[QUOTE=chalsall;527768]However, if your team is trying to identify Mersenne prime numbers by using Kaggle GPUs, that is totally fine.[/QUOTE]

Wow!

[QUOTE=chalsall;527768]So, way back in mid-September, I reached out to both Colab and Kaggle, trying to get some clarity on "fair use". I pointed to the beginning of this thread; just got the official word back:[/QUOTE]
:ttu:

[QUOTE=Uncwilly;527775][Two thumbs up!][/QUOTE]

I giggle every time I see that emoji. That guy looks very much like my pot dealer back in Canada did twenty years ago... :wink:

I am playing on Google colab and successfully run GFN on Tesla80. Now I try to run TPU on mprime doing P-1
And I got it , but that is run on

[QUOTE]Optimizing for CPU architecture: Core i3/i5/i7, L2 cache size: 256 KB, L3 cache size: 45 MB[/QUOTE], so for my point of view it is not TPU is is ordinary older ( FMA3 ) Intel
Does anyone know where is catch , can mprime use TPU?


Second I register on Kaggle, I would like to start GFN on GPU.
Is any script available?

[QUOTE=pepi37;527778]I am playing on Google colab and successfully run GFN on Tesla80. Now I try to run TPU on mprime doing P-1
And I got it , but that is run on

, so for my point of view it is not TPU is is ordinary older ( FMA3 ) Intel
Does anyone know where is catch , can mprime use TPU?


Second I register on Kaggle, I would like to start GFN on GPU.
Is any script available?[/QUOTE]

TPU is designed to run Neural Network and AI things, and they are quite specialized and designed for low precision workload (correct me if i'm wrong). But so far I don't think any software can use them so I guess firing up a purely CPU instance would help.

CUDALucas 2.06 CUDA 10.1 Linux Build?
 

Since I can't get GPUOWL to run on kaggle (neither can I compile as I think I need rocm stuff and gcc version 8 (which I think kaggle only provides gcc-6)), I wanted to run CUDALucas instead, but I can't find a linux build with CUDA 10.1 anywhere. Does anyone know where I can download one?

[QUOTE=xx005fs;527786]Since I can't get GPUOWL to run on kaggle (neither can I compile as I think I need rocm stuff and gcc version 8 (which I think kaggle only provides gcc-6)), I wanted to run CUDALucas instead, but I can't find a linux build with CUDA 10.1 anywhere. Does anyone know where I can download one?[/QUOTE]

I take it back. I actually figured out how to run GPUOWL on kaggle and it runs like a champ! I am netting 1162us/it on the P100, which completes an exponents at just under the 30 hour weekly quota, which means 1 exponent per week (impressive stuff). However, I need to figure out a way to store the checkpoints and worktodo files, and I have yet to find a solution.

Just look at this speed!
[CODE]2019-10-12 04:24:50 OpenCL compilation in 1594 ms
2019-10-12 04:24:51 /kaggle/working/gpuowl/90396473/90396473.owl not found
2019-10-12 04:24:51 /kaggle/working/gpuowl/90396473/90396473-old.owl not found
2019-10-12 04:24:51 starting from the beginning.
2019-10-12 04:24:54 90396473 OK 1000 0.00%; 1162 us/sq; ETA 1d 05:11; bcdd6dc4e29c5fac (check 0.79s)
2019-10-12 04:25:51 90396473 50000 0.06%; 1162 us/sq; ETA 1d 05:10; ab64be3254ba0225
2019-10-12 04:26:49 90396473 100000 0.11%; 1162 us/sq; ETA 1d 05:09; 9c6befc766902a1f
2019-10-12 04:27:47 90396473 150000 0.17%; 1162 us/sq; ETA 1d 05:08; 3ebfb1883b141232
2019-10-12 04:28:45 90396473 200000 0.22%; 1162 us/sq; ETA 1d 05:07; 787b6520a862cfc3
2019-10-12 04:29:44 90396473 OK 250000 0.28%; 1162 us/sq; ETA 1d 05:06; 69901e0dfa55dce4 (check 0.77s)
2019-10-12 04:30:42 90396473 300000 0.33%; 1162 us/sq; ETA 1d 05:05; dbd6611c9e73b0c2
2019-10-12 04:31:40 90396473 350000 0.39%; 1162 us/sq; ETA 1d 05:04; dc0f70c29515daa2
2019-10-12 04:32:39 90396473 400000 0.44%; 1162 us/sq; ETA 1d 05:03; ec87482d5f4693e6
2019-10-12 04:33:37 90396473 450000 0.50%; 1162 us/sq; ETA 1d 05:02; 59e15057572f9aa3[/CODE]

