Hi everyone,
\n\nCan you tell me which hardware (CPU, RAM, SSD, GPU) are you using for this competition?
\n\nI'm currently thinking of buying a new computer for Kaggle competitions, but one question remains: Which CPU is best between the Threadripper 1950X and the Core i9-7900X (both at $999) for running numpy / scikit-learn / LightGBM / XGBoost computations in a heavily multi-threaded setup?
", "messages": [ { "id": "221887", "postDate": "09/16/2017 17:00:25", "content": "Hi everyone,
\n\nCan you tell me which hardware (CPU, RAM, SSD, GPU) are you using for this competition?
\n\nI'm currently thinking of buying a new computer for Kaggle competitions, but one question remains: Which CPU is best between the Threadripper 1950X and the Core i9-7900X (both at $999) for running numpy / scikit-learn / LightGBM / XGBoost computations in a heavily multi-threaded setup?
", "rawMarkdown": "Hi everyone,\n\nCan you tell me which hardware (CPU, RAM, SSD, GPU) are you using for this competition?\n\nI'm currently thinking of buying a new computer for Kaggle competitions, but one question remains: Which CPU is best between the Threadripper 1950X and the Core i9-7900X (both at $999) for running numpy / scikit-learn / LightGBM / XGBoost computations in a heavily multi-threaded setup?", "votes": null }, { "id": "221904", "postDate": "09/16/2017 18:18:04", "content": "Threadripper is faster for parallel workloads (which for me has been the case in Kaggle competitions so far) and has 64 instead of 48 PCIe-lanes.\nI would pick Threadripper :)
\n\n<-- Ryzen 1700 + GTX1080TI
", "rawMarkdown": "Threadripper is faster for parallel workloads (which for me has been the case in Kaggle competitions so far) and has 64 instead of 48 PCIe-lanes.\nI would pick Threadripper :)\n\n<-- Ryzen 1700 + GTX1080TI", "votes": null }, { "id": "221906", "postDate": "09/16/2017 18:31:07", "content": "I am far from being an expert, I am waitinag for more informed opinions too.
\n\nHere are my 2c:
\n\nFirst of all if you want to do deep learning you should invest in a GPU, and don't worry much about the CPU. If you can wait, then wait for the Volta architecture based cards which was designed with DL in mind, supposedly they are expected very soon. If you are in a hurry, then the current sweet spot for price/performance is between GTX 1070 and GTX 1080, depending on what deal you can find. Otherwise, if you are looking for the best, then look for 1080Ti < Titan X < Titan Xp. Arguably the price premium is not justified.
\n\nBetween 1950X and 7900X it is a difficult choice. 1950X is much faster in multi core applications, but slower per core. The speed boost diminishes with each core, and 20 and 32 threads are a lot. Also, there are other operations that are single threaded, so when all is said and done it is difficult to predict which one will end up on top.
\n\nAlso, LightGBM and XGBoost both support GPU training.
\n\nFor SSD I have a separate SSD for my system and important files, and RAID 0 for the data. My option is to go for cheap SSDs in RAID 0, and add as many as I need. I have 2 now, but I'm thinking of adding another 2.
", "rawMarkdown": "I am far from being an expert, I am waitinag for more informed opinions too.\n\nHere are my 2c:\n\nFirst of all if you want to do deep learning you should invest in a GPU, and don't worry much about the CPU. If you can wait, then wait for the Volta architecture based cards which was designed with DL in mind, supposedly they are expected very soon. If you are in a hurry, then the current sweet spot for price/performance is between GTX 1070 and GTX 1080, depending on what deal you can find. Otherwise, if you are looking for the best, then look for 1080Ti < Titan X < Titan Xp. Arguably the price premium is not justified. \n\nBetween 1950X and 7900X it is a difficult choice. 1950X is much faster in multi core applications, but slower per core. The speed boost diminishes with each core, and 20 and 32 threads are a lot. Also, there are other operations that are single threaded, so when all is said and done it is difficult to predict which one will end up on top.\n\nAlso, LightGBM and XGBoost both support GPU training.\n\nFor SSD I have a separate SSD for my system and important files, and RAID 0 for the data. My option is to go for cheap SSDs in RAID 0, and add as many as I need. I have 2 now, but I'm thinking of adding another 2.", "votes": null }, { "id": "221907", "postDate": "09/16/2017 18:48:30", "content": "Thanks for your answers.
