Hi all,
\n\nThis morning only I observed that on the console in Kaggle kernels, GPU is mentioned as \"Off\". Kaggle kernels have never had this GPU thing mentioned altogether before.
\n\nThe following information is available:\nCPU 0% GPU OFF `RAM 214MB/ 17.2 GB Disk 55.4MB/1GB
\n\nI am interested in knowing how to turn this GPU on? \nAny idea?
", "messages": [ { "id": 309141, "postDate": "2018-04-04T17:57:15.320Z", "content": "You can add a GPU from the \"Settings\" tab in the kernel (expand the menu on the right)
", "rawMarkdown": "You can add a GPU from the \"Settings\" tab in the kernel (expand the menu on the right)", "votes": 7, "replies": [ { "id": 309155, "postDate": "2018-04-04T18:22:56.537Z", "content": "That's awesome Ben...Thanks !
\n\nCould we have more RAM for this competition ? :)
\n\nIt becomes nearly impossible to test my new features due to lack of RAM :p
", "rawMarkdown": "That's awesome Ben...Thanks !\n\nCould we have more RAM for this competition ? :)\n\nIt becomes nearly impossible to test my new features due to lack of RAM :p", "votes": 4 }, { "id": 309156, "postDate": "2018-04-04T18:29:28.210Z", "content": "Any info on what kind of GPU is it? i.e., the amount of GPU RAM, how many Tflops, etc.
", "rawMarkdown": "Any info on what kind of GPU is it? i.e., the amount of GPU RAM, how many Tflops, etc. ", "votes": 9 }, { "id": 309159, "postDate": "2018-04-04T18:39:36.280Z", "rawMarkdown": "", "isDeleted": true }, { "id": 309164, "postDate": "2018-04-04T18:46:05.547Z", "content": "\n\n\nSerigne wrote
\n \n\n\n \nThat's awesome Ben...
\nCould we have more RAM for this competition ? :)
\n \nIt becomes nearly impossible to test my new features due to lack of RAM :p
\n
You can shard your df, blend the different models, and get the same result.
", "rawMarkdown": "> **Serigne wrote**\n> \n>> That's awesome Ben...\n> \n> Could we have more RAM for this competition ? :)\n> \n> It becomes nearly impossible to test my new features due to lack of RAM :p\n\nYou can shard your df, blend the different models, and get the same result." }, { "id": 309185, "postDate": "2018-04-04T19:20:00.603Z", "content": "When I tried using GPU, it reduced my memory limit to 6 GB (oddly using the non-digital definition of \"giga\" as 10^9 rather than 2^30). Also when I tried to run another kernel at the same time, it applied a 6GB limit to that kernel too. If that limit applies, it makes the GPU pretty useless for this particular competition. And if the limit applies to all concurrently running kernels, it means we can't experiment while running something else. (6 GB would make sense as a limit on GPU memory, but I don't see any reason to limit system memory that way. I'm wondering if there's some kind of bug here.)
", "rawMarkdown": "When I tried using GPU, it reduced my memory limit to 6 GB (oddly using the non-digital definition of \"giga\" as 10^9 rather than 2^30). Also when I tried to run another kernel at the same time, it applied a 6GB limit to that kernel too. If that limit applies, it makes the GPU pretty useless for this particular competition. And if the limit applies to all concurrently running kernels, it means we can't experiment while running something else. (6 GB would make sense as a limit on GPU memory, but I don't see any reason to limit system memory that way. I'm wondering if there's some kind of bug here.)", "votes": 1 }, { "id": 309187, "postDate": "2018-04-04T19:24:59.373Z", "content": "Even, I have faced the same issue. However, Nvidia Tesla k80 would be a great add on for this competion.
", "rawMarkdown": "Even, I have faced the same issue. However, Nvidia Tesla k80 would be a great add on for this competion. " }, { "id": 309211, "postDate": "2018-04-04T20:08:52.473Z", "content": "@Steven
\n\nI was trying it ...but was longer to run . And I was facing the time limit issue .