Unfortunately I probably fluked it last time when I made it work, so now I need to figure out what I did.

UPDATE: Figured out the problem. Here's the code (it might throw some error when installing nvidia-opencl-icd the first time, but when running it again it should work. clinfo is there to make sure it detects opencl device)
[CODE]!chmod 777 /tmp
!echo "deb http://ftp.us.debian.org/debian testing main contrib non-free" | tee -a /etc/apt/sources.list
!apt update
!apt install gcc g++ clinfo ocl-icd-libopencl1 nvidia-opencl-icd -y
!apt install nvidia-opencl-icd -y
!clinfo
!cp -r /kaggle/input/gpuowl /kaggle/working
!cd /kaggle/working/gpuowl && chmod 777 gpuowl && ./gpuowl -use ORIG_X2[/CODE]

So my second question arise: can I on Kaggle run GFN for GPU?
Linus app exist and can be downloaded from Primegrid: so what else I need?

Now I have Colab up and running, I wondered, can you use Kaggle to get a seperate notebook running using the same access code?

It would be great to get both running simulatneously :)

[QUOTE=bayanne;527800]Now I have Colab up and running, I wondered, can you use Kaggle to get a seperate notebook running using the same access code? It would be great to get both running simulatneously :)[/QUOTE]

I'm inferring you're referring to the GPU72_TF notebook?

You can use the same AKey multiple times, on as many instances as you'd like. I would suggest, however, that you instead create new AKeys for the different instances, so you can keep track of what's running and what has stopped. You may have as many AKeys as you'd like.

And, further, Kaggle is quite nice in that you can "Commit" up to two GPU instances at a time, which runs in the background -- even if you close your browser. Some around here burn through their 30-hour weekly quota in less than a day...

[QUOTE=pepi37;527798]So my second question arise: can I on Kaggle run GFN for GPU?
Linus app exist and can be downloaded from Primegrid: so what else I need?[/QUOTE]


Yes you can run GFN for GPU on Kaggle:


[CODE]#get the OpenCL packages needed to run the opencl executable
!chmod 777 /tmp
!echo "deb http://ftp.us.debian.org/debian testing main contrib non-free" | tee -a /etc/apt/sources.list
!apt update
!apt install gcc g++ clinfo ocl-icd-libopencl1 nvidia-opencl-icd -y
!apt install nvidia-opencl-icd -y
#import os.path
import os.path
#useful things to see where we are (uncomment to use)
#!pwd
#!ls -l
#see if we have a GFN folder in the working directory, if not, create it
if not os.path.exists('/kaggle/working/GFN'):
!mkdir GFN
#cd to the GFN directory and get the latest version of genefer
%cd /kaggle/working/GFN
!svn checkout https://subversion.assembla.com/svn/genefer/tags/3.3.4/linux
#and now we run genefer
%cd /kaggle/working/GFN/linux
#we need a GPU here, and a sieve file
!clinfo
#for testing purposes here I am running the benchmark
!./geneferocl_linux64 -b -x ocl4
#to test a file you would probably need to use
#!./geneferocl_linux64 -d 0 <filename>
#where <filename> is the name of the file to test[/CODE]this will get the app, and all needed packages for the app to run properly. The only thing you will need is a sieve file and then the only changes you will need to make is to uncomment the second to last line and replace <filename> with the name of the file.