![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F494641%2F81207b4bcb6f2949804091a4fed6a235%2Fgpu_low.png?generation=1707576020679909&alt=media\")
This is often the picture that I see:
\n
Sometimes GPU utilization jumps to 7% at maximum.
\nThese are the parameters I use:
\n = {\n : ,\n : ,\n : ,\n : ,\n : ,\n : ,\n : , \n : ,\n : ,\n : ,\n : \n}\nIt seems to me it does not utilize GPU. Do you experience something similar?
", "messages": [ { "id": "2645915", "postDate": "02/10/2024 14:44:06", "content": "This is often the picture that I see:
\n
Sometimes GPU utilization jumps to 7% at maximum.
\nThese are the parameters I use:
\n = {\n : ,\n : ,\n : ,\n : ,\n : ,\n : ,\n : , \n : ,\n : ,\n : ,\n : \n}\nIt seems to me it does not utilize GPU. Do you experience something similar?
", "rawMarkdown": "This is often the picture that I see:\n\n\n\nSometimes GPU utilization jumps to 7% at maximum.\n\nThese are the parameters I use:\n```\nparams = {\n \"boosting_type\": \"gbdt\",\n \"objective\": \"binary\",\n \"metric\": \"auc\",\n \"max_depth\": 8,\n \"learning_rate\": 0.1,\n \"n_estimators\": 100,\n \"colsample_bytree\": 0.8, \n \"colsample_bynode\": 0.8,\n \"verbose\": -1,\n \"random_state\": 42,\n \"device\": 'gpu'\n}\n```\n\nIt seems to me it does not utilize GPU. Do you experience something similar?", "votes": null }, { "id": "2645938", "postDate": "02/10/2024 14:56:44", "content": "yup, same here
", "rawMarkdown": "yup, same here", "votes": null }, { "id": "2646072", "postDate": "02/10/2024 16:15:44", "content": "For the LGB GPU model, you need to reduce the num_threads, from my experience, use 2 threads fewer than all of your available threads.
If you need more speed up (lower performance), you can change max_bin and gpu_use_dp.
https://lightgbm.readthedocs.io/en/latest/GPU-Performance.html
", "rawMarkdown": "For the LGB GPU model, you need to reduce the `num_threads`, from my experience, use 2 threads fewer than all of your available threads.\n\nIf you need more speed up (lower performance), you can change `max_bin` and `gpu_use_dp`.\n\nhttps://lightgbm.readthedocs.io/en/latest/GPU-Performance.html", "votes": null }, { "id": "2646400", "postDate": "02/10/2024 21:26:38", "content": "Thanks, I will try it
", "rawMarkdown": "Thanks, I will try it", "votes": null }, { "id": "2648890", "postDate": "02/12/2024 13:34:58", "content": "I did try it, however it did not change almost anything in my case.
", "rawMarkdown": "I did try it, however it did not change almost anything in my case.", "votes": null }, { "id": "2648933", "postDate": "02/12/2024 14:01:50", "content": "maybe you can also try using the actual number cores instead of threads.
\nFrom the link above:
\nDuring on CPU we used only physical cores of the CPU, not use hyper-threading cores, we found that using too many threads actually makes performance worse\nFor example, most CPUs nowadays have 2 threads per core, for a 2 core 4 thread machine, use num_threads=2.
\nI haven't tried this on Kaggle, but this has been my experience with my local machine.
", "rawMarkdown": "maybe you can also try using the actual number cores instead of threads.\n\nFrom the link above:\n```\nDuring benchmarking on CPU we used only 28 physical cores of the CPU, and did not use hyper-threading cores, because we found that using too many threads actually makes performance worse\n```\nFor example, most CPUs nowadays have 2 threads per core, for a 2 core 4 thread machine, use num_threads=2.\n\nI haven't tried this on Kaggle, but this has been my experience with my local machine.", "votes": null }, { "id": "2749236", "postDate": "04/12/2024 23:43:12", "content": "Did you find any explanation or solution for this?
