AI SaaS companies, AI engineers
SONM can provide:
- Unlimited resources for both neural network training and appliance. Hundreds instances with 4-8 GPUs GPUs can be involved instantly
- Low, competing prices. For example, a PC with Nvidia Geforce 1070 GPU card will cost about 0,12USD/hour, comparing to minimal 0,4USD/hour on the market
- Strict security for customer's data. SONM partners that provide computing resources are assessed for security and sign NDA with customers. All data transfers are encrypted using TLS and dedicated public key infrastructure (MVP protects only docker image transfer).
- Compatibility with professional machine learning software: TensorFlow, Theano, Caffe, MLib (Spark) and others
- Integration with cloud services for data storage (Amazon S3, Dropbox, etc.). There is also built-in p2p efficient data transfer component, optimized for simultaneous multipoint upload (not implemented in MVP).
- Ability to run custom micro services architecture. Companies that implement AI SaaS projects can use SONM to deploy their own application packed as docker containers (not implemented in MVP). SONM will provide an API to start and stop containers with GPU access.
- Neural network training tasks need big amount of data to be downloaded to remote PCs. The size of training data vary from several gigabytes to hundreds of terabytes. SONM's resources are distributed across the world, their Internet connections can be poor, so it could make quite long time to download data and perform machine learning. Still Internet connection speeds of some dedicated SONM partners are comparable to local network, reaching 10 Gbit/s. Besides, training data are often collected and hosted on external Internet resources (cloud storages, online services), so the data transfer will not be limited to customer's Internet connection speed. Besides, SONM implements a BitTorrent-protocol based filesystem that can effectively distribute the files, prioritizing the transfers to start the jobs immediately (not implemented in MVP). SONM nodes will exchange data with each other, increasing overall load speed. The more SONM nodes are involved in machine learning task, the more total transfer speed can be achieved, as SONM will automatically balance data exchanges to minimize latencies.