0.25
0.5
0.75
1.25
1.5
1.75
2
Deep Learning (hopefully faster)
Published on Sep 13, 20157055 Views
Related categories
Chapter list
Deep Learning (hopefully faster)00:00
Scope - 100:05
Scope - 201:38
Overview - 103:08
Cycle time arguement04:01
Scaling arguement06:10
Approach10:32
Workload - 113:51
Workload - 214:51
Caveat16:09
Single Note Principles17:50
Setting goals18:12
The speed of light - 119:21
The speed of light - 220:42
The speed of light - 322:39
Performence modeling23:16
Model of a compute note - 125:02
Model of a compute note - 225:42
Model of a compute note - 327:24
Example: Matrix-vector multipy - 128:39
Example: Matrix-vector multipy - 230:20
Arithmetic intensity32:29
The ''Roofline'' model - 133:40
The ''Roofline'' model - 235:15
The ''Roofline'' model - 336:32
Example: matrix-matrix multiply - 137:08
Example: matrix-matrix multiply - 238:43
Example: matrix-matrix multiply - 339:35
The ''Roofline'' model - 440:13
Roofine in practice - 141:45
Roofine in practice - 242:54
Summary46:42
Single Note Issues47:24
Minibatch size - 147:28
Minibatch size - 248:14
Minibatch size - 349:26
Moral of story51:07
Optimizing software51:42
Things to try52:02
Note on code cocmplexity52:46
Multinode53:43
Training with clusters53:57
Wgat can we hope to achive? - 154:50
Wgat can we hope to achive? - 255:26
Wgat can we hope to achive? - 356:07
Example: weak scaling01:00:18
Weak vs. strong01:01:49
Performance modeling - 101:02:40
Example: Data Parallelism - 101:03:09
Example: Data Parallelism - 201:04:01
Local throughput01:05:40
Performance modeling - 201:09:06
Performance modeling - 301:09:26
''Roofline'' model - 101:10:18
''Roofline'' model - 201:10:45
Example: Data Parallelism - 101:13:44
Example: Data Parallelism - 201:14:00
Overall throughput01:14:51
Assumptions01:15:41
Putting everything together - 101:19:40
Putting everything together - 201:20:21
Optimization strategy01:22:24
Conclusion01:24:04
Key ideas01:24:06
Thank you!01:25:03