Serverless Deep Learning SNARK HYPER
Scale any machine learning pipeline from a single server to hundreds of cloud instances with zero friction.
With the exponential increase of training data and the computational complexity of machine learning models, Deep Learning on the cloud has become very engineering heavy. Single training experiment of a production ready model may take up to 2 weeks. If one wants to explore more variations or fine-tune hyper-parameters, production lifecycle becomes really slow. Picking the right instance, managing cloud instances for optimal utility rate, handling spot/pre-emptive instances, running multiple experiments at the same time, all require a lot of DevOps work from deep learning engineers, who may better spend their time developing models.
Scale Any Machine Learning Pipeline to Elastic Cloud Servers
Any Framework
Compatible with any machine learning framework. Run large-scale computation on PetaBytes of data with any R / Python / Java / C++ code.
Easy
Run any local machine learning pipeline on a fleet of cloud instances without any change in your code. Zero hassles in data migration and environment setup for new cloud servers.
Monitor
Cloud resource (CPU/GPU/RAM) utilization tracking and machine learning experiment analysis. Transparent overview of your cloud spending for each of your task.
Scalable
Unlimited persistent storage for all of your tasks. Never worried about hard drive filling up during the execution of your jobs. Scaling your task to thousands of CPUs and hundreds of GPUs with zero friction.
Separation of Persistent Data Storage and Elastic Compute Fleet.
- Your code base and data stay in persistent storage which is cost efficient (~20$/TB per month) and infinitely scalable.
- The persistent storage can be mount as local folders to each new cloud instance saving the trouble of data migration and environment setup.
Execute Parameterized Commands Across a Fleet of Cloud Instances
- Hyper-parameter Search. Run machine learning training with different hyper-parameters on different cloud instances.
- Batch Prediction. Run any trained model on Petabytes of data across a fleet of cloud instances in parallel. Instant model evaluation on any big dataset.
- Feature transformation. Compute features from Petabytes of raw data across thousands of cloud instances in parallel. Never limited to Spark libraries anymore. Use any R / Python / other package available.
Machine Learning Experiment Tracking and Insight Analysis
- Track and evaluate thousands of models and quickly surface the best-performing ones across your organization.
- Automatic model reporting tools.
- Monitor the CPU/GPU/RAM utilization of each cloud resource.
Enabling elastic ML compute to run big data ETL, feature transformations, machine learning and deep learning pipelines with any R/Python/Matlab/C++ code.
Up to 100x speed up for your machine learning pipeline
- Scaling any machine learning pipeline from a single server to an elastic group of 100 cloud instances to achieve 100x speed up.
- Scale your favorite R/Python packages to thousands of CPUs across hundreds of machines. Never limited to Spark libraries anymore for large scale computation.
- Typical use cases include hyper-parameter search, batch prediction, and feature transformation.
Save 2 hours/day for each machine learning engineer
- Saving ML engineers’ time for configuring cloud infrastructure, monitoring cloud resource utilization and ML environment setup in each new cloud instance.
- Let ML engineers easily create model reports from training logs.
Saving up to 80% on your cloud spending
- Choosing the most cost-efficient hardware from cross-cloud including AWS/Azure/GCP.
- Snark support Pre-emptible / Spot instances which are 70% cheaper than on-demand instances. Snark reschedules the jobs automatically for any spot interruption/instance pre-emption.
Hiç yorum yok:
Yorum Gönder
Not: Yalnızca bu blogun üyesi yorum gönderebilir.