", "rawMarkdown": "@Steven \n\nI was trying it ...but was longer to run . And I was facing the time limit issue . " }, { "id": 309229, "postDate": "2018-04-04T20:59:54.270Z", "content": "@Andy we switched it to 6GiB. Notice that this is still subject to change (we may need to lower it 500MB or so). Regarding the 6GB limit applying to another (non-GPU) kernel, that is certainly not intended. We also fixed a bug that may relate with that just a few minutes ago. Can you let me know if you see it again, and if you do, post or DM me with details (kernel urls, ...).
", "rawMarkdown": "@Andy we switched it to 6GiB. Notice that this is still subject to change (we may need to lower it 500MB or so). Regarding the 6GB limit applying to another (non-GPU) kernel, that is certainly not intended. We also fixed a bug that may relate with that just a few minutes ago. Can you let me know if you see it again, and if you do, post or DM me with details (kernel urls, ...)." }, { "id": 309230, "postDate": "2018-04-04T21:01:17.713Z", "content": "@Bojan, as indicated in the Settings tab where you enable the GPU, it's a Tesla k80. See specs / info at
\n\nhttps://www.nvidia.com/en-us/data-center/tesla-k80/
", "rawMarkdown": "@Bojan, as indicated in the Settings tab where you enable the GPU, it's a Tesla k80. See specs / info at\n\nhttps://www.nvidia.com/en-us/data-center/tesla-k80/", "votes": 2 }, { "id": 309234, "postDate": "2018-04-04T21:14:58.890Z", "content": "@Seb Nice! Do we actually get the full 24 GB of GDDR5 memory? Because it would be pretty pointless to have that much GPU RAM if our Kernels can only handle 6 GB. :/
", "rawMarkdown": "@Seb Nice! Do we actually get the full 24 GB of GDDR5 memory? Because it would be pretty pointless to have that much GPU RAM if our Kernels can only handle 6 GB. :/", "votes": -1 }, { "id": 309237, "postDate": "2018-04-04T21:29:54.637Z", "content": "@Bojan You get one GPU out of a k80 card. So it's limited to 12GB. The actual limit is 11GB for some reason. I'm not sure it's pointless (in all cases) to have less RAM than VRAM.
", "rawMarkdown": "@Bojan You get one GPU out of a k80 card. So it's limited to 12GB. The actual limit is 11GB for some reason. I'm not sure it's pointless (in all cases) to have less RAM than VRAM." }, { "id": 309241, "postDate": "2018-04-04T21:38:43.350Z", "content": "@Seb, I don't know about \"most cases\", but certainly from everything I've seen the usual workflow for GPU enabled deep learning problems involves getting data into RAM and then pushing it to GPU.
", "rawMarkdown": "@Seb, I don't know about \"most cases\", but certainly from everything I've seen the usual workflow for GPU enabled deep learning problems involves getting data into RAM and then pushing it to GPU. ", "votes": -1 }, { "id": 309257, "postDate": "2018-04-04T22:32:00.667Z", "content": "@Bojan. Thanks for that feedback. I'm trying to find out if there's ever a way to make full use of the GPU memory in this case. At this point we're unfortunately only able to offer 6GiB RAM in this environment.
", "rawMarkdown": "@Bojan. Thanks for that feedback. I'm trying to find out if there's ever a way to make full use of the GPU memory in this case. At this point we're unfortunately only able to offer 6GiB RAM in this environment.", "votes": 1 }, { "id": 309262, "postDate": "2018-04-04T22:40:59.487Z", "content": "Thanks @Seb. I'm not trying to be difficult, just pointing out how our usual ML workflows have been structured. I'd just like to take the full potential of the all available GPU resources in as streamlined way as possible.
", "rawMarkdown": "Thanks @Seb. I'm not trying to be difficult, just pointing out how our usual ML workflows have been structured. I'd just like to take the full potential of the all available GPU resources in as streamlined way as possible. " }, { "id": 309263, "postDate": "2018-04-04T22:47:37.453Z", "content": "I understood it that way, thanks @Bojan and we'll look into it!
