PROFET Anchor Prediction
PROFET Anchor Prediction System can forecast latencies on various target GPU instances based on profiling data measured on the anchor GPU instance. Therefore, JSON type profiling data is required to use this system, and example data can be used in the link below for testing. You can check documentation for more information.
Model: VGG16
Pixel size: 128x128x3
Batch size: 16 batchsize
Anchor instance: g3s.xlarge
Anchor latency: 122400 us
Model: VGG16
Pixel size: 128x128x3
Batch size: 256 batchsize
Anchor instance: g3s.xlarge
Anchor latency: 1529870 us
Profiling Feature
| Instance Type |
Latency (us) |
Cost (1000 batches) |
|---|---|---|
| g3s.xlarge 0.75 $/hr |
||
| g4dn.xlarge 0.526 $/hr |
||
| p2.xlarge 0.9 $/hr |
||
| p3.2xlarge 3.06 $/hr |
Anchor Prediction Result
Latency and Cost of Target Instances
PROFET Scaler Prediction
PROFET Scaler Prediction System can forecast latencies for any batchsize or data size between maximum and minimum, and it requires only latencies at minimum and maximum size. First, you can specify the target GPU instance and type of size (batch or data) that you want to predict. Next, enter the minimum and maximum latencies as actual or anchor predicted values. You can check documentation for more information.
Target Size
Scaler Prediction Result
Size-Latency Chart
Size:
Latency: 0