github.com/sentienttechnologies/studio-go-runner@v0.0.0-20201118202441-6d21f2ced8ee/docs/gpus.md (about)

     1  GPU Allocation
     2  
     3  When using StudioML gpus specified in the resource section of the config.yaml being used are typically done using slots.
     4  
     5  ```
     6  resources_needed:
     7      cpus: 1
     8      gpus: 1
     9      hdd: 3gb
    10      ram: 2gb
    11      gpuMem: 2gb
    12  ```
    13  
    14  When using cards such as the GTX 1050, or GTX 1060 then the slots assigned will be in single digit increments.  When using cloud and data center deployments where higher powered cards are being used the gpu value when expressed as a 1 will result in usage of the higher powered GTX 1070 and GTX 1080 without any problems.
    15  
    16  However as the power of the cards deployed within your infrastructure increases it becomes more important to express the gpus values as slots representing the desired upper bound.  For example
    17  
    18  |Slots|Card Types|
    19  |---|---|
    20  |2|GTX 1050, GTX 1060|
    21  |2|GTX 1070, GTX 1080|
    22  |4|Titan X, Tesla P40|
    23  |8|Tesla P100|
    24  
    25  If the number of slots you define is above what is available then the system will attempt to create your desired configuration from smaller units of GPUs.  However it will not drop below units of 4 slots when larger quantities are specified.  For example it is possible when using 8 slots that 2 Tesla P40s might be used instead.  In the future the resources_needed block will be used to allow you to specify the smallest slots that are permitted.
    26  
    27  Copyright &copy 2019-2020 Cognizant Digital Business, Evolutionary AI. All rights reserved. Issued under the Apache 2.0 license.