", "rawMarkdown": "I understood it that way, thanks @Bojan and we'll look into it!", "votes": 1 }, { "id": 309654, "postDate": "2018-04-05T18:43:35.737Z", "content": "@Bojan This may not be a realistic example or a statement that in practice having more CPU memory wouldn't be a big help, but FWIW, one of our resident experts, @DanB, put together a quick example that shows that running out of the 11GB GPU memory is in fact possible with 6GB of system memory:
\n\nhttps://www.kaggle.com/dansbecker/running-out-of-ram-on-gpu-before-cpu
", "rawMarkdown": "@Bojan This may not be a realistic example or a statement that in practice having more CPU memory wouldn't be a big help, but FWIW, one of our resident experts, @DanB, put together a quick example that shows that running out of the 11GB GPU memory is in fact possible with 6GB of system memory:\n\nhttps://www.kaggle.com/dansbecker/running-out-of-ram-on-gpu-before-cpu", "votes": 4 }, { "id": 309666, "postDate": "2018-04-05T19:15:22.660Z", "content": "I might dispute being called a \"resident expert,\" but I can give a little background on what's happening in that example.
\n\nWhen you fit a model, TensorFlow caches the forward propagated value for every node, for every item in your batch. It then also calculates and stores the gradients at each node. If you are fitting on a GPU, all this is being stored on a GPU.
\n\nFor a big network, that can mean a much bigger storage requirement on the GPU than the the original data.
\n\nYou obviously have a lot of control over what's stored on CPU vs GPU... so CPU RAM may be the limiting factor for your workflow. And I don't know anything about what PyTorch does for memory management.
\n\nBut, fwiw, my experience has been that I frequently hit the GPU limit before the CPU limit.
", "rawMarkdown": "I might dispute being called a \"resident expert,\" but I can give a little background on what's happening in that example.\n\nWhen you fit a model, TensorFlow caches the forward propagated value for every node, for every item in your batch. It then also calculates and stores the gradients at each node. If you are fitting on a GPU, all this is being stored on a GPU.\n\nFor a big network, that can mean a much bigger storage requirement on the GPU than the the original data.\n\nYou obviously have a lot of control over what's stored on CPU vs GPU... so CPU RAM may be the limiting factor for your workflow. And I don't know anything about what PyTorch does for memory management.\n\nBut, fwiw, my experience has been that I frequently hit the GPU limit before the CPU limit.", "votes": 3 }, { "id": 309839, "postDate": "2018-04-06T03:42:39.023Z", "content": "Thanks for the info. I just tried re-running one of my Toxic kernels with GPU support and execution time went from 2400 seconds to just over 400!
\n\nhttps://www.kaggle.com/tunguz/new-script-with-gpu/code
", "rawMarkdown": "Thanks for the info. I just tried re-running one of my Toxic kernels with GPU support and execution time went from 2400 seconds to just over 400! \n\nhttps://www.kaggle.com/tunguz/new-script-with-gpu/code", "votes": 3 } ] }, { "id": 309251, "postDate": "2018-04-04T22:14:26.113Z", "content": "On the first day we are given GPU access we are complaining about memory and resources being useless or not being enough to run multiple GPU-jobs simultaneously. And to think that literally yesterday we couldn't run any GPU-based jobs.
\n\nPerspective, anyone?
", "rawMarkdown": "On the first day we are given GPU access we are complaining about memory and resources being useless or not being enough to run multiple GPU-jobs simultaneously. And to think that literally yesterday we couldn't run any GPU-based jobs.\n\nPerspective, anyone?", "votes": 6, "replies": [ { "id": 309259, "postDate": "2018-04-04T22:35:03.853Z", "content": "Was anyone complaining about resources \"not being enough to run multiple GPU-jobs simultaneously\"? The problem for me was that I couldn't run CPU jobs simultaneously with a single GPU job (but that was probably a bug), and I couldn't do anything significant with a GPU without rebuilding my workflow from the ground up. So at this point the GPU is just something to play with. Hopefully the bug is fixed, and I can play with the GPU while continuing what I was doing in CPU jobs. But it's weird to have all this power and still have a major bottleneck in the system RAM that makes the power largely inaccessible.
", "rawMarkdown": "Was anyone complaining about resources \"not being enough to run multiple GPU-jobs simultaneously\"? The problem for me was that I couldn't run CPU jobs simultaneously with a single GPU job (but that was probably a bug), and I couldn't do anything significant with a GPU without rebuilding my workflow from the ground up. So at this point the GPU is just something to play with. Hopefully the bug is fixed, and I can play with the GPU while continuing what I was doing in CPU jobs. But it's weird to have all this power and still have a major bottleneck in the system RAM that makes the power largely inaccessible.", "votes": 1 }, { "id": 309261, "postDate": "2018-04-04T22:39:16.740Z", "content": "\n\n\nBut it's weird to have all this power and still have a major bottleneck in the system RAM that makes the power largely inaccessible.