\n\nCurrently, I have an old Core i7 2600k with 16GB of RAM and a GTX1070. I'm planning to buy a new PC near Xmas 2017 or Q1 2018.
\n\nMaybe I'll wait for Volta GPUs, but Nvidia said they won't be launched until 2018. So, maybe I'll think about taking two GTX1080Ti or one Volta card (If we consider that Volta cards will be more expensive and that we can get rebates on GTX1080Ti).
\n\nI've tried XGBoost on GPU few months ago, but the performances are not as good as the CPU version. So, I stick to CPU for Gradient boosting algorithms at the moment. So, CPU is also important.
\n\nMoreover, I've noticed that when I do Deep Learning on my GTX1070, my CPU runs at 100% and the GPU is only used at 70% max, because the CPU is needed to do synchronization tasks between the GPU SMs.
\n\nFinally, for the choice of CPU, one question remains about using scikit-learn / numpy: as numpy relies on Intel's MKL library, will the performances be lower on AMD CPUs due to the fact that MKL is not supported?
", "rawMarkdown": "Thanks for your answers.\n\nCurrently, I have an old Core i7 2600k with 16GB of RAM and a GTX1070. I'm planning to buy a new PC near Xmas 2017 or Q1 2018.\n\nMaybe I'll wait for Volta GPUs, but Nvidia said they won't be launched until 2018. So, maybe I'll think about taking two GTX1080Ti or one Volta card (If we consider that Volta cards will be more expensive and that we can get rebates on GTX1080Ti).\n\nI've tried XGBoost on GPU few months ago, but the performances are not as good as the CPU version. So, I stick to CPU for Gradient boosting algorithms at the moment. So, CPU is also important.\n\nMoreover, I've noticed that when I do Deep Learning on my GTX1070, my CPU runs at 100% and the GPU is only used at 70% max, because the CPU is needed to do synchronization tasks between the GPU SMs.\n\nFinally, for the choice of CPU, one question remains about using scikit-learn / numpy: as numpy relies on Intel's MKL library, will the performances be lower on AMD CPUs due to the fact that MKL is not supported?", "votes": null }, { "id": "222031", "postDate": "09/17/2017 09:38:30", "content": "Make sure you get a m2 SSD as I/O is often what limits GPU use.
", "rawMarkdown": "Make sure you get a m2 SSD as I/O is often what limits GPU use.", "votes": null }, { "id": "222053", "postDate": "09/17/2017 10:25:53", "content": "Here is a great post you may find useful http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/
", "rawMarkdown": "Here is a great post you may find useful http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/", "votes": null }, { "id": "255131", "postDate": "12/08/2017 11:56:21", "content": "HI @thomasseleck
\n\nWhich CPU that you chose, and why?
", "rawMarkdown": "HI @thomasseleck \n\nWhich CPU that you chose, and why?", "votes": null }, { "id": "255149", "postDate": "12/08/2017 13:04:30", "content": "Hi @M Hendra Herviawan,
\n\nI've bought the AMD Threadripper 1950X for the following reasons:
\n\nDrawbacks:\nFor the moment, I have some BSOD on Windows 10 only at boot (when arriving on the desktop), when virtualization is enabled in BIOS (SVM/IOMMU setting).
