Using RapidMiner with CUDA GPU

Hi Phivu,

 

There is currently no built-in support for parallel computation using CUDA.  You can use multiple CPU cores however.  You will need our commercial version for doing so: https://rapidminer.com/pricing/

 

This white paper here explains the performance improvements thanks to this.  Depending on your hardware and process, we have seen speed improvements of up to 16x.  More information here: https://rapidminer.com/resource/performance-improvements-rapidminer-studio/

 

Best,

Ingo

Hi Ingo,

 

Thanks for your reply, do you have a plan for GPU support built-in in a near future, or do you think it is necessary to have that? I'm actually using an RM licenced version under my company account, running on a server with 16 logical cores, but sometimes found that i need to run it faster especially with the "Optimize selection" and "Optimize parameters" operators (it took me a few days for sometimes).

 

Thank you very much.

 

Best,

Phivu

Hi Phivu,

 

We are definitely not ruling it out :-)

 

Right now the demand is somewhat limited and most of the times people who ask for this are looking for parallelized learners running on GPUs, not for those meta operators.  But we also have looked into using GPUs for this kind of functions as well.  I can't promise anything for any time soon but we are on top of things here ;-)

 

Cheers,

Ingo

way too late reply, but some researchers sped up K-NN using both CUDA and OpenCL: http://ijates.org/index.php/ijates/article/download/142/119

 

The alternative way (if you are using deep learning) is to use the Keras extension and use the CUDA support built in to the CNTK or Tensorflow backends (or if you want to use OpenCL then use the plaidML backend)

 

There was a short-lived DeepLearning4J plugin (which supports CUDA) but it died 3 years ago in favour of H20 ... https://github.com/LostSummer233/rapidminer-extension-dl4j-pack 

 

 

Hi @wongcr,

RapidMiner 9 has a Deep Learning extension which is based on DL4J and supports CUDA.

 

BR,

Martin

Thats great to hear about DL4J (DeepLearning 4 Java) - but a couple of questions/observations:
(1) Why did you swap from H2O to DL4J ?  
(2) RM9.0 states that it requires CUDA 9.1; however RM9.1 update seems to now only work with CUDA 9.0 ! 
(3) where in the documentation does it state how to setup CUDA suitable for GPU operation (Settings > Preference> General > Deep Learning Backend  )  eg Paths, exact CUDNN and CUDA versions supported etc

Ok, I've answered part of my own question.

Rapidminer 9.0 uses dl-windows-libs-0.8.0-all.jar in your %HOMEPATH%\.rapidminer\extensions\workspace\rmx_deeplearning directory.
This is hardcoded to use the nd4j-cuda-9.1 package (referenced by gpu-backend-0.8.0-all.jar). You MUST have exactly CUDA 9.1 and CuDNN 7.1  installed and the runtime library cudart64_91.dll and in your path.

Rapidminer 9.1 used dl-windows-libs-0.9.0-all.jar. This is hardcoded to use CUDA 9.0 and CUDnn 7.0 
Again, you need to have the right path setup to point to the appropriate CUDA toolkit.

Best to update the marketplace extension documentation to be explicit about the CUDA/CUDNN versions ie: "Lowered CUDA version requirement from 9.1 to 9.0" is not correct as it implies 9.1 is supported.
Should state "Only supports CUDA 9.0 and CUDNN 7.0."

UPDATE: I added a comment to the rapidminer marketplace extension with these details

thank you @wongcr. cc'ing @pschlunder who's in charge of the DL extension. May as well go right to the source.

Scott