\n
That doesn't strike me as the weirdest thing going on today in this thread. Not by a long shot.
", "rawMarkdown": "> But it's weird to have all this power and still have a major bottleneck in the system RAM that makes the power largely inaccessible.\n\nThat doesn't strike me as the weirdest thing going on today in this thread. Not by a long shot." } ] }, { "id": 309130, "postDate": "2018-04-04T17:29:29.067Z", "content": "Hi all,
\n\nThis morning only I observed that on the console in Kaggle kernels, GPU is mentioned as \"Off\". Kaggle kernels have never had this GPU thing mentioned altogether before.
\n\nThe following information is available:\nCPU 0% GPU OFF `RAM 214MB/ 17.2 GB Disk 55.4MB/1GB
\n\nI am interested in knowing how to turn this GPU on? \nAny idea?
", "rawMarkdown": "Hi all, \n\nThis morning only I observed that on the console in Kaggle kernels, GPU is mentioned as \"Off\". Kaggle kernels have never had this GPU thing mentioned altogether before.\n\nThe following information is available:\nCPU 0% GPU OFF `RAM 214MB/ 17.2 GB Disk 55.4MB/1GB \n\nI am interested in knowing how to turn this GPU on? \nAny idea?\n", "votes": 6 }, { "id": 309242, "postDate": "2018-04-04T21:41:08.187Z", "content": "Great news, I do not have any GPU (non-CPU integrated) on my laptop, and I have not started to play with AWS. I am glad I can start toying with them here.
", "rawMarkdown": "Great news, I do not have any GPU (non-CPU integrated) on my laptop, and I have not started to play with AWS. I am glad I can start toying with them here." }, { "id": 309688, "postDate": "2018-04-05T20:03:39.740Z", "content": "Apparently it works.
", "rawMarkdown": "Apparently [it works][1].\n\n [1]: https://www.kaggle.com/aharless/talkingdata-gpu-example-from-prepared-data", "votes": 1, "replies": [ { "id": 310257, "postDate": "2018-04-06T23:54:05.133Z", "content": "Turns out it works quite well. You can do a lot with a good GPU even with only a little bit of system memory.
", "rawMarkdown": "Turns out [it works quite well][1]. You can do a lot with a good GPU even with only a little bit of system memory.\n\n [1]: https://www.kaggle.com/aharless/talkingdata-gpu-example-with-multiple-runs", "votes": 2 } ] }, { "id": 499869, "postDate": "2019-03-25T10:38:56.710Z", "content": "Yes, go to the settings and slide GPU to on and it will show you the GPU in On State at bottom right corner of console. It is NVIDIA-K80 which 12.5X faster than usual CPUs.
", "rawMarkdown": "Yes, go to the settings and slide GPU to on and it will show you the GPU in On State at bottom right corner of console. It is NVIDIA-K80 which 12.5X faster than usual CPUs. " }, { "id": 309233, "postDate": "2018-04-04T21:13:30.593Z", "content": "I see the Settings tab, but when I select it, nothing happens. The other tabs are working fine. Is it happening to me only? I am using the Safari browser on a iMac.
", "rawMarkdown": "I see the Settings tab, but when I select it, nothing happens. The other tabs are working fine. Is it happening to me only? I am using the Safari browser on a iMac.", "replies": [ { "id": 309239, "postDate": "2018-04-04T21:34:59.550Z", "content": "Your interactive session should automatically restart when you check that box, in an environment with a GPU available. You can run !nvidia-smi in a Python kernel to see info on it. If that doesn't work for you, let me know (and include the URL to your kernel please).
So we can run very deep, very wide networks with fairly large batches, but only on fairly small data files? Maybe an occasion to start working with undersampled negatives, I guess..
