, so this measurement seems rather correct).

    My CPU is an older i5 2500K,but the newest consumer i7 6700K is only between 50% and 100% faster, enthusiast CPU's like the 5820K will be about 100% faster (if the 6 cores are used fully). This gives a big margin on the GPU before it becomes the bottleneck.


    kind regards,

    Alain

    Getting a quicker CPU will help you a good bit (if your disk is fast enough to handle the increase throughput from the GPU, as the CPU will be able to feed it faster)


    Thanks, but I suppose that the disk (for me a SSD) won't change a thing while "brushing 1 image.

    I'm still curious what would be the speed difference between the i5 500k and a i76700k.
    0
  • Christian Gruner
    [quote="Alain" wrote:
    [quote="Christian Gruner" wrote:
    [quote="Alain" wrote:
    Hi

    I don't get the effect op GPU's. When I do test with 9.0.1 while creating local masks (simple brush with zero hardness), the CPU is clearly the bottleneck at 95-100%, my GPU -AMD 7870- is running at 10-25% (and is not getting hotter, so this measurement seems rather correct).

    My CPU is an older i5 2500K,but the newest consumer i7 6700K is only between 50% and 100% faster, enthusiast CPU's like the 5820K will be about 100% faster (if the 6 cores are used fully). This gives a big margin on the GPU before it becomes the bottleneck.


    kind regards,

    Alain

    Getting a quicker CPU will help you a good bit (if your disk is fast enough to handle the increase throughput from the GPU, as the CPU will be able to feed it faster)


    Thanks, but I suppose that the disk (for me a SSD) won't change a thing while "brushing 1 image.

    I'm still curious what would be the speed difference between the i5 500k and a i76700k.


    Correct, the disk is only really used to max potential during processing.
    0
  • Fernando Javier Giménez Cepero
    I have an i7 920 @2,67ghz stock and 18gb of ram with two r9 280x Asus.
    On the log I have a 0.081185 y 0.081334 on each gpu. One is @ 1000th clock and 1500mhz memory and the other a@ 1050mhz clock and 1600mhz.
    Exporting from an ssd 850 evo to a 840 evo with catalog on the 850 I get cpu activity between 80-100 (average 85) and gpus between 20 to 65 ( average 30).
    Exporting 111 NEFS 14 bit lossless compressed of d800 to jpg 300ppp resized to 30cm in widest takes 92 seconds.
    I would like to know if it scales well adding more cores ( upgrading to a 6 or 8 core) and how efficient is. (Adhams law)I pwould think of lan upgrade because of ram...Capture one 9 uses more ram than capture one 8.
    I got errors of using all ram with only capture one.
    I can't link capture ones affinity to only 1 2 3 4 5 6 7 cores instead of all to calculate how efficient capture one is ( with open cl disabled to make calculations comparable). I raised a support case to clarify
    0
  • Alain Decamps
    [quote="fgimenezc" wrote:
    I have an i7 920 @2,67ghz stock and 18gb of ram with two r9 280x Asus.
    On the log I have a 0.081185 y 0.081334 on each gpu. One is @ 1000th clock and 1500mhz memory and the other a@ 1050mhz clock and 1600mhz.
    Exporting from an ssd 850 evo to a 840 evo with catalog on the 850 I get cpu activity between 80-100 (average 85) and gpus between 20 to 65 ( average 30).
    Exporting 111 NEFS 14 bit lossless compressed of d800 to jpg 300ppp resized to 30cm in widest takes 92 seconds.
    I would like to know if it scales well adding more cores ( upgrading to a 6 or 8 core) and how efficient is. (Adhams law)I pwould think of lan upgrade because of ram...Capture one 9 uses more ram than capture one 8.
    I got errors of using all ram with only capture one.
    I can't link capture ones affinity to only 1 2 3 4 5 6 7 cores instead of all to calculate how efficient capture one is ( with open cl disabled to make calculations comparable). I raised a support case to clarify


    Hi normaly you can disable hyperthreading inside the bios to see what difference it makes.

    A i7 6700k will be a faster cpu, probably about 2times faster. I don't know if that would have a big impact on c1?
    0

Post is closed for comments.