Request Callback
AI Workbench

AI Workbench: GPU-Powered AI environments ready in minutes

GPU Virtual Machine NVIDIA
Request Callback

Overview of AI Workbench

With AI Workbench, you get a GPU-backed compute environment on CloudXP, purpose-built for AI development and accelerated workloads. It delivers a fully configured virtual machine with your choice of GPU hardware — NVIDIA H200 NVL, NVIDIA H200 SXM, or AMD MI300X — with pre-installed GPU drivers and an up-to-date CUDA or ROCm stack, so your team can run model training, deep learning inference, RAG pipelines, and data-intensive experiments without manually setting up the underlying infrastructure.

What makes AI Workbench unique is the AI software layer: at the time you place your provisioning order, you choose the exact AI tools you want installed — including inference and serving (Ollama, vLLM, Open WebUI), notebooks and MLOps (JupyterLab, MLflow, LangFlow), vector search (ChromaDB, Weaviate, pgvector), deep learning (PyTorch, TensorFlow), observability (Grafana, Prometheus, InfluxDB), workflow automation (n8n), and data infrastructure (PostgreSQL, Redis, RabbitMQ) — or select a named bundle like the Generative AI Stack to deploy a complete, pre-validated toolkit in one step. Everything you select arrives on your VM configured, compatible, and running the moment provisioning completes.

How it works

Select your tools

Choose your AI tools and bundle from the AI Workbench product on CloudXP Marketplace at order time.

AI Workbench | CloudXP Market place

Your VM gets provisioned

Your VM is created on Ubuntu 24.04 LTS, and your GPU is configured with CUDA or ROCm.

Ubuntu 24.04 | H200 NVL | H200 SXM | MI300X

Your tools are ready

Your AI tools go live, pre-installed and running, with nothing left to set up.

Pre-installed | Ready in <30m

18+
AI Tools
3
GPU Models
<30 m
Ready time
0
Manual Installs

GPU layout variants

AI Workbench gives you three GPU layouts on CloudXP for training, inference, and memory-heavy jobs. When you provision, you pick your accelerator and, in the same step, the AI tools or pre-validated bundle you want — notebooks, inference servers, vector databases, ML frameworks, observability, and more — version-tested for that GPU so services are ready when the VM activates. Every instance runs Ubuntu 24.04 LTS with drivers pre-installed; NVIDIA layouts include CUDA, AMD MI300X includes ROCm. Use on-demand billing or a reserved plan (1-, 6-, or 12-month).

NVIDIA H200 NVL

Best for - Large-scale inference and NVLink-connected multi-GPU deployments

GPU profile - 141 GB HBM3 per GPU with NVLink interconnect - layout sizes and instance counts follow your Marketplace SKU and quota

Typical workloads: Serving, RAG pipelines, and experimentation across linked GPUs

NVIDIA H200 SXM

Best for - Intensive training and jobs that need SXM5 ultra-high bandwidth

GPU profile - 141 GB HBM3 per GPU on the SXM5 form factor

Typical workloads - Model training, fine-tuning, and compute-heavy deep learning

AMD MI300X

Best for - Memory-bound models, large context windows, and ROCm-accelerated pipelines

GPU profile - 192 GB HBM3 per accelerator - the largest on-board memory option in the lineup

Typical workloads - Enterprise training and inference when VRAM capacity is the limiting factor

Core features of AI Workbench

AI tool selection at order time
Choose individual AI tools or a complete software bundle when you place your provisioning order. Tools arrive pre-installed — no post-provisioning setup, no manual configuration.
Pre-validated software bundles
Deploy a complete AI stack - such as the Generative AI Stack - in one selection. Every tool in the bundle is validated at compatible versions for the selected GPU layout
Real-time milestone tracking
Track provisioning progress through every step - VM Creation, GPU Driver Installation, CUDA Setup, Software Installation, VM Activation - with live status and error detail.
Multi-instance provisioning
Set the instance count and provision multiple identically configured workbenches in a single order. Every researcher on the team gets the same environment, every time.
Flexible data disk management
Attach multiple SSD or HDD data disks at provisioning time for datasets, model weights, and experiment outputs. Storage survives VM reprovisioning — your data stays safe.
Three GPU layout options
Select the GPU that fits your workload: NVIDIA H200 NVL (141 GB, NVLink), NVIDIA H200 SXM (141 GB, SXM5 high-bandwidth), or AMD MI300X (192 GB HBM3). All layouts include pre-installed drivers and the full compute stack — CUDA or ROCm.

