Adding to this video: I paid for collab pro and I didn't use my credits during the month then I canceled the pro plan to stop paying for something was not using. The compute credits can be used up to 90 days. I used the copute credits with the T4 GPU for feature extraction then after 1 h runtime the execution was terminated by collab. and my credits were used for nothing. It is frustrating that I am experiencing my paid compute credits in the same way that a free plan experiences.
With gpu prices a bit more manageable now, it seems like building a system makes a lot more sense with deep learning now that colab is no longer worth it.
do you think it is cheaper to just build a powerful desktop PC with a 4090, or to just buy a really nice laptop (like an M4 Pro MacBook Pro) and pay for a cloud service gpu instance like through lambda labs, or paperspace to do ML/deep learning work? Really irritated, was in a great workflow testing and troubleshooting a CNN I am building, and I guess I ran out of access to the T4 GPU through the free tier I am using. I want to be able to do more more/research in my field, but don't know whether I should build a beefy desktop PC or just buy a new laptop and get used to connecting via ssh to a gpu instance. What does everyone think is the future of computing in terms of an individual doing this kind of work?
Jeff - This is exactly my frustration with Colab right now. Do you happen to know of a source that compares consumer-grade GPUs to purchase for machine learning? I'm thinking of just buying an Nvidia or AMD GPU to run locally, but all benchmarking sites focus on gaming performance, which may or may not be similar to machine learning.
2024 i get to choose between A100 gpu and V100 gpu, I believe the A100 gpu is more efficient right? Tbh I only bought colab pro+ because of the background running since some of my notebooks tend to have extensive execution times.
I'm debating upgrading to Pro+ for this exact reason, but concerned it'll still unexpectedly disconnect my runtime. How's your experience been since you commented this?
@@mrkmja it hasn’t disconnected for a single time when background running, you don’t even need to be connected to the internet. It has been really good so far. Your notebook will keep running regardless but keep in mind that the limit is 24 hours of background running, after that your notebook will automatically stop running.
Just discovered your channel. I’m going to try out Paperspace as well. Could you also do a video on how you’re managing your experiment results, model training runs or dataset versions?
hello, ı have a 12 GB video dataset that I will train with deep learning which environment should I choose? (colab, AWS, jupyter notebook, pycharm, spider)
I am experimenting Colab Pro+ for CV and object identification training and inferencing. I have found loading images from my Google Drive really slows down the training. Is there a way to speed up the image loading process?
you need to make a zip or tar from your training data and copy large files (archives) to your colab training machine and unpack there. Best is to buy 4090 though. Colab pro + became too expensive for serious work! (24/7 training)
@@denysivanov3364 How do I copy the files into the training machine? I currently have all my images on GoogleDrive (that I have to replicate from GCP bucket storage).
@@Jerry-kh8tz I tried to comment but I dont see it. Just google "google colaboratory mount google drive" then copy code snippet into your notebook and voila :) After its mounted cp -rf will work! But google limits number of files transferred per day, so files should be packed into +- large tar or zip files.
I would like to build my own PC with dual RTX 4090 for machine learning, may I ask which motherboard and case you would recommend? I can hardly find a board that can fit two RTX 4090.
I would advise against it. I tried this dual setup a long time ago and it didn't compare to just a single high-end GPU, in the end I had to sell the GPUs and buy a new one. Maybe they fixed the issue, I'm not up to date on this, but I'd do a lot of research before doing it.
thank you , please can you explain, what computer units means? its said you have 1300 unites, and for example if I purchase 100 unites, how long can I use that .. lets say I spend 5hr a day to work with Stable Diff.
It depends. I don't send it through GDrive or even map it. GDrive is slow, I just use GDrive to archive. I will either place the data somewhere I can wget/curl it. Or I will use the colab Python package's upload feature and place it into the /content area, realizing that area is gone when the VM dies. I copy the final work product to GDrive.
I found a way to use large datasets in Colab from an article called, How to successfully add large data sets to Google Drive and use them in Google Colab, by Magdalena Konkiewicz Hope it helps
Lambda labs charges $1.1 per hour for an A100 instance. Colab pro is more expensive at $1.3 per hour. Either way, if someone needs a GPU for full time research they're still better off building a PC with a dedicated GPU.
when you are subscribed as pro+ can you use a non-GPU runtime and leave it in the background (for example for a long gridsearch for a randomforest) and will it not consume your compute units?
They might be consuming a very miniscule amount. After signing up, I only used the CPU / T4 (supposedly free) to login and still had some credits deducted. If it is free to non-subscribers, it should be free to subscribers.
