Spell, an artificial intelligence development platform founded by a former Facebook AI engineering chief, has rolled out its services in earnest after snagging $15m in funding.
Spell was founded in 2016 as Top 1 Networks by Serkan Piantino, former director of engineering for Facebook AI research, as a way for AI developers to get their hands onto up to the minute hardware.
The company appears to have spend much of last year setting up an office and hiring in New York, as well as demo-ing its nascent service.
Yesterday it went public on $15m of funding led by Eclipse Ventures and Two Signma Ventures, and fleshed out its offering with Spell for Teams.
In a blog post, it said practitioners had told it “The tools available for building real products with AI were immature, fragmented and costing them time.” Hence, it continued, “the next version of Spell [is] a complete end-to-end system for exploring, training, building, automating and serving models built with deep learning.
New features include dedicated clusters and hyperparameter searches. Also included is the ability to turn a trained model into a REST API, which can be deployed into production. The rollout also includes Organisations, which allows teams to see each others’ work in one place.
Spell’s rationale rollout chimes with other efforts. The Databricks-backed open source MLFLow framework aims to address the workflow issues, and make it easier for datascientists to compare and reproduce results.
If you haven’t got an industry strength hardware setup, the company can rent you a range of configurations, all at cost, ranging across CPU and GPU configurations
Individuals can get free access to a basic tier, described as “everything you need to get going in ML/AI” consisting of a default 2 vCPUs and 4GB of RAM.
Prices kick in at £0.68 per hour for the cpu-big package of 16 vCPUs and 32GB of RAM, ranging up to a $6.05 per hour ram-huge configuration of 96 vCPUs and 768 GB of RAM.
GPU options range from a 4.3TFlops K80 with 4 vCPUs and 61GB of RAM at $0.90 an hour to a 8xTesla V100, 125.6 TFlops set up with 64 vCPUs and 488GB of RAM for $24.48 per hour.
Give that this is all “no markup”, we presume the company will be living off the Paid Features options, starting at $99 per host per month.