Colab Pro Plus - Page 4

slandrum · 2022-10-06, 21:36

Quote:

Originally Posted by Uncwilly

On one of my 2 unpaid Colab accounts I have not seen a GPU in about a week (but have been getting 2 concurrent sessions, which is a change). On the other, no problem, T4 all the time.

This happens regularly to my free accounts - an account will not be able to get a GPU for a period ranging from 3 days to 20 days, but then it will go back to normal after that. None of my other accounts are affected while this happens. The first couple days after it can get GPU again the sessions may be shorter or longer than they were before, but within a couple days it goes back to normal.

ATH · 2022-10-06, 22:15

Strange, on my free account I have been able to get a T4 roughly every 24 hours for 3-4 hours regularly for a long time (except when I forget to try).

chalsall · 2022-10-06, 23:08

Quote:

Originally Posted by ATH

Strange, on my free account I have been able to get a T4 roughly every 24 hours for 3-4 hours regularly for a long time (except when I forget to try).

Ditto. Ten different accounts / Notebooks.

Edit: Also, a CPU whenever I ask for them. Up to 12 hours each.

ATH · 2022-10-06, 23:28

Quote:

Originally Posted by chalsall

Ditto. Ten different accounts / Notebooks.

Edit: Also, a CPU whenever I ask for them. Up to 12 hours each.

Getting a CPU on the same account does that limit your GPU access? I have been wondering that but haven't tested it.

chalsall · 2022-10-07, 01:39

Quote:

Originally Posted by ATH

Getting a CPU on the same account does that limit your GPU access? I have been wondering that but haven't tested it.

Not at all. Basically, I can (if I'm pulling an all-nighter) get ~3 hours of GPU and ~21 hours of CPU each day for all of the accounts. And, actually, when the GPU becomes available for the day the CPU instance continues for almost exactly an hour before being killed.

LaurV · 2022-10-07, 07:45

To add some more info:

Running two "premium" GPUs will "cost" you 7.6 compute units per hour. YMMV, depending on the server you connect to, or availability, bid/ask, or whatever (this is just a guess), but it means, with a pro plus (50 bucks for 500 cu) you can go as long as 65 hours or so. Which, again, sucks, considering that now, to get to the same "performance" we used to, i.e. run two "premium" GPUs almost continuous, chained for 30 days, you will have to pay about 11 times more (720 hours per month divided by 65 hours per 50 bucks), i.e. be prepared to part with ~550 bucks.

We were lucky to get two "premiums" overnight, and that's how we know, but they only lasted until the remaining ~18 CU's of this month went to zero (a bit longer actually, it seems like they don't check in real time, nor get a feedback/non-maskable-interrupt-line when this happens, so the notebooks may run few [minutes|hours|?] longer. Maybe subject to availability, when other people with more CU's request it.

Now, we are butt-naked again, no GPUs and no CU's.

We don't think we will continue in this conditions...

chalsall · 2022-10-07, 17:00

Quote:

Originally Posted by LaurV

We were lucky to get two "premiums" overnight, and that's how we know, but they only lasted until the remaining ~18 CU's of this month went to zero (a bit longer actually, it seems like they don't check in real time, nor get a feedback/non-maskable-interrupt-line when this happens, so the notebooks may run few [minutes|hours|?] longer. Maybe subject to availability, when other people with more CU's request it.

Interesting on the credits thing being introduced on the Pro tier.

This /might/ have already been experimented with on the Free tier. To offer my empirical... While I was traveling several of my Notebooks were not launched. After two weeks those who hadn't been used in a while were given longer than nominal GPU periods, and for the first day or two would actually allow two GPU launches in a single day.

I'm sure they enjoy watching us as much as we do trying to figure out their algorithms... To me, it's worth the free compute! 8-)

LaurV · 2022-11-12, 03:45

Update: my former 7.x cu/h for "premium" card(s) was for P100. Don't know if still actual.
Few minutes ago I paid the 10.7 bucks (0.7 is the VAT) for a "pay as you go" 100 cu's, which I need for few hours (not prime-related), got the A100-40GB "first strike" (beginner's luck!), and the cu's go down with a rate of 13.16 cu/hour. This is only one card, one session, and it doesn't seem to depend of what you are doing with it. If you don't use it, you have to close the session completely and the rate goes to zero. So I don't believe it will last long enough to try running any owls, but I may give a try with GPU72 TF towards the end.

I assume V100 is in between.

much later edit: one A100 to run, costs 13.08 cu/h, regardless of what you do with it. In the case above, the additional 0.08 came from another, forgotten CPU-only session which was running in background from a different computer.
.