@@praveenbehara yeah it does consume a bit and it’s annoying because as you say it should be free to subscribers as well. I use 2 google accounts now to avoid that.
I'm subscribed to Colab Pro, not Colab Pro+. I get A100-40GB all the time too. Does anyone get any higher? I actually need more VRAM for some heavy deep-learning training.
summarized by Claude+ "Here is a summary of the key points from the video: - Google Colab started as a free service but now charges for GPU access and additional features. The free tier now provides limited, intermittent GPU access. - Colab Pro and Pro Plus are paid subscription plans that provide faster, dedicated GPUs and additional features like a terminal window. However, the specific GPUs and resources provided are not clearly specified. - The video creator finds the lack of transparency around the GPUs and resources provided to be frustrating. The paid plans can end up being expensive if a lot of compute is used, but it's hard to determine if the cost is worthwhile without more details. - Colab Pro Plus provides continuous execution, so notebooks can be stopped and restarted without losing progress. This makes it easy to accidentally run up high compute costs if unchecked. - As of early 2023, Colab Pro Plus appears to primarily provide access to A100 GPUs, which can be quite expensive. The creator would prefer to see lower-cost options like V100 GPUs available as well. - Overall, the creator feels Colab started as a very useful free service but has become less valuable and transparent over time. The paid plans may be worth it for some limited use cases, like accessing the terminal window, but are hard to evaluate fully without more details on the resources provided. - Alternative options mentioned include Kaggle, SageMaker, and managing your own EC2 instances. The key takeaways are: 1) Google Colab's free GPU access and resources have become more limited over time. The paid plans provide faster, dedicated resources but lack transparency into what is actually provided. 2) Colab Pro and Pro Plus can end up being expensive, especially if high-end GPUs like the A100 are used. More cost-effective options would be preferable for many use cases. 3) While still useful for some limited purposes, Colab is not as compelling of an option as it used to be due to the lack of details provided about the paid resources. Alternative services may be better options depending on your needs. 4) If you do use Colab, be very aware of how much you are spending on compute to avoid surprise charges from continuously running or resource-intensive notebooks. Does this summary accurately reflect the key highlights and takeaways from the video? Let me know if you have any feedback or additions. I'm happy to clarify or expand on any part of the summary."
Hi Jeff. How did you get the Googl Colab Runtime type to offer "Python 3.9 (tensorflow)"? I'm trying to get Tensforflow 2.9 with Python 3.8 and am facing challenges. It just defaults to Python3.10 (latest)
In my stupid country called India, the gpus are like 3* higher priced than compared to USA. So I'll rather stick with colab. I still use a frickin GT 610 as a programmer and a 3D artist. SMH
Dude dare you call your country stupid! Forgot about India you don't deserve to be a resident of Pakistan 🤣 And GPU prices have been reduced a lot since the downfall of crypto. Stop complain stupid!
Adding to this video: I paid for collab pro and I didn't use my credits during the month then I canceled the pro plan to stop paying for something was not using. The compute credits can be used up to 90 days. I used the copute credits with the T4 GPU for feature extraction then after 1 h runtime the execution was terminated by collab. and my credits were used for nothing. It is frustrating that I am experiencing my paid compute credits in the same way that a free plan experiences.
With gpu prices a bit more manageable now, it seems like building a system makes a lot more sense with deep learning now that colab is no longer worth it.
Yes, I agree.
@stevedevney7322 can you share which GPU you have used for building your system
@@abdennacerlachiheb319 I have a 1080, and a 3090. With the 3090 being for nets.
@@stevedevney7322thanks. With the GPU cost n cost of energy, how much does it approximately come upto
Can anyone tell me what exactly does compute unit actually mean?
Paperspace subscription is quite good, cost 8$ a month and you get to choose the GPU, they also have a free tier
what are the GPUs offered in $8 plans?
Yes, I am wondering this myself. Hope to try them out soon.
Paperspace looks really interesting.
I had 3 Colab Pro+, but dropped all three after their pricing change. Just wasn’t worth it any more. Paperspace gave more bang.
Thank you for the alt suggestions, please keep them coming.
im a student and google collab pro is not for student, ill recommend google remove the "student" tag from their packages list plan. Too expensive.
do you think it is cheaper to just build a powerful desktop PC with a 4090, or to just buy a really nice laptop (like an M4 Pro MacBook Pro) and pay for a cloud service gpu instance like through lambda labs, or paperspace to do ML/deep learning work? Really irritated, was in a great workflow testing and troubleshooting a CNN I am building, and I guess I ran out of access to the T4 GPU through the free tier I am using. I want to be able to do more more/research in my field, but don't know whether I should build a beefy desktop PC or just buy a new laptop and get used to connecting via ssh to a gpu instance. What does everyone think is the future of computing in terms of an individual doing this kind of work?
