Google Diet Colab Notebook - Page 101

moebius · 2020-12-24, 18:49

Quote:

Originally Posted by petrw1

No P100s lately.

Same at me since 24 Hours, only T4's, P4's. For LL-DC's in the 56M range it is sufficient enough (~ 3000 us/it). Merry Christmas

petrw1 · 2020-12-24, 19:13

Quote:

Originally Posted by moebius

Same at me since 24 Hours, only T4's, P4's. For LL-DC's in the 56M range it is sufficient enough (~ 3000 us/it). Merry Christmas

Maybe Scrooge and the Grinch are running Colab...... LOL

Merry Christmas to you too.

moebius · 2020-12-26, 23:23

Quote:

Originally Posted by petrw1

Maybe Scrooge and the Grinch are running Colab...... LOL Merry Christmas to you too.

Krampus and Knecht Ruprecht should finally release the P100's again. Christmas is already over in Middle European Time (MET), but l still only get T4 bestows.

moebius · 2021-01-04, 12:53

Quote:

Originally Posted by petrw1

Maybe Scrooge and the Grinch are running Colab...... LOL

Now the P100 runtimes are slowly coming back to me. It was actually the holidays.

LaurV · 2021-01-04, 14:15

Maybe they phase out them related to power requirements? All the newer cards (turing, volta, ampere), are a lot less power hungry for the same flops.

moebius · 2021-01-04, 14:41

Quote:

Originally Posted by LaurV

Maybe they phase out them related to power requirements? All the newer cards (turing, volta, ampere), are a lot less power hungry for the same flops.

That is partly true, but the Turing TU100 chip has not yet been invented as far as I know. So your FLOPS/Watt calculation should only work for TF, but not for LL/PRP with Turing cards.

LaurV · 2021-01-05, 07:00

That's also true. I was only trying to guess a reason.

petrw1 · 2021-01-05, 15:33

Quote:

Originally Posted by moebius

Now the P100 runtimes are slowly coming back to me. It was actually the holidays.

Coming back for me too but only getting 4 or 5 hours instead of 8 or 9.

ric · 2021-01-07, 19:09

This is some (bad/funny) news: yesterday I had 2 bad PRP-CFs from the same Colab instance, namely 2747161 and 4222763 (using mprime 30.3b6). To put things in perspective, so far I've managed to complete a bit more than a thousand candidates, with no other glitches.

Retcodes in both cases were all-zeros, no other warnings whatsoever (at least, not written on results.txt). Even funnier, next attempt within the same session had a

Code:

Possible error: round off (0.4906197548) > 0.40625

but recovered itself and was successful at its end.

With all plausibility, a case of a bad underlying piece of metal - what puzzles me is the fact that, despite PRP's better error checking, those were undetected. Alas, since that session is longtime gone, I can't think of a way to (try to) reproduce the issue, nor investigate further.

<edit> those two candidates weren't strictly marked bad, they had mismatching residues (and I don't know what proof processing was like)

Prime95 · 2021-01-07, 22:35

You'll need to run PRP with type-5 residues to get Gerbicz error checking

LaurV · 2021-01-08, 02:43

After about two weeks of absence, I reopened Chris' TF notebook yesterday and got exactly 34 minutes of T4 from first trial. All subsequent trials had "no GPU" and only run for few hours. Now (24 hours later) I have a T4 again (for 5 minutes), let's see how long it'll last.

Additional, I am starting to get a big red error message, saying that "output-height is not a valid annotation. Supported annotations are: vertical-output, run, form-width, display-mode" but the work seems not affected. It seems they changed some names of the lib calls and Chris didn't keep up with (or me, with him?). Advice?

2021-01-07, 19:09	#1109
ric Jul 2004 Milan, Ita 3·61 Posts	2 bad PRP-CFs from Colab This is some (bad/funny) news: yesterday I had 2 bad PRP-CFs from the same Colab instance, namely 2747161 and 4222763 (using mprime 30.3b6). To put things in perspective, so far I've managed to complete a bit more than a thousand candidates, with no other glitches. Retcodes in both cases were all-zeros, no other warnings whatsoever (at least, not written on results.txt). Even funnier, next attempt within the same session had a Code: Possible error: round off (0.4906197548) > 0.40625 but recovered itself and was successful at its end. With all plausibility, a case of a bad underlying piece of metal - what puzzles me is the fact that, despite PRP's better error checking, those were undetected. Alas, since that session is longtime gone, I can't think of a way to (try to) reproduce the issue, nor investigate further. <edit> those two candidates weren't strictly marked bad, they had mismatching residues (and I don't know what proof processing was like) Last fiddled with by ric on 2021-01-07 at 19:14 Reason: last line

2021-01-08, 02:43	#1111
LaurV Romulan Interpreter Jun 2011 Thailand 59·157 Posts	After about two weeks of absence, I reopened Chris' TF notebook yesterday and got exactly 34 minutes of T4 from first trial. All subsequent trials had "no GPU" and only run for few hours. Now (24 hours later) I have a T4 again (for 5 minutes), let's see how long it'll last. Additional, I am starting to get a big red error message, saying that "output-height is not a valid annotation. Supported annotations are: vertical-output, run, form-width, display-mode" but the work seems not affected. It seems they changed some names of the lib calls and Chris didn't keep up with (or me, with him?). Advice? Last fiddled with by LaurV on 2021-01-08 at 02:44 Reason: s/lonh/long/

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Alternatives to Google Colab	kriesel	Cloud Computing	11	2020-01-14 18:45
Notebook	enzocreti	enzocreti	0	2019-02-15 08:20
Computer Diet causes Machine Check Exception -- need heuristics help	Christenson	Hardware	32	2011-12-25 08:17
Computer diet - Need help	garo	Hardware	41	2011-10-06 04:06
Workunit diet ?	dsouza123	NFSNET Discussion	5	2004-02-27 00:42

2021-01-04, 14:15	#1105
LaurV Romulan Interpreter Jun 2011 Thailand 59×157 Posts	Maybe they phase out them related to power requirements? All the newer cards (turing, volta, ampere), are a lot less power hungry for the same flops.

2021-01-05, 07:00	#1107
LaurV Romulan Interpreter Jun 2011 Thailand 59×157 Posts	That's also true. I was only trying to guess a reason.

2021-01-07, 22:35	#1110
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 7347₁₀ Posts	You'll need to run PRP with type-5 residues to get Gerbicz error checking