Updating benchmarks for recent cuDNN v6 #126
Description
cuDNN v6 is now released and there are clear improvements. How can I go about updating these numbers (and adding more frameworks)? Is it best to just submit a PR with README tables updated? Or did you have different plans for new cuDNN versions?
So far I've got the following for TensorFlow:
2017-04-12 11:53:22.136747: I tensorflow/core/common_runtime/gpu/gpu_device.cc:908] DMA: 0 1
2017-04-12 11:53:22.136756: I tensorflow/core/common_runtime/gpu/gpu_device.cc:918] 0: Y Y
2017-04-12 11:53:22.136759: I tensorflow/core/common_runtime/gpu/gpu_device.cc:918] 1: Y Y
2017-04-12 11:53:22.136765: I tensorflow/core/common_runtime/gpu/gpu_device.cc:977] Creating TensorFlow device (/gpu:0) -> (device: 0, name: TITAN X (Pascal), pci bus id: 0000:01:00.0)
2017-04-12 11:53:22.136768: I tensorflow/core/common_runtime/gpu/gpu_device.cc:977] Creating TensorFlow device (/gpu:1) -> (device: 1, name: TITAN X (Pascal), pci bus id: 0000:03:00.0)
2017-04-12 11:53:23.636099: step 10, duration = 0.017
2017-04-12 11:53:23.808993: step 20, duration = 0.017
2017-04-12 11:53:23.981334: step 30, duration = 0.017
2017-04-12 11:53:24.155021: step 40, duration = 0.017
2017-04-12 11:53:24.329057: step 50, duration = 0.017
2017-04-12 11:53:24.503106: step 60, duration = 0.017
2017-04-12 11:53:24.677380: step 70, duration = 0.017
2017-04-12 11:53:24.851187: step 80, duration = 0.017
2017-04-12 11:53:25.023797: step 90, duration = 0.017
2017-04-12 11:53:25.179516: Forward across 100 steps, 0.017 +/- 0.002 sec / batch
2017-04-12 11:53:26.523299: step 10, duration = 0.050
2017-04-12 11:53:27.033520: step 20, duration = 0.051
2017-04-12 11:53:27.538539: step 30, duration = 0.050
2017-04-12 11:53:28.039816: step 40, duration = 0.050
2017-04-12 11:53:28.547799: step 50, duration = 0.050
2017-04-12 11:53:29.049781: step 60, duration = 0.050
2017-04-12 11:53:29.551407: step 70, duration = 0.050
2017-04-12 11:53:30.055291: step 80, duration = 0.050
2017-04-12 11:53:30.562754: step 90, duration = 0.050
2017-04-12 11:53:31.013226: Forward-backward across 100 steps, 0.050 +/- 0.005 sec / batch