Thanks for this review, man. Really helpful for me
Jeff - This is exactly my frustration with Colab right now. Do you happen to know of a source that compares consumer-grade GPUs to purchase for machine learning? I'm thinking of just buying an Nvidia or AMD GPU to run locally, but all benchmarking sites focus on gaming performance, which may or may not be similar to machine learning.
Gaming perf should be an okay-ish benchmark honestly. And memory ofc
quadro cards are 48gb for machine learning... expensive
can you redo the test now?
2024 i get to choose between A100 gpu and V100 gpu, I believe the A100 gpu is more efficient right? Tbh I only bought colab pro+ because of the background running since some of my notebooks tend to have extensive execution times.
I'm debating upgrading to Pro+ for this exact reason, but concerned it'll still unexpectedly disconnect my runtime. How's your experience been since you commented this?
@@mrkmja it hasn’t disconnected for a single time when background running, you don’t even need to be connected to the internet. It has been really good so far. Your notebook will keep running regardless but keep in mind that the limit is 24 hours of background running, after that your notebook will automatically stop running.
@@mateolopez6205 Okay perfect because I just made the move. 😂 thanks for letting me know!
Just discovered your channel. I’m going to try out Paperspace as well. Could you also do a video on how you’re managing your experiment results, model training runs or dataset versions?
Thank you very much colab pro how much ram available
hello, ı have a 12 GB video dataset that I will train with deep learning which environment should I choose? (colab, AWS, jupyter notebook, pycharm, spider)
I am experimenting Colab Pro+ for CV and object identification training and inferencing. I have found loading images from my Google Drive really slows down the training. Is there a way to speed up the image loading process?
you need to make a zip or tar from your training data and copy large files (archives) to your colab training machine and unpack there. Best is to buy 4090 though. Colab pro + became too expensive for serious work! (24/7 training)
@@denysivanov3364 Thank you.
@@denysivanov3364 How do I copy the files into the training machine? I currently have all my images on GoogleDrive (that I have to replicate from GCP bucket storage).
@@Jerry-kh8tz I tried to comment but I dont see it. Just google "google colaboratory mount google drive" then copy code snippet into your notebook and voila :) After its mounted cp -rf will work! But google limits number of files transferred per day, so files should be packed into +- large tar or zip files.
Pack images in tensor and save it.
I would like to build my own PC with dual RTX 4090 for machine learning, may I ask which motherboard and case you would recommend? I can hardly find a board that can fit two RTX 4090.
I would advise against it. I tried this dual setup a long time ago and it didn't compare to just a single high-end GPU, in the end I had to sell the GPUs and buy a new one. Maybe they fixed the issue, I'm not up to date on this, but I'd do a lot of research before doing it.
thank you , please can you explain, what computer units means? its said you have 1300 unites, and for example if I purchase 100 unites, how long can I use that .. lets say I spend 5hr a day to work with Stable Diff.
100 did one video only
How do you handle big data uploads to CoLab ? Also with Google Drive?
It depends. I don't send it through GDrive or even map it. GDrive is slow, I just use GDrive to archive. I will either place the data somewhere I can wget/curl it. Or I will use the colab Python package's upload feature and place it into the /content area, realizing that area is gone when the VM dies. I copy the final work product to GDrive.
I found a way to use large datasets in Colab from an article called, How to successfully add large data sets to Google Drive and use them in Google Colab, by Magdalena Konkiewicz
Hope it helps
you can parse your data into TFRecord something like that, which will make it run much more faster in your project
Isn't sagemaker studio free?
Runpod good?
How much intense we can run in Google colab pro+ in one time
Lambda labs charges $1.1 per hour for an A100 instance. Colab pro is more expensive at $1.3 per hour. Either way, if someone needs a GPU for full time research they're still better off building a PC with a dedicated GPU.
I have one question, which one is better collab pro + Services or buy NVIDA RTX 3060 for Research purspose
Colab pro! V100 is better than 3060. But your max execution time is 24hrs keep that in mind
@@RajdeepBorgohainRajdeep thanks for your reply. Do you mean to say everyday limit in collab
Can you run multiple colab sessions at once with Colab Pro?
Yes, I frequently do that.
@@HeatonResearch How many sessions? 2 or 3 or?
