Anyscale has announced a GPU-native architecture for Ray, a new RayTurbo engine and the general availability of Ray Data at its Ray Summit conference in San Francisco, the company said on 1 October 2024.
RayTurbo is an optimised version of Ray available only on the Anyscale platform. Compared with open-source Ray, the company said it offers up to 4.5 times faster data processing, up to 90% lower costs through spot instance support and 60% higher serving throughput.
Ray now uses static task graphs to cut overhead on GPUs for workloads such as distributed training and model serving. Ray Data adds support for lakehouse formats including Hudi, Iceberg and Delta Lake.
The new Anyscale Operator for Kubernetes launched with Amazon EKS, Google GKE, Azure AKS and OCI Kubernetes Engine. Anyscale also added quota management, workload prioritisation and observability tools for enterprise teams.
“We’ve seen Anyscale usage soar 44x over the past year, and Ray is now orchestrating more than 1 million clusters per month,” said Robert Nishihara, co-founder of Anyscale.