", "rawMarkdown": "Hi @M Hendra Herviawan,\n\nI've bought the AMD Threadripper 1950X for the following reasons:\n\n - Same performances as Core i9 7960X but cheaper\n - 16 cores instead of 10 for the same price ( against the Core i9 7900X)\n - No heating and VRM issues (with Enermax Liqtech TR4 cooler, my CPU does not go beyond 55 Celsius degrees)\n - More PCIe lanes for NVMe drives or GPUs\n\nDrawbacks:\nFor the moment, I have some BSOD on Windows 10 only at boot (when arriving on the desktop), when virtualization is enabled in BIOS (SVM/IOMMU setting).", "votes": null }, { "id": "353304", "postDate": "07/06/2018 11:53:51", "content": "EDIT: never mind... I read the topic top down and didn't notice it was sorted in reverse time order.... I just found the answer for my question.
\n\nHi @Thomas!
\n\nThis is a little bit old thread, but the topic is of my interest....
\n\nI'm facing the same problem that you said: choosing a good CPU mostly for XGBoost and LightGBM models.
\n\nFurthermore, I have the same hardware configuration you had: an i7 2600k with 16 Gb and a more recent GPU.
\n\nDo you mind to share your decision on which CPU to buy? Did you went to Intel or AMD? How fast it is, compared to the i7 2600k?
\n\nThanks for your time!
", "rawMarkdown": "*EDIT: never mind... I read the topic top down and didn't notice it was sorted in reverse time order.... I just found the answer for my question.*\n\n\nHi @Thomas!\n\nThis is a little bit old thread, but the topic is of my interest....\n\nI'm facing the same problem that you said: choosing a good CPU mostly for XGBoost and LightGBM models.\n\nFurthermore, I have the same hardware configuration you had: an i7 2600k with 16 Gb and a more recent GPU.\n\nDo you mind to share your decision on which CPU to buy? Did you went to Intel or AMD? How fast it is, compared to the i7 2600k?\n\nThanks for your time!", "votes": null }, { "id": "353403", "postDate": "07/06/2018 17:06:12", "content": "I bought the Threadripper 1950X. It's nearly 5 times the power of the i7 2600k. It's great, especially for LightGBM, where GPU doesn't bring a significant speedup.
\n\nAnd if you can wait a little bit, AMD will release a 32 cores version of Threadripper that will be much better: https://www.techpowerup.com/245570/amd-ryzen-threadripper-2990x-makes-an-appearance-on-3dmark
", "rawMarkdown": "I bought the Threadripper 1950X. It's nearly 5 times the power of the i7 2600k. It's great, especially for LightGBM, where GPU doesn't bring a significant speedup.\n\nAnd if you can wait a little bit, AMD will release a 32 cores version of Threadripper that will be much better: https://www.techpowerup.com/245570/amd-ryzen-threadripper-2990x-makes-an-appearance-on-3dmark", "votes": null }, { "id": "353518", "postDate": "07/06/2018 23:26:20", "content": "Look like TH have memory compatibility issue, which memory type & speed do you use ?
", "rawMarkdown": "Look like TH have memory compatibility issue, which memory type & speed do you use ?", "votes": null }, { "id": "353657", "postDate": "07/07/2018 10:57:29", "content": "I use the G.Skill F4-2400C15Q-64GFX memory (64 GB DDR4 @ 2400 MHz). Note that I solved the BSOD problem. It was an issue related to an immature BIOS / UEFI that created BSODs at Windows startup. I upgraded the BIOS to the last version (601 for ASUS Prime X399-A) and I don't have BSODs anymore.
\n\nI also uninstall ASUS AURA (software that controls MB and GPU useless RGB LEDs) and ASUS AI Suite 3, both known to cause BSODs.
