{
  "id": 64667,
  "title": "Inference Speed Prize: NCHW vs NHWC ?",
  "url": "/competitions/airbus-ship-detection/discussion/64667",
  "author_name": "Michael Schimmer",
  "post_date": "2018-08-31T10:21:54.974000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Does anyone know how the inference speed gets testet? </p>\n\n<p>Especially, if i can optimize everything to run on GPU (and i would therefore use the NCHW - Format) or if we are bound to use CPU-Kernels?</p>\n\n<blockquote>\n  <p>NCHW should always be used when training with GPUs. NHWC is sometimes faster on CPU</p>\n</blockquote>\n\n<p>(see: <a href=\"https://www.tensorflow.org/performance/performance_guide#data_formats\">Tensorflow Performance Guide</a>)</p>",
  "messages": [
    {
      "id": 379410,
      "postDate": "2018-08-31T10:21:54.973Z",
      "content": "<p>Does anyone know how the inference speed gets testet? </p>\n\n<p>Especially, if i can optimize everything to run on GPU (and i would therefore use the NCHW - Format) or if we are bound to use CPU-Kernels?</p>\n\n<blockquote>\n  <p>NCHW should always be used when training with GPUs. NHWC is sometimes faster on CPU</p>\n</blockquote>\n\n<p>(see: <a href=\"https://www.tensorflow.org/performance/performance_guide#data_formats\">Tensorflow Performance Guide</a>)</p>",
      "rawMarkdown": "Does anyone know how the inference speed gets testet? \n\nEspecially, if i can optimize everything to run on GPU (and i would therefore use the NCHW - Format) or if we are bound to use CPU-Kernels?\n\n&gt;  NCHW should always be used when training with GPUs. NHWC is sometimes faster on CPU\n\n(see: [Tensorflow Performance Guide][1])\n\n\n  [1]: https://www.tensorflow.org/performance/performance_guide#data_formats",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "379410": "Does anyone know how the inference speed gets testet? \n\nEspecially, if i can optimize everything to run on GPU (and i would therefore use the NCHW - Format) or if we are bound to use CPU-Kernels?\n\n&gt;  NCHW should always be used when training with GPUs. NHWC is sometimes faster on CPU\n\n(see: [Tensorflow Performance Guide][1])\n\n\n  [1]: https://www.tensorflow.org/performance/performance_guide#data_formats"
  }
}