Implemented in Q4: our GPU cloud’s updates

Compute CloudCompute Cloud

• Added AMD platform configuration (cpu-d3).

• Added new monitoring dashboards to the console: CPU, GPU, RAM, NVLink metrics, InfiniBand metrics, Ethernet metrics.

• Selecting the type of public IP address (static or dynamic) is now available when creating a VM in GUI, CLI or Terraform.

Cluster managementCluster management

Soperator

• Enabled sshd on worker nodes for direct, secure access and streamlined troubleshooting.

• Introduced support for enroot without requiring root privileges, enabling more flexible and secure containerized workflows.

• Added dockerd support for a container runtime management.

• Integrated apparmor for enhanced security profiles and workload isolation.

• Introduced Slurm partitions to provide logical separation of resources and improve scheduling efficiency.

• Launched a Slurm REST API, allowing programmatic cluster management, job submission and querying cluster states.

• Supported CPU-only and GPU-only cluster types, enabling users to tailor their infrastructure precisely to their workload requirements.

Managed Service for Kubernetes

• Added load balancer support to expose services to the internet and internal networks.

• Added new monitoring dashboards to the console: CPU, GPU, RAM, NVLink metrics, InfiniBand metrics, Ethernet metrics.

• Launched a node autoscaler to dynamically add or remove nodes based on resource demands.

• Introduced integration with our Container Registry service for seamless image management.

• Enabled high availability for clusters by default, ensuring a redundant control plane at no additional cost.

• Added support for ReadWriteOnce block volumes with CSI over block storage in preview mode. Please contact support or your cloud solutions architect to start using them.

• Added support for setting up clusters in custom subnets, enabling Kubernetes clusters to connect to client’s private address space via VPN.

Container Registry

• Introduced Container Registry in preview stage, available free of charge.

Data storeData store

Shared Filesystem

• Launched filesystem resize feature. Performance of the filesystem will increase for every 4 TB of size.

• Added support for retrieving available platforms and presets via API and CLI.

• Enabled resizing of filesystems and disks without requiring recreation.

Object Storage

• Added performance and consumption metrics to the console.

Managed Service for PostgreSQL

• Added performance metrics to the console.

• Added support for Run:ai.

• Launched private cluster endpoints (available only from the user’s VPC).

• Added an ability to enable and disable public cluster’s endpoints.

• Added pool size settings and some minor settings related to PostgreSQL cluster parameters.

• Added an ability to update the created cluster by changing cluster parameters (number of hosts, number of CPU, memory size), PostgreSQL parameters (password, autovacuum settings, etc.) or pooler settings.

MLOps services and appsMLOps services and apps

Managed Service for MLflow

• Private and public endpoints are now supported.

• Added MLflow logs to the web console.

• Added MLflow performance metrics to the web console.

Cloud platform featuresCloud platform features

Network

• Added “number of available public IP addresses” as a public quota in the console.

• Ability to change the range of IP addresses for a private network is now available.

• Wireguard VPN solution has been added for secure remote access.

Identity and Access Management

• Added GitHub authentication to the console.

• Customers now have the ability to configure their custom federations via SAML2 protocol via API.

API

• Released the API repository.

• Go SDK and Python SDK are released in preview mode. Please contact support or your cloud solutions architect to start using them.

Status page

• Now you can subscribe on incidents on the status page.

Nebius AI StudioNebius AI Studio

• Our portfolio of LLMs now includes more than 30 models.

• The platform now supports increased rate limits of 100M+ tokens per minute and beyond.

• Added the newest Llama-3.3-70B-Instruct model.

• Added Guard models.

• Added specialized Med42 and Llama3-OpenBioLLM-8B models.

• Vision models are now also available.

• LoRA is available in preview mode.

To learn about the latest updates in Nebius AI Studio, check out a separate post.