", "rawMarkdown": "I use the G.Skill F4-2400C15Q-64GFX memory (64 GB DDR4 @ 2400 MHz). Note that I solved the BSOD problem. It was an issue related to an immature BIOS / UEFI that created BSODs at Windows startup. I upgraded the BIOS to the last version (601 for ASUS Prime X399-A) and I don't have BSODs anymore.\n\nI also uninstall ASUS AURA (software that controls MB and GPU useless RGB LEDs) and ASUS AI Suite 3, both known to cause BSODs.", "votes": null } ], "comments": [ { "id": 221904, "author_name": "timjoseph", "author_url": "", "post_date": "09/16/2017 18:18:04", "content": "Threadripper is faster for parallel workloads (which for me has been the case in Kaggle competitions so far) and has 64 instead of 48 PCIe-lanes.\nI would pick Threadripper :)
\n\n<-- Ryzen 1700 + GTX1080TI
", "votes": null, "replies": [] }, { "id": 221906, "author_name": "radustoicescu", "author_url": "", "post_date": "09/16/2017 18:31:07", "content": "I am far from being an expert, I am waitinag for more informed opinions too.
\n\nHere are my 2c:
\n\nFirst of all if you want to do deep learning you should invest in a GPU, and don't worry much about the CPU. If you can wait, then wait for the Volta architecture based cards which was designed with DL in mind, supposedly they are expected very soon. If you are in a hurry, then the current sweet spot for price/performance is between GTX 1070 and GTX 1080, depending on what deal you can find. Otherwise, if you are looking for the best, then look for 1080Ti < Titan X < Titan Xp. Arguably the price premium is not justified.
\n\nBetween 1950X and 7900X it is a difficult choice. 1950X is much faster in multi core applications, but slower per core. The speed boost diminishes with each core, and 20 and 32 threads are a lot. Also, there are other operations that are single threaded, so when all is said and done it is difficult to predict which one will end up on top.
\n\nAlso, LightGBM and XGBoost both support GPU training.
\n\nFor SSD I have a separate SSD for my system and important files, and RAID 0 for the data. My option is to go for cheap SSDs in RAID 0, and add as many as I need. I have 2 now, but I'm thinking of adding another 2.
", "votes": null, "replies": [] }, { "id": 221907, "author_name": "thomasseleck", "author_url": "", "post_date": "09/16/2017 18:48:30", "content": "Thanks for your answers.
\n\nCurrently, I have an old Core i7 2600k with 16GB of RAM and a GTX1070. I'm planning to buy a new PC near Xmas 2017 or Q1 2018.
\n\nMaybe I'll wait for Volta GPUs, but Nvidia said they won't be launched until 2018. So, maybe I'll think about taking two GTX1080Ti or one Volta card (If we consider that Volta cards will be more expensive and that we can get rebates on GTX1080Ti).
\n\nI've tried XGBoost on GPU few months ago, but the performances are not as good as the CPU version. So, I stick to CPU for Gradient boosting algorithms at the moment. So, CPU is also important.
\n\nMoreover, I've noticed that when I do Deep Learning on my GTX1070, my CPU runs at 100% and the GPU is only used at 70% max, because the CPU is needed to do synchronization tasks between the GPU SMs.
\n\nFinally, for the choice of CPU, one question remains about using scikit-learn / numpy: as numpy relies on Intel's MKL library, will the performances be lower on AMD CPUs due to the fact that MKL is not supported?
", "votes": null, "replies": [] }, { "id": 222031, "author_name": "cpmpml", "author_url": "", "post_date": "09/17/2017 09:38:30", "content": "Make sure you get a m2 SSD as I/O is often what limits GPU use.
", "votes": null, "replies": [] }, { "id": 222053, "author_name": "sheriytm", "author_url": "", "post_date": "09/17/2017 10:25:53", "content": "Here is a great post you may find useful http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/
", "votes": null, "replies": [] }, { "id": 255131, "author_name": "hendraherviawan", "author_url": "", "post_date": "12/08/2017 11:56:21", "content": "HI @thomasseleck
\n\nWhich CPU that you chose, and why?