On the free version I notice I can only run 1
Yeah hwo man? @@HeatonResearch
Did you get the answer?@@prchaser2081
when you are subscribed as pro+ can you use a non-GPU runtime and leave it in the background (for example for a long gridsearch for a randomforest) and will it not consume your compute units?
They might be consuming a very miniscule amount. After signing up, I only used the CPU / T4 (supposedly free) to login and still had some credits deducted. If it is free to non-subscribers, it should be free to subscribers.
@@praveenbehara yeah it does consume a bit and it’s annoying because as you say it should be free to subscribers as well. I use 2 google accounts now to avoid that.
I'm subscribed to Colab Pro, not Colab Pro+. I get A100-40GB all the time too. Does anyone get any higher? I actually need more VRAM for some heavy deep-learning training.
Bro no you will not get atleast for now
does anyone know how much disk colabs pro plus provide to us?
How about paperspace they have a free start to pay as well?
Ooops just realized someone beat me to this comment by 4h 😂
Oh this is good, just another vote for paperspace. I will certainly check them out.
Im trying to export 5100 frames 2/3 times a week, do you think Colab pro will work?
I render video/frames on it, and have rendered in that size. Depends on the amount of processing per frame.
Can we connect snowflake database to collab?
If you have a public URL for Snowflake, I don't see why not.
summarized by Claude+
"Here is a summary of the key points from the video:
- Google Colab started as a free service but now charges for GPU access and additional features. The free tier now provides limited, intermittent GPU access.
- Colab Pro and Pro Plus are paid subscription plans that provide faster, dedicated GPUs and additional features like a terminal window. However, the specific GPUs and resources provided are not clearly specified.
- The video creator finds the lack of transparency around the GPUs and resources provided to be frustrating. The paid plans can end up being expensive if a lot of compute is used, but it's hard to determine if the cost is worthwhile without more details.
- Colab Pro Plus provides continuous execution, so notebooks can be stopped and restarted without losing progress. This makes it easy to accidentally run up high compute costs if unchecked.
- As of early 2023, Colab Pro Plus appears to primarily provide access to A100 GPUs, which can be quite expensive. The creator would prefer to see lower-cost options like V100 GPUs available as well.
- Overall, the creator feels Colab started as a very useful free service but has become less valuable and transparent over time. The paid plans may be worth it for some limited use cases, like accessing the terminal window, but are hard to evaluate fully without more details on the resources provided.
- Alternative options mentioned include Kaggle, SageMaker, and managing your own EC2 instances.
The key takeaways are:
1) Google Colab's free GPU access and resources have become more limited over time. The paid plans provide faster, dedicated resources but lack transparency into what is actually provided.
2) Colab Pro and Pro Plus can end up being expensive, especially if high-end GPUs like the A100 are used. More cost-effective options would be preferable for many use cases.
3) While still useful for some limited purposes, Colab is not as compelling of an option as it used to be due to the lack of details provided about the paid resources. Alternative services may be better options depending on your needs.
4) If you do use Colab, be very aware of how much you are spending on compute to avoid surprise charges from continuously running or resource-intensive notebooks.
Does this summary accurately reflect the key highlights and takeaways from the video? Let me know if you have any feedback or additions. I'm happy to clarify or expand on any part of the summary."
This is not summary 🤣🤣
Thanks for your insight. Please do a follow up video on better alternatives.
Yes, I am planning on that.
Hi Jeff. How did you get the Googl Colab Runtime type to offer "Python 3.9 (tensorflow)"?
I'm trying to get Tensforflow 2.9 with Python 3.8 and am facing challenges. It just defaults to Python3.10 (latest)
Wow, this can get hopefully expensive in a hurry
In my stupid country called India, the gpus are like 3* higher priced than compared to USA. So I'll rather stick with colab. I still use a frickin GT 610 as a programmer and a 3D artist. SMH
Dude dare you call your country stupid! Forgot about India you don't deserve to be a resident of Pakistan 🤣
And GPU prices have been reduced a lot since the downfall of crypto. Stop complain stupid!
India isn't stupid. It's a land of opportunities. Last time I checked, it's the same price
@@lubdubguyyou probably live in a cave
Well I know they're hard to get except for a place called "Nehru Place" ironically. Book a trip to Delhi if you want 1
Thank you for this instructive video
Glad it was helpful!
Runpod is about 6 times cheaper.
Oh thanks! Will check that one out for sure.
I find ram is much better in Colab Plus
I agree. Have not specifically tested it though.
51 GB of system memory is available in Colab Plus now. How is it in Colab Pro Plus?
49,9$ for one month or for three months! Because auto subscription to new 500 computation unit
thx