", "rawMarkdown": "So we can run very deep, very wide networks with fairly large batches, but only on fairly small data files? Maybe an occasion to start working with undersampled negatives, I guess.." }, { "id": 539425, "postDate": "2019-05-30T03:29:41.323Z", "content": "!nvidia-smi not running on my kernel\nhttps://www.kaggle.com/rbales/cactus-detection-fastai
", "rawMarkdown": "!nvidia-smi not running on my kernel\nhttps://www.kaggle.com/rbales/cactus-detection-fastai" } ] } ], "comments": [ { "id": 309141, "author_name": "Ben Hamner", "author_url": "", "post_date": "2018-04-04T17:57:15.320000", "content": "You can add a GPU from the \"Settings\" tab in the kernel (expand the menu on the right)
", "votes": 7, "replies": [ { "id": 309155, "author_name": "Serigne ", "author_url": "", "post_date": "2018-04-04T18:22:56.537000", "content": "That's awesome Ben...Thanks !
\n\nCould we have more RAM for this competition ? :)
\n\nIt becomes nearly impossible to test my new features due to lack of RAM :p
", "votes": 4, "replies": [] }, { "id": 309156, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2018-04-04T18:29:28.210000", "content": "Any info on what kind of GPU is it? i.e., the amount of GPU RAM, how many Tflops, etc.
", "votes": 9, "replies": [] }, { "id": 309159, "author_name": "", "author_url": "", "post_date": "2018-04-04T18:39:36.280000", "content": "", "votes": 0, "replies": [] }, { "id": 309164, "author_name": "Alan Khoa Nguyen", "author_url": "", "post_date": "2018-04-04T18:46:05.547000", "content": "\n\n\nSerigne wrote
\n \n\n\n \nThat's awesome Ben...
\nCould we have more RAM for this competition ? :)
\n \nIt becomes nearly impossible to test my new features due to lack of RAM :p
\n
You can shard your df, blend the different models, and get the same result.
", "votes": 0, "replies": [] }, { "id": 309185, "author_name": "Andy Harless", "author_url": "", "post_date": "2018-04-04T19:20:00.603000", "content": "When I tried using GPU, it reduced my memory limit to 6 GB (oddly using the non-digital definition of \"giga\" as 10^9 rather than 2^30). Also when I tried to run another kernel at the same time, it applied a 6GB limit to that kernel too. If that limit applies, it makes the GPU pretty useless for this particular competition. And if the limit applies to all concurrently running kernels, it means we can't experiment while running something else. (6 GB would make sense as a limit on GPU memory, but I don't see any reason to limit system memory that way. I'm wondering if there's some kind of bug here.)
", "votes": 1, "replies": [] }, { "id": 309187, "author_name": "Gaurav Arora", "author_url": "", "post_date": "2018-04-04T19:24:59.373000", "content": "Even, I have faced the same issue. However, Nvidia Tesla k80 would be a great add on for this competion.
", "votes": 0, "replies": [] }, { "id": 309211, "author_name": "Serigne ", "author_url": "", "post_date": "2018-04-04T20:08:52.473000", "content": "@Steven
\n\nI was trying it ...but was longer to run . And I was facing the time limit issue .
", "votes": 0, "replies": [] }, { "id": 309229, "author_name": "Seb Boving", "author_url": "", "post_date": "2018-04-04T20:59:54.270000", "content": "@Andy we switched it to 6GiB. Notice that this is still subject to change (we may need to lower it 500MB or so). Regarding the 6GB limit applying to another (non-GPU) kernel, that is certainly not intended. We also fixed a bug that may relate with that just a few minutes ago. Can you let me know if you see it again, and if you do, post or DM me with details (kernel urls, ...).
", "votes": 0, "replies": [] }, { "id": 309230, "author_name": "Seb Boving", "author_url": "", "post_date": "2018-04-04T21:01:17.713000", "content": "@Bojan, as indicated in the Settings tab where you enable the GPU, it's a Tesla k80. See specs / info at
\n\nhttps://www.nvidia.com/en-us/data-center/tesla-k80/
", "votes": 2, "replies": [] }, { "id": 309234, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2018-04-04T21:14:58.890000", "content": "@Seb Nice! Do we actually get the full 24 GB of GDDR5 memory? Because it would be pretty pointless to have that much GPU RAM if our Kernels can only handle 6 GB. :/
", "votes": -1, "replies": [] }, { "id": 309237, "author_name": "Seb Boving", "author_url": "", "post_date": "2018-04-04T21:29:54.637000", "content": "@Bojan You get one GPU out of a k80 card. So it's limited to 12GB. The actual limit is 11GB for some reason. I'm not sure it's pointless (in all cases) to have less RAM than VRAM.
