____  _             
| __ )| | ___   __ _ 
|  _ \| |/ _ \ / _` |
| |_) | | (_) | (_| |
|____/|_|\___/ \__, |
               |___/ 
     _               _            
 _/_| |__   ___ _ __| |_   _  _/_ 
/ _ \ '_ \ / _ \ '__| | | | |/ _ \
  __/ |_) |  __/ |  | | |_| |  __/
\___|_.__/ \___|_|  |_|\__,_|\___|

< Previous post | Next post >

[2025-12-18][Daily]

More experiments comparing pytorch intraop vs dahl

For these experiments we only ran pytorch in intraop mode, because it makes way more sense in terms of comparability as seen in previous day experiment.

Here we clearly see the confidence intervals on cifar and mnist which is great! On big-fashion we don't see it cause runtimes are ranging from 1200 - 100s compared to 80 - 10s on fashion-mnist for example. And here on big-fashion results are always better than pytorch in terms of runtime/consumption, which is encouraging.

Scheduling overhead is getting compensated when using bigger workloads, so at some point using a TGCS becomes more intersting than using raw pytorch intraop.

More info on the hyperparameters:

• LR: 0.001
• Epochs: 5
• Batch size: 32 when fixed (varying num cpu)
• Num cpu: 28 when fixed (varying batch size)
• Num samples: 8192
• Num repeated runs: 10

< Previous post | Next post >