{
  "id": 243993,
  "title": "Sometimes a good harware can be a game changer",
  "url": "/competitions/bms-molecular-translation/discussion/243993",
  "author_name": "",
  "post_date": "2021-06-04T19:01:24.931773200Z",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In most of the top solutions, I see that they have used resnet or effNet with good tricks. These models cross the 9hrs limits of Kaggle after a few epochs. So we need either our own hardware or cloud to train large models.</p>\n<p>The aim of this thread is to collect the list of hardware or Cloud used by top teams. <br>\nTSo the kagglers will have some idea the next time they are about to purchase any machine or subscribe to a cloud for Deep Learning.</p>\n<ol>\n<li>I found one by <a href=\"https://www.kaggle.com/haqishen\" target=\"_blank\">@haqishen</a>. He has 5th place and used Z8G4 Workstation with dual RTX6000 GPU and a ZBook with RTX5000 GPU.</li>\n</ol>",
  "messages": [
    {
      "id": "1336284",
      "postDate": "06/04/2021 19:01:24",
      "content": "<p>In most of the top solutions, I see that they have used resnet or effNet with good tricks. These models cross the 9hrs limits of Kaggle after a few epochs. So we need either our own hardware or cloud to train large models.</p>\n<p>The aim of this thread is to collect the list of hardware or Cloud used by top teams. <br>\nTSo the kagglers will have some idea the next time they are about to purchase any machine or subscribe to a cloud for Deep Learning.</p>\n<ol>\n<li>I found one by <a href=\"https://www.kaggle.com/haqishen\" target=\"_blank\">@haqishen</a>. He has 5th place and used Z8G4 Workstation with dual RTX6000 GPU and a ZBook with RTX5000 GPU.</li>\n</ol>",
      "rawMarkdown": "In most of the top solutions, I see that they have used resnet or effNet with good tricks. These models cross the 9hrs limits of Kaggle after a few epochs. So we need either our own hardware or cloud to train large models.\n\nThe aim of this thread is to collect the list of hardware or Cloud used by top teams. \nTSo the kagglers will have some idea the next time they are about to purchase any machine or subscribe to a cloud for Deep Learning.\n\n1. I found one by @haqishen. He has 5th place and used Z8G4 Workstation with dual RTX6000 GPU and a ZBook with RTX5000 GPU.",
      "votes": null
    },
    {
      "id": "1336779",
      "postDate": "06/05/2021 07:48:32",
      "content": "<p>I don't think it is \"sometimes\". It should be \"99.99% of the time\". Having a good hardware means you train faster, have more ideas to try, and can fit big models and batch size, etc.</p>\n<p>Of course there is still (gold) winning solution with Kaggle's spec, but that lies in the 0.01%.</p>",
      "rawMarkdown": "I don't think it is \"sometimes\". It should be \"99.99% of the time\". Having a good hardware means you train faster, have more ideas to try, and can fit big models and batch size, etc.\n\nOf course there is still (gold) winning solution with Kaggle's spec, but that lies in the 0.01%.",
      "votes": null
    },
    {
      "id": "1336830",
      "postDate": "06/05/2021 08:31:46",
      "content": "<p>May be, than I belong to that 0.01% as I have a gold with only using Kaggle TPU. :)</p>",
      "rawMarkdown": "May be, than I belong to that 0.01% as I have a gold with only using Kaggle TPU. :)",
      "votes": null
    },
    {
      "id": "1338483",
      "postDate": "06/06/2021 13:43:34",
      "content": "<p>In terms of TPU we are mostly fair :) I don't think people will buy a TPU given that they can use cloud TPUs.</p>",
      "rawMarkdown": "In terms of TPU we are mostly fair :) I don't think people will buy a TPU given that they can use cloud TPUs.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1336779,
      "author_name": "aeryss",
      "author_url": "",
      "post_date": "06/05/2021 07:48:32",
      "content": "<p>I don't think it is \"sometimes\". It should be \"99.99% of the time\". Having a good hardware means you train faster, have more ideas to try, and can fit big models and batch size, etc.</p>\n<p>Of course there is still (gold) winning solution with Kaggle's spec, but that lies in the 0.01%.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1336830,
          "author_name": "vikrant06",
          "author_url": "",
          "post_date": "06/05/2021 08:31:46",
          "content": "<p>May be, than I belong to that 0.01% as I have a gold with only using Kaggle TPU. :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1338483,
          "author_name": "aeryss",
          "author_url": "",
          "post_date": "06/06/2021 13:43:34",
          "content": "<p>In terms of TPU we are mostly fair :) I don't think people will buy a TPU given that they can use cloud TPUs.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1336284": "In most of the top solutions, I see that they have used resnet or effNet with good tricks. These models cross the 9hrs limits of Kaggle after a few epochs. So we need either our own hardware or cloud to train large models.\n\nThe aim of this thread is to collect the list of hardware or Cloud used by top teams. \nTSo the kagglers will have some idea the next time they are about to purchase any machine or subscribe to a cloud for Deep Learning.\n\n1. I found one by @haqishen. He has 5th place and used Z8G4 Workstation with dual RTX6000 GPU and a ZBook with RTX5000 GPU.",
    "1336779": "I don't think it is \"sometimes\". It should be \"99.99% of the time\". Having a good hardware means you train faster, have more ideas to try, and can fit big models and batch size, etc.\n\nOf course there is still (gold) winning solution with Kaggle's spec, but that lies in the 0.01%.",
    "1336830": "May be, than I belong to that 0.01% as I have a gold with only using Kaggle TPU. :)",
    "1338483": "In terms of TPU we are mostly fair :) I don't think people will buy a TPU given that they can use cloud TPUs."
  },
  "source": "meta"
}