Wow! I just got P100 on Colab.:surprised:

[CODE]CUDA device info
name Tesla P100-PCIE-16GB
compute capability 6.0
max threads per block 1024
max shared memory per MP 65536 byte
number of multiprocessors 56
clock rate (CUDA cores) 1328MHz
memory clock rate: 715MHz
memory bus width: 4096 bit[/CODE]

[CODE]
Date Time | class Pct | time ETA | GHz-d/day Sieve Wait
Nov 12 14:09 | 0 0.1% | 4.191 1h06m | 988.82 82485 n.a.%
[/CODE]

[QUOTE=kriesel;530278]Cool. Does that mean you could go back and do that "exercise left for the reader"?[/QUOTE]

Sure... Right after you show me how to square a circle, using only a straightedge and a compass... :wink:

In actuality, this should be possible using Kaggle's API.

Short of using "meta" software acting on behalf of the human in a browser, I don't think this is possible under Colab. (Happy to be proven incorrect, of course.)

[QUOTE=LaurV;530355]2.0.9. Who the hack do you think you are? NASA?
* see a former post of mine about working in a shoes factory[/QUOTE]

ROFLMAO...

With regard to your story about the shoes factory, could you please retell it?

While computers have indelible memory, we simple "wetware" devices sometimes need retriggering of the gateways.

And why not ask the source, rather than run queries against an AI?

This is meant to be funny, and serious, at the same time. Please do share knowledge. :smile:

[QUOTE=LaurV;530355]This would be, at least, my reply to number 2. For me, that was always the most stupid excuse, and only a snobby infatuated crank could forward such a reason.[/QUOTE]

Actually, what I most often encounter is those arguing points 3# through 5#.

They tend to be managers, and some are terrified of being "found out". (Some can't even type.)

Unfortunately, they tend to be "The powers that be, but shouldn't be."

It's all about risk management for them; cover thy butt.

But, every once in awhile, you'll find a client who actually listens to reason.

Such clients are those I keep. To the rest I politely suggest they find other assistance.

Today my CoLab session died quickly
 

I have 2 instances I run whenever I get a backend.
Today I've started them 4 times already and they don't stay up more than 30 minutes.

[QUOTE=chalsall;530398]ROFLMAO...

With regard to your story about the shoes factory, could you please retell it? [/QUOTE]
Maybe it's [URL="https://www.mersenneforum.org/showpost.php?p=521577&postcount=16"]this one...[/URL]

I´m currently having a "funny" situation running mfaktc on Colab (manual assignments).

I´m using a script to launch mfaktc from the Google Drive (mounting it on the Google VM) and all goes pretty well when it assigns a K80. When I am lucky enough to get a P100, mfaktc refuses to start, giving out the "Invalid device function" error.
I guess there is a mismatch between the CUDA version used for the compilation and the compute capability (6.0) of the P100. The K80 has CC 3.7 and works like a charm. The binary I have was compiled for CUDA 6.5, as per the folloowing section of the mfaktc output file:

[B]CUDA version info
binary compiled for CUDA 6.50
CUDA runtime version 6.50
CUDA driver version 10.10[/B]

Will someone be kind enough to point me to a binary that will run on the P100?
I searched around but couldn´t find any, and it´s really frustrating that I can only use the slower GPU.

Thx in advance.

Update: I just tried another connection and this time I got a T4 (CC 7.5). Mfaktc didn´t start, giving out the same error-

[QUOTE=lycorn;530498]Will someone be kind enough to point me to a binary that will run on the P100?[/QUOTE]

This goes back a little while, but you might try the executable I made available [URL="https://mersenneforum.org/showpost.php?p=525235&postcount=19"]way back in early September[/URL].

Please let us know if this helps you.

@Chalsall: Thanks very much. For some reason I had overlooked that post while browsing the forum for a solution.

I tried it and it is running perfectly. Thing is, this time I have been given a K80 (LOL!) so I can´t assure the outcome will be satisfactory on a P100 or T4. I guess it will be, but need the Google gods to have mercy on me and provide a P100. As soon as they do it, I´ll let know.

Update: The Gods were merciful, and starting a new session I got a P100, that is happily chugging along @ ~1300 GHz-d/day.

Many thanks!

I've updated the boinc script. Now the script will fetch virtualbox (needed for some projects such as LHC@Home), and the script will be killed if the user does not supply a project.


[CODE][COLOR=#d4d4d4][FONT=monospace][COLOR=Black]#@title BOINC test
import os.path
import subprocess
import getpass
#Use apt-get to get boinc
!apt-get update
!apt-get install boinc boinc-client
#Some projects require virtualbox to work
#but it requires that the machine has virtualization turned on first
!echo deb https://download.virtualbox.org/virtualbox/debian bionic contrib >> /etc/apt/sources.list
!wget -q https://www.virtualbox.org/download/oracle_vbox_2016.asc -O- | sudo apt-key add -
!wget -q https://www.virtualbox.org/download/oracle_vbox.asc -O- | sudo apt-key add -
!apt-get update
!apt-get install virtualbox-6.0
#cp boinc, boinccmd to working directory
!cp /usr/bin/boinc /content
!cp /usr/bin/boinccmd /content
#create a slots directory if it doesn't exist(otherwise boinc doesn't work)
if not os.path.exists('/content/slots'):
!mkdir slots
#launch the client
#attach to a project as desired
project = input("Enter a URL of a BOINC project: ")
#we hide the account key with getpass for security purposes
acct_key = getpass.getpass("Enter your account key: ")
#we now launch boinc
if project == "":
raiseValueError("No project link specified. Stopping.")
if not os.path.exists('/content/slots/0'):
subprocess.run(['boinc',"--attach_project", project, acct_key])
else:
subprocess.run('boinc')[/COLOR]
[/FONT][/COLOR]
[/CODE]While testing this, I ran into another thing: the machines on Colab do not have virtualization enabled, so this would render virtualbox and other virtualization software useless (unless you use something like bochs, which is a lot slower and isn't compatible with BOINC projects).

Over 600 TF Results and only 2 factors.
 

Not panicked but suspicious.

Anyone can figure out how to get CUDALucas to work on the Colab - Tesla P100 ?

It compiled for me but is not working: [url]https://mersenneforum.org/showthread.php?t=12576[/url]