", "votes": 0, "replies": [] }, { "id": 309241, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2018-04-04T21:38:43.350000", "content": "@Seb, I don't know about \"most cases\", but certainly from everything I've seen the usual workflow for GPU enabled deep learning problems involves getting data into RAM and then pushing it to GPU.
", "votes": -1, "replies": [] }, { "id": 309257, "author_name": "Seb Boving", "author_url": "", "post_date": "2018-04-04T22:32:00.667000", "content": "@Bojan. Thanks for that feedback. I'm trying to find out if there's ever a way to make full use of the GPU memory in this case. At this point we're unfortunately only able to offer 6GiB RAM in this environment.
", "votes": 1, "replies": [] }, { "id": 309262, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2018-04-04T22:40:59.487000", "content": "Thanks @Seb. I'm not trying to be difficult, just pointing out how our usual ML workflows have been structured. I'd just like to take the full potential of the all available GPU resources in as streamlined way as possible.
", "votes": 0, "replies": [] }, { "id": 309263, "author_name": "Seb Boving", "author_url": "", "post_date": "2018-04-04T22:47:37.453000", "content": "I understood it that way, thanks @Bojan and we'll look into it!
", "votes": 1, "replies": [] }, { "id": 309654, "author_name": "Seb Boving", "author_url": "", "post_date": "2018-04-05T18:43:35.737000", "content": "@Bojan This may not be a realistic example or a statement that in practice having more CPU memory wouldn't be a big help, but FWIW, one of our resident experts, @DanB, put together a quick example that shows that running out of the 11GB GPU memory is in fact possible with 6GB of system memory:
\n\nhttps://www.kaggle.com/dansbecker/running-out-of-ram-on-gpu-before-cpu
", "votes": 4, "replies": [] }, { "id": 309666, "author_name": "DanB", "author_url": "", "post_date": "2018-04-05T19:15:22.660000", "content": "I might dispute being called a \"resident expert,\" but I can give a little background on what's happening in that example.
\n\nWhen you fit a model, TensorFlow caches the forward propagated value for every node, for every item in your batch. It then also calculates and stores the gradients at each node. If you are fitting on a GPU, all this is being stored on a GPU.
\n\nFor a big network, that can mean a much bigger storage requirement on the GPU than the the original data.
\n\nYou obviously have a lot of control over what's stored on CPU vs GPU... so CPU RAM may be the limiting factor for your workflow. And I don't know anything about what PyTorch does for memory management.
\n\nBut, fwiw, my experience has been that I frequently hit the GPU limit before the CPU limit.
", "votes": 3, "replies": [] }, { "id": 309839, "author_name": "Bojan Tunguz", "author_url": "", "post_date": "2018-04-06T03:42:39.023000", "content": "Thanks for the info. I just tried re-running one of my Toxic kernels with GPU support and execution time went from 2400 seconds to just over 400!
\n\nhttps://www.kaggle.com/tunguz/new-script-with-gpu/code
", "votes": 3, "replies": [] } ] }, { "id": 309251, "author_name": "Tilii", "author_url": "", "post_date": "2018-04-04T22:14:26.113000", "content": "On the first day we are given GPU access we are complaining about memory and resources being useless or not being enough to run multiple GPU-jobs simultaneously. And to think that literally yesterday we couldn't run any GPU-based jobs.
\n\nPerspective, anyone?
", "votes": 6, "replies": [ { "id": 309259, "author_name": "Andy Harless", "author_url": "", "post_date": "2018-04-04T22:35:03.853000", "content": "Was anyone complaining about resources \"not being enough to run multiple GPU-jobs simultaneously\"? The problem for me was that I couldn't run CPU jobs simultaneously with a single GPU job (but that was probably a bug), and I couldn't do anything significant with a GPU without rebuilding my workflow from the ground up. So at this point the GPU is just something to play with. Hopefully the bug is fixed, and I can play with the GPU while continuing what I was doing in CPU jobs. But it's weird to have all this power and still have a major bottleneck in the system RAM that makes the power largely inaccessible.