", "votes": null, "replies": [ { "id": 255149, "author_name": "thomasseleck", "author_url": "", "post_date": "12/08/2017 13:04:30", "content": "Hi @M Hendra Herviawan,
\n\nI've bought the AMD Threadripper 1950X for the following reasons:
\n\nDrawbacks:\nFor the moment, I have some BSOD on Windows 10 only at boot (when arriving on the desktop), when virtualization is enabled in BIOS (SVM/IOMMU setting).
", "votes": null, "replies": [] }, { "id": 353518, "author_name": "hendraherviawan", "author_url": "", "post_date": "07/06/2018 23:26:20", "content": "Look like TH have memory compatibility issue, which memory type & speed do you use ?
", "votes": null, "replies": [] }, { "id": 353657, "author_name": "thomasseleck", "author_url": "", "post_date": "07/07/2018 10:57:29", "content": "I use the G.Skill F4-2400C15Q-64GFX memory (64 GB DDR4 @ 2400 MHz). Note that I solved the BSOD problem. It was an issue related to an immature BIOS / UEFI that created BSODs at Windows startup. I upgraded the BIOS to the last version (601 for ASUS Prime X399-A) and I don't have BSODs anymore.
\n\nI also uninstall ASUS AURA (software that controls MB and GPU useless RGB LEDs) and ASUS AI Suite 3, both known to cause BSODs.
", "votes": null, "replies": [] } ] }, { "id": 353304, "author_name": "bguberfain", "author_url": "", "post_date": "07/06/2018 11:53:51", "content": "EDIT: never mind... I read the topic top down and didn't notice it was sorted in reverse time order.... I just found the answer for my question.
\n\nHi @Thomas!
\n\nThis is a little bit old thread, but the topic is of my interest....
\n\nI'm facing the same problem that you said: choosing a good CPU mostly for XGBoost and LightGBM models.
\n\nFurthermore, I have the same hardware configuration you had: an i7 2600k with 16 Gb and a more recent GPU.
\n\nDo you mind to share your decision on which CPU to buy? Did you went to Intel or AMD? How fast it is, compared to the i7 2600k?
\n\nThanks for your time!
", "votes": null, "replies": [ { "id": 353403, "author_name": "thomasseleck", "author_url": "", "post_date": "07/06/2018 17:06:12", "content": "I bought the Threadripper 1950X. It's nearly 5 times the power of the i7 2600k. It's great, especially for LightGBM, where GPU doesn't bring a significant speedup.
\n\nAnd if you can wait a little bit, AMD will release a 32 cores version of Threadripper that will be much better: https://www.techpowerup.com/245570/amd-ryzen-threadripper-2990x-makes-an-appearance-on-3dmark
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "221887": "Hi everyone,\n\nCan you tell me which hardware (CPU, RAM, SSD, GPU) are you using for this competition?\n\nI'm currently thinking of buying a new computer for Kaggle competitions, but one question remains: Which CPU is best between the Threadripper 1950X and the Core i9-7900X (both at $999) for running numpy / scikit-learn / LightGBM / XGBoost computations in a heavily multi-threaded setup?", "221904": "Threadripper is faster for parallel workloads (which for me has been the case in Kaggle competitions so far) and has 64 instead of 48 PCIe-lanes.\nI would pick Threadripper :)\n\n<-- Ryzen 1700 + GTX1080TI", "221906": "I am far from being an expert, I am waitinag for more informed opinions too.\n\nHere are my 2c:\n\nFirst of all if you want to do deep learning you should invest in a GPU, and don't worry much about the CPU. If you can wait, then wait for the Volta architecture based cards which was designed with DL in mind, supposedly they are expected very soon. If you are in a hurry, then the current sweet spot for price/performance is between GTX 1070 and GTX 1080, depending on what deal you can find. Otherwise, if you are looking for the best, then look for 1080Ti < Titan X < Titan Xp. Arguably the price premium is not justified. \n\nBetween 1950X and 7900X it is a difficult choice. 