Overview

What are Instances?

Instances are ready to use, containerized workspaces designed for AI/ML development and experimentation. Each instance combines a container environment, optional JupyterLab interface, and pre-installed ML libraries eliminating the need for manual environment setup. Your team can immediately run scripts, manage datasets, and conduct model experiments without any infrastructure configuration.

TIR Instances are high performance, managed compute resources built for modern AI and cloud workloads. Rather than traditional virtual machines that virtualize an entire hardware stack, TIR Instances use OS level virtualization running as high performance containers on specialized host nodes. This reduces overhead and improves efficiency.

What this means in practice

Capability	Benefit
Dedicated Linux environment	Full control without shared OS interference
Near bare metal performance	Minimal virtualization overhead
Direct hardware accelerator access	GPU resources available without abstraction layers
Near instant provisioning	Faster startup compared to traditional VMs

In short: you get the isolation of a virtual machine with the performance of bare metal hardware purpose built for AI workloads.

Core Infrastructure Specifications

Execution Model: Kernel shared, namespace isolated containers ensuring deterministic performance for HPC workloads.
Hardware Acceleration: Direct path NVIDIA GPU Passthrough via the NVIDIA Container Toolkit, supporting H200, H100, A100, and L40S architectures.
Persistence Layer: Stateful data retention via a persistent volume mount at /home/jovyan, decoupling the user's workspace from the ephemeral container lifecycle.
Networking & Access: Secure ingress via OpenSSH (Port 22) and encrypted WebSockets for JupyterLab access.

Use Case

Use Case	Key Benefits
Web Application Hosting	High performance & reliability: Utilizes high IOPS and SSD storage to ensure fast page loads, combined with scalable vCPUs that can be adjusted to match fluctuating user traffic.
LLM Training & Fine-Tuning	Massive throughput: Leverages NVLink for high speed communication between GPUs, massive VRAM (up to 141 GB) for large model weights, and NVMe storage to eliminate data bottlenecks during training.
Secure Network	Cloud Native Agility: Offers seamless VPC integration for secure networking, MetalLB for load balancing services, and native autoscaling to automatically adjust resources based on cluster demand.
Computer Vision & NLP	Ready to Use AI Stack: Features multi GPU support for parallel processing and comes with pre configured CUDA/cuDNN environments to skip the complex driver installation process.

Why use TIR Instances?

Rapid Deployment: Launch environments with pre installed software stacks like PyTorch, TensorFlow, and Hugging Face Transformers.
Hardware Acceleration: Native access to industry leading NVIDIA GPUs (H200, H100, A100, L40S).
Development Ready: Every instance includes integrated access via JupyterLab for interactive coding and SSH for terminal based automation.
Persistence: A dedicated directory at /home/jovyan ensures your code and notebooks survive instance restarts.

Supported Instance Types

TIR provides four instance types, each optimized for specific workload requirements.

CPU Instances

High-performance compute instances for general-purpose workloads, data preprocessing, and non-GPU applications.

Plan: C3 Series
Configurations: 4–48 vCPUs with 8–256 GB RAM

Available options: 4/8 GB, 8/16 GB, 16/32 GB, 20/64 GB, 24/96 GB, 32/128 GB, 40/192 GB, 48/256 GB.

Designed to provide scalable compute and memory capacity to meet varying workload requirements.

GPU Instances

Dedicated NVIDIA GPU instances built for AI training and inference workloads.

Available GPUs: H100 SXM, H200 SXM (141GB), A100 (80GB / 40GB), L40S (48GB), L4 (24GB).
Scaling: 1 to 8 GPUs per instance.

Spot Instances

Cost efficient instances that leverage unused cloud capacity at reduced rates.

Savings: Up to 70% lower than On-Demand (e.g., H200 Spot at ₹88.00/hr vs. ₹300.14/hr On-Demand).
Note: Instances may be interrupted when capacity is reclaimed for On Demand users. Best suited for flexible or fault tolerant workloads.

Private Cluster Instances

Dedicated infrastructure for enterprise grade security and isolated networking.

Private Clusters provide a fixed cost GPU resource pool with guaranteed hardware availability and predictable billing regardless of utilization. Teams can share compute across multiple projects and sub services with no additional per node charges.

Key Benefit: Deploy within a private network boundary, ideal for sensitive or regulated workloads.

Image Categories & Available Options

Pre-built

Ready-to-use environments: PyTorch, TensorFlow, Diffusers, Triton, and more. CUDA and dependencies pre-configured. For more info click here.

+ Recommended

Base OS

Minimal Ubuntu image. Install exactly what you need full control, more setup time.

Custom

1. Public Images: Pre-built images maintained by the community or vendors, available through public registries such as Docker Hub.
2. Private Images: Your own images stored in the TIR Container Registry or built using the TIR Image Builder.

To learn more about instance images such as Pre-built, Base OS, and Custom images click here.

To view more about the availabe images view the full list on the Supported Image List page.

Plans

Plan	Billing Model	Best For
On-Demand	Hourly, pay-as-you-go	Short-term, unpredictable, or test workloads
Committed	Upfront for a fixed period (1-month, 6-month, etc.)	Long running or predictable production workloads