", "rawMarkdown": "Did you find any explanation or solution for this?", "votes": null }, { "id": "2749472", "postDate": "04/13/2024 04:33:37", "content": "Unfortunately not
", "rawMarkdown": "Unfortunately not", "votes": null } ], "comments": [ { "id": 2645938, "author_name": "ags299", "author_url": "", "post_date": "02/10/2024 14:56:44", "content": "yup, same here
", "votes": null, "replies": [] }, { "id": 2646072, "author_name": "kingychiu", "author_url": "", "post_date": "02/10/2024 16:15:44", "content": "For the LGB GPU model, you need to reduce the num_threads, from my experience, use 2 threads fewer than all of your available threads.
If you need more speed up (lower performance), you can change max_bin and gpu_use_dp.
https://lightgbm.readthedocs.io/en/latest/GPU-Performance.html
", "votes": null, "replies": [ { "id": 2646400, "author_name": "narsil", "author_url": "", "post_date": "02/10/2024 21:26:38", "content": "Thanks, I will try it
", "votes": null, "replies": [ { "id": 2648890, "author_name": "narsil", "author_url": "", "post_date": "02/12/2024 13:34:58", "content": "I did try it, however it did not change almost anything in my case.
", "votes": null, "replies": [ { "id": 2648933, "author_name": "kingychiu", "author_url": "", "post_date": "02/12/2024 14:01:50", "content": "maybe you can also try using the actual number cores instead of threads.
\nFrom the link above:
\nDuring on CPU we used only physical cores of the CPU, not use hyper-threading cores, we found that using too many threads actually makes performance worse\nFor example, most CPUs nowadays have 2 threads per core, for a 2 core 4 thread machine, use num_threads=2.
\nI haven't tried this on Kaggle, but this has been my experience with my local machine.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2749236, "author_name": "manuelandersen", "author_url": "", "post_date": "04/12/2024 23:43:12", "content": "Did you find any explanation or solution for this?
", "votes": null, "replies": [ { "id": 2749472, "author_name": "narsil", "author_url": "", "post_date": "04/13/2024 04:33:37", "content": "Unfortunately not
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2645915": "This is often the picture that I see:\n\n\n\nSometimes GPU utilization jumps to 7% at maximum.\n\nThese are the parameters I use:\n```\nparams = {\n \"boosting_type\": \"gbdt\",\n \"objective\": \"binary\",\n \"metric\": \"auc\",\n \"max_depth\": 8,\n \"learning_rate\": 0.1,\n \"n_estimators\": 100,\n \"colsample_bytree\": 0.8, \n \"colsample_bynode\": 0.8,\n \"verbose\": -1,\n \"random_state\": 42,\n \"device\": 'gpu'\n}\n```\n\nIt seems to me it does not utilize GPU. Do you experience something similar?", "2645938": "yup, same here", "2646072": "For the LGB GPU model, you need to reduce the `num_threads`, from my experience, use 2 threads fewer than all of your available threads.\n\nIf you need more speed up (lower performance), you can change `max_bin` and `gpu_use_dp`.\n\nhttps://lightgbm.readthedocs.io/en/latest/GPU-Performance.html", "2646400": "Thanks, I will try it", "2648890": "I did try it, however it did not change almost anything in my case.", "2648933": "maybe you can also try using the actual number cores instead of threads.\n\nFrom the link above:\n```\nDuring benchmarking on CPU we used only 28 physical cores of the CPU, and did not use hyper-threading cores, because we found that using too many threads actually makes performance worse\n```\nFor example, most CPUs nowadays have 2 threads per core, for a 2 core 4 thread machine, use num_threads=2.\n\nI haven't tried this on Kaggle, but this has been my experience with my local machine.", "2749236": "Did you find any explanation or solution for this?", "2749472": "Unfortunately not" }, "source": "meta" }