1950X is much faster in multi core applications, but slower per core. The speed boost diminishes with each core, and 20 and 32 threads are a lot. Also, there are other operations that are single threaded, so when all is said and done it is difficult to predict which one will end up on top.\n\nAlso, LightGBM and XGBoost both support GPU training.\n\nFor SSD I have a separate SSD for my system and important files, and RAID 0 for the data. My option is to go for cheap SSDs in RAID 0, and add as many as I need. I have 2 now, but I'm thinking of adding another 2.", "221907": "Thanks for your answers.\n\nCurrently, I have an old Core i7 2600k with 16GB of RAM and a GTX1070. I'm planning to buy a new PC near Xmas 2017 or Q1 2018.\n\nMaybe I'll wait for Volta GPUs, but Nvidia said they won't be launched until 2018. So, maybe I'll think about taking two GTX1080Ti or one Volta card (If we consider that Volta cards will be more expensive and that we can get rebates on GTX1080Ti).\n\nI've tried XGBoost on GPU few months ago, but the performances are not as good as the CPU version. So, I stick to CPU for Gradient boosting algorithms at the moment. So, CPU is also important.\n\nMoreover, I've noticed that when I do Deep Learning on my GTX1070, my CPU runs at 100% and the GPU is only used at 70% max, because the CPU is needed to do synchronization tasks between the GPU SMs.\n\nFinally, for the choice of CPU, one question remains about using scikit-learn / numpy: as numpy relies on Intel's MKL library, will the performances be lower on AMD CPUs due to the fact that MKL is not supported?", "222031": "Make sure you get a m2 SSD as I/O is often what limits GPU use.", "222053": "Here is a great post you may find useful http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/", "255131": "HI @thomasseleck \n\nWhich CPU that you chose, and why?", "255149": "Hi @M Hendra Herviawan,\n\nI've bought the AMD Threadripper 1950X for the following reasons:\n\n - Same performances as Core i9 7960X but cheaper\n - 16 cores instead of 10 for the same price ( against the Core i9 7900X)\n - No heating and VRM issues (with Enermax Liqtech TR4 cooler, my CPU does not go beyond 55 Celsius degrees)\n - More PCIe lanes for NVMe drives or GPUs\n\nDrawbacks:\nFor the moment, I have some BSOD on Windows 10 only at boot (when arriving on the desktop), when virtualization is enabled in BIOS (SVM/IOMMU setting).", "353304": "*EDIT: never mind... I read the topic top down and didn't notice it was sorted in reverse time order.... I just found the answer for my question.*\n\n\nHi @Thomas!\n\nThis is a little bit old thread, but the topic is of my interest....\n\nI'm facing the same problem that you said: choosing a good CPU mostly for XGBoost and LightGBM models.\n\nFurthermore, I have the same hardware configuration you had: an i7 2600k with 16 Gb and a more recent GPU.\n\nDo you mind to share your decision on which CPU to buy? Did you went to Intel or AMD? How fast it is, compared to the i7 2600k?\n\nThanks for your time!", "353403": "I bought the Threadripper 1950X. It's nearly 5 times the power of the i7 2600k. It's great, especially for LightGBM, where GPU doesn't bring a significant speedup.\n\nAnd if you can wait a little bit, AMD will release a 32 cores version of Threadripper that will be much better: https://www.techpowerup.com/245570/amd-ryzen-threadripper-2990x-makes-an-appearance-on-3dmark", "353518": "Look like TH have memory compatibility issue, which memory type & speed do you use ?", "353657": "I use the G.Skill F4-2400C15Q-64GFX memory (64 GB DDR4 @ 2400 MHz). Note that I solved the BSOD problem. It was an issue related to an immature BIOS / UEFI that created BSODs at Windows startup. I upgraded the BIOS to the last version (601 for ASUS Prime X399-A) and I don't have BSODs anymore.\n\nI also uninstall ASUS AURA (software that controls MB and GPU useless RGB LEDs) and ASUS AI Suite 3, both known to cause BSODs." }, "source": "meta" }