Secure AI Workloads in ComputerVault Virtual Infrastructure
Overview
This report benchmarks NVIDIA A2 vGPU profiles in ComputerVault’s demo environment to showcase the
capabilities of the ComputerVault AI Platform. The tests were deliberately run on the lowest-tier GPU card
and vGPU profiles to establish a clear performance baseline.
The findings demonstrate the platform’s suitability for AI workloads in life sciences and pharmaceutical
research within a fully managed, on-premises ComputerVault private cloud and the AI expertise of
ComputerVaut’s team. Testing covered inference, image generation, and tool-augmented reasoning,
mapping performance tiers to specific GPU configurations. The study also addresses intellectual property
(IP) protection and data privacy, underscoring ComputerVault’s ability to deliver bare-metal-like
performance in a secure, compliant, and virtualized infrastructure.
1. Benchmark Summary
VM Image (AI-relevant components)
• Ubuntu version: 24.04.2 LTS
• NVIDIA driver version: 550.90.07
• CUDA Toolkit version: 12.4
• Python version: 3.10.18
• Conda version: 25.5.1
Python AI Libraries by Maintainer
• PyTorch: torch, torchvision, torchaudio
• Hugging Face: transformers, diffusers, accelerate, huggingface_hub, safetensors
• LangChain: langchain, langsmith
• NVIDIA: nvidia-cudnn-cu12, nvidia-cublas-cu12
Tested Hardware Profiles
• A24Q (4 GB VRAM): Entry-level AI capability
• A28Q (8 GB VRAM): Mid-tier AI capability
• A216Q (16 GB VRAM): High-tier AI capability
Workloads Benchmarked
• Vision tasks: Google ViT Base
• LLM inference: Qwen2 1.5B, Qwen3 4B, Qwen3 8B
• Image generation: Stable Diffusion 2.x
• Tool-augmented reasoning: Command summarization with Qwen3 8B
Key Findings by Profile:
Profile
Max Model
Size
Vision
LLM
Image Gen
Training
4 GB
≤1.5B
(quant.)
Yes
Yes
No
No
8 GB
≤4B (quant.)
Yes
Yes
Yes
(512×512)
No
16 GB
≤8B (quant.)
/ ≤7B LoRA
Yes
Yes
Yes
Yes (light)
2. Strategic Advantages
Technical Benefits
Bare-metal-like performance in virtual machines with NVIDIA vGPU profiles
• Flexible scaling from basic inference to small-scale training
• Support for drug discovery, molecular modeling, and complex simulations
• Dynamic resource allocation without needing dedicated hardware solely for AI
Security & Compliance
In pharmaceutical R&D, data sovereignty and IP control are as important as raw compute power.
ComputerVault’s architecture addresses this directly:
1. Local Execution – All inference, finetuning, and data processing occur inside your own VMs, backed by
dedicated NVIDIA vGPU hardware.
2. No Public Cloud Dependency – No prompts, datasets, or results sent to third-party AI APIs.
3. Data Residency Control – You choose the datacenter or colocation; hardware is owned by you.
4. Tenant Isolation – vGPU profiles and NVIDIA licenses are dedicated to your infrastructure.
5. IP Ownership Assurance – All trained or finetuned models remain your property.
6. Controlled Access – ComputerVault staff manage infrastructure from a separate NOC domain; no access
to customer’s user domain or research data.
3. Why It Matters
With VMware, Citrix, and Nutanix introducing higher complexity and cost — and public cloud AI costing up
to 5× or more — ComputerVault offers a simpler, more secure, and more cost-effective AI