What you get with AI Workbench

Complete AI stack, zero setup
From order to running AI tools in under 30 minutes. JupyterLab opens in your browser, Ollama serves your LLM, MLflow tracks your experiments — every service started and configured automatically. Your team starts building on day one, not day three.
Compatibility guaranteed by design
Software bundles define pre-tested version combinations — every tool in the bundle is validated to work with the selected GPU layout and with each other. No CUDA version hunting, no dependency conflicts, no version-mismatch problems across your team.
Three GPU layouts: pick your power
Choose the GPU layout that fits your workload: NVIDIA H200 NVL (141 GB, NVLink interconnect) for large multi-GPU inference, NVIDIA H200 SXM (141 GB, SXM5 ultra-high bandwidth) for intensive training runs, or AMD MI300X (192 GB HBM3) for the largest context-window and memory-bound workloads. All ship with drivers and compute stack pre-installed.
Enterprise governance from day one
Every workbench is assigned to a hierarchy, project, and subscription at provisioning time. Billing allocation, resource quotas, and access boundaries are set before the VM starts — giving platform teams full visibility and cost control across every AI environment in the organisation.
Team environment parity
Provision multiple instances in a single order. Every researcher receives an identical, reproducible environment — same tools, same versions, same GPU configuration. Eliminate environment drift and onboarding delays across your entire AI team with one provisioning order.
Generative AI stack bundle
Deploy a complete Generative AI development environment in one bundle selection: Ollama, Open WebUI, JupyterLab, ChromaDB, MLflow, LangFlow, Grafana, and Prometheus — all installed, all compatible, all running. Go from order to a working RAG pipeline in under 30 minutes.

FAQs about AI Workbench

18+ tools are selectable when you order, across inference, notebooks, MLOps, vectors, monitoring, orchestration, frameworks, and data services: LLM inference (Ollama, vLLM); Notebooks (JupyterLab, TensorFlow-Jupyter); MLOps & data (MLflow; ChromaDB, Weaviate, pgvector; PostgreSQL, Redis, RabbitMQ); Monitoring (Grafana, Prometheus, InfluxDB); Orchestration & UI (LangFlow, n8n, Open WebUI); Frameworks (PyTorch, TensorFlow, CUDA-ready on NVIDIA layouts).
Both install during provisioning; the difference is how versions are chosen. Individual tools: you pick tools and versions — best when you need specific stacks or uncommon combinations. Bundles: pre-validated sets (e.g. Generative AI Stack: Ollama, Open WebUI, JupyterLab, ChromaDB, MLflow, LangFlow, Grafana, Prometheus) tested together on your GPU layout — best for teams and fast ramp-up. If you choose a bundle, it replaces conflicting individual selections so the tested configuration stays intact.
Base VM plus GPU drivers and CUDA or ROCm usually finishes in 5–15 minutes for any supported layout (H200 NVL, H200 SXM, MI300X). Heavy tool sets or bundles can run up to ~30 minutes while images install. Track progress in Track Order (VM creation through activation). Every milestone must succeed before the workbench is usable.
Yes. Set No. of Instances on the order — every VM shares the same GPU layout, tools or bundle, and disk layout; each instance gets its own disks and a hostname suffix (e.g. ai-workbench-01, ai-workbench-02). One order and milestone view covers all instances. GPU quota in the zone must cover every instance. Pricing scales with instance count — review Check Price before submit.
After milestones succeed, SSH in with the admin credentials from provisioning. Selected tools run as Docker services that start on boot. Examples: JupyterLab http://:8888; Ollama :11434; Open WebUI often :3000; MLflow UI often :5000. Run docker ps for ports and mappings; open matching inbound rules on your security group.

Ready to power your AI team

Please provide the necessary information and our team will reach out with GPU layout options, bundle details, pricing, and a personalised demo.
image
Captcha
By selecting 'Submit', you authorise Jio Platforms Limited to store your contact details for further communication.
Submit
Cancel