LaurV · 2022-11-12, 10:01

@Chris: something is smelly again with that A100 on your side - it couldn't get any assignments... I wanted to put the last cu's on GPU72, but the loop you know, was rolling and rolling... Aren't any assignment available on your side? Or the old version (before A100 fix) is back?

chalsall · 2022-11-12, 19:00

Quote:

Originally Posted by LaurV

@Chris: something is smelly again with that A100 on your side - it couldn't get any assignments...

@LaurV... Yes... Google Colab "regressed" a while ago such that the free tier was no longer compatible with the version I build for you while you were giving me access to Pro. It was fine when I deployed it.

This is documented somewhere here on the Forum...

Sorry... I just don't have time at the moment to deal with this.

LaurV · 2022-11-13, 09:36

Absolutely no problem if it was "intentional" or at least "known". My concern was that is could be SPE. I don't need any fix right now, let it be till you get the time and mood for it. I only bought (and used) an A100 for few hours, I don't grow them in my garden...

(albeit I would like to...)

2022-10-07, 07:45	#39
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 2³×5×257 Posts	To add some more info: Running two "premium" GPUs will "cost" you 7.6 compute units per hour. YMMV, depending on the server you connect to, or availability, bid/ask, or whatever (this is just a guess), but it means, with a pro plus (50 bucks for 500 cu) you can go as long as 65 hours or so. Which, again, sucks, considering that now, to get to the same "performance" we used to, i.e. run two "premium" GPUs almost continuous, chained for 30 days, you will have to pay about 11 times more (720 hours per month divided by 65 hours per 50 bucks), i.e. be prepared to part with ~550 bucks. We were lucky to get two "premiums" overnight, and that's how we know, but they only lasted until the remaining ~18 CU's of this month went to zero (a bit longer actually, it seems like they don't check in real time, nor get a feedback/non-maskable-interrupt-line when this happens, so the notebooks may run few [minutes\|hours\|?] longer. Maybe subject to availability, when other people with more CU's request it. Now, we are butt-naked again, no GPUs and no CU's. We don't think we will continue in this conditions... Last fiddled with by LaurV on 2022-10-07 at 07:51

2022-11-12, 03:45	#41
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 24050₈ Posts	Update: my former 7.x cu/h for "premium" card(s) was for P100. Don't know if still actual. Few minutes ago I paid the 10.7 bucks (0.7 is the VAT) for a "pay as you go" 100 cu's, which I need for few hours (not prime-related), got the A100-40GB "first strike" (beginner's luck!), and the cu's go down with a rate of 13.16 cu/hour. This is only one card, one session, and it doesn't seem to depend of what you are doing with it. If you don't use it, you have to close the session completely and the rate goes to zero. So I don't believe it will last long enough to try running any owls, but I may give a try with GPU72 TF towards the end. I assume V100 is in between. much later edit: one A100 to run, costs 13.08 cu/h, regardless of what you do with it. In the case above, the additional 0.08 came from another, forgotten CPU-only session which was running in background from a different computer. . Last fiddled with by LaurV on 2022-11-13 at 09:37

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
GPU72/Colab Error	Awesomeotts	GPU to 72	38	2022-08-13 19:40
Colab question	David703	GPU to 72	302	2022-07-01 03:41
I'm getting maybe 10 or 15 minutes with Colab before having to reset	mrk74	GPU to 72	21	2021-06-17 20:00
Revisiting msieve GPU with Colab	EdH	Msieve	4	2020-08-24 02:22
Alternatives to Google Colab	kriesel	Cloud Computing	11	2020-01-14 18:45

2022-10-06, 22:15	#35
ATH Einyen Dec 2003 Denmark 6553₈ Posts	Strange, on my free account I have been able to get a T4 roughly every 24 hours for 3-4 hours regularly for a long time (except when I forget to try).

2022-11-12, 10:01	#42
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 2³·5·257 Posts	@Chris: something is smelly again with that A100 on your side - it couldn't get any assignments... I wanted to put the last cu's on GPU72, but the loop you know, was rolling and rolling... Aren't any assignment available on your side? Or the old version (before A100 fix) is back?

2022-11-13, 09:36	#44
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 24050₈ Posts	Absolutely no problem if it was "intentional" or at least "known". My concern was that is could be SPE. I don't need any fix right now, let it be till you get the time and mood for it. I only bought (and used) an A100 for few hours, I don't grow them in my garden... (albeit I would like to...)