", "votes": 1, "replies": [] }, { "id": 309261, "author_name": "Tilii", "author_url": "", "post_date": "2018-04-04T22:39:16.740000", "content": "\n\n\nBut it's weird to have all this power and still have a major bottleneck in the system RAM that makes the power largely inaccessible.
\n
That doesn't strike me as the weirdest thing going on today in this thread. Not by a long shot.
", "votes": 0, "replies": [] } ] }, { "id": 309242, "author_name": "giim", "author_url": "", "post_date": "2018-04-04T21:41:08.187000", "content": "Great news, I do not have any GPU (non-CPU integrated) on my laptop, and I have not started to play with AWS. I am glad I can start toying with them here.
", "votes": 0, "replies": [] }, { "id": 309688, "author_name": "Andy Harless", "author_url": "", "post_date": "2018-04-05T20:03:39.740000", "content": "Apparently it works.
", "votes": 1, "replies": [ { "id": 310257, "author_name": "Andy Harless", "author_url": "", "post_date": "2018-04-06T23:54:05.133000", "content": "Turns out it works quite well. You can do a lot with a good GPU even with only a little bit of system memory.
", "votes": 2, "replies": [] } ] }, { "id": 499869, "author_name": "SamDesai", "author_url": "", "post_date": "2019-03-25T10:38:56.710000", "content": "Yes, go to the settings and slide GPU to on and it will show you the GPU in On State at bottom right corner of console. It is NVIDIA-K80 which 12.5X faster than usual CPUs.
", "votes": 0, "replies": [] }, { "id": 309233, "author_name": "FabrizioFedericoni", "author_url": "", "post_date": "2018-04-04T21:13:30.593000", "content": "I see the Settings tab, but when I select it, nothing happens. The other tabs are working fine. Is it happening to me only? I am using the Safari browser on a iMac.
", "votes": 0, "replies": [ { "id": 309239, "author_name": "Seb Boving", "author_url": "", "post_date": "2018-04-04T21:34:59.550000", "content": "Your interactive session should automatically restart when you check that box, in an environment with a GPU available. You can run !nvidia-smi in a Python kernel to see info on it. If that doesn't work for you, let me know (and include the URL to your kernel please).
So we can run very deep, very wide networks with fairly large batches, but only on fairly small data files? Maybe an occasion to start working with undersampled negatives, I guess..
", "votes": 0, "replies": [] }, { "id": 539425, "author_name": "Ryan Bales", "author_url": "", "post_date": "2019-05-30T03:29:41.323000", "content": "!nvidia-smi not running on my kernel\nhttps://www.kaggle.com/rbales/cactus-detection-fastai
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "309141": "You can add a GPU from the \"Settings\" tab in the kernel (expand the menu on the right)", "309251": "On the first day we are given GPU access we are complaining about memory and resources being useless or not being enough to run multiple GPU-jobs simultaneously. And to think that literally yesterday we couldn't run any GPU-based jobs.\n\nPerspective, anyone?", "309130": "Hi all, \n\nThis morning only I observed that on the console in Kaggle kernels, GPU is mentioned as \"Off\". Kaggle kernels have never had this GPU thing mentioned altogether before.\n\nThe following information is available:\nCPU 0% GPU OFF `RAM 214MB/ 17.2 GB Disk 55.4MB/1GB \n\nI am interested in knowing how to turn this GPU on? \nAny idea?\n", "309242": "Great news, I do not have any GPU (non-CPU integrated) on my laptop, and I have not started to play with AWS. I am glad I can start toying with them here.", "309688": "Apparently [it works][1].\n\n [1]: https://www.kaggle.com/aharless/talkingdata-gpu-example-from-prepared-data", "499869": "Yes, go to the settings and slide GPU to on and it will show you the GPU in On State at bottom right corner of console. It is NVIDIA-K80 which 12.5X faster than usual CPUs. ", "309233": "I see the Settings tab, but when I select it, nothing happens. The other tabs are working fine. Is it happening to me only? I am using the Safari browser on a iMac." } }