NucleusResearch.com
4
Document Number: T176 December 2019
on machine learning-associated hardware and staff by up to 80 percent after moving the
workloads to Amazon SageMaker. Based on interviews with AWS customers leveraging
SageMaker for machine learning over the past two years, developer productivity increased
by 20 to 25 percent by automating administrative tasks and accelerated model training and
tuning. Additionally, internal Amazon Web Services' research found that using SageMaker
lowers the TCO of machine learning projects by at least 54 percent. The research also found
that using Amazon Managed Spot Training reduced model training costs by up to 90
percent and using Amazon Elastic Inference reduced inference costs by up to 75 percent.
INCREASED MACHINE LEARNING AGILITY
Using Amazon SageMaker, developers were able to implement more machine learning
models faster, increasing the speed of their research by enabling teams to test new ideas,
tune parameters, and troubleshoot models, then rapidly deploy them on real data to assess
the results. Customers said using SageMaker makes training and deployment 2-3 times
faster than on self-managed infrastructure. As a service from AWS, SageMaker integrates
with other AWS services such as Amazon S3, AWS Lambda, and others. This reduces
complexity for developers and reduces development times, further increasing organizational
agility to react more effectively to issues or implement new ideas. In conversations with
customers, the integration with other AWS services was cited as a key factor in the decision
to choose Amazon in over 75 percent of conversations.
C U S T O M E R P R O F I L E S
Over the past two years, Nucleus has interviewed over 50 organizations encompassing over
400 unique machine learning projects on AWS. The included customer profiles are written
without identifiable information but should provide a clear understanding of the use cases
and the value provided by Amazon.
Using SageMaker makes training and
deployment 2-3 times faster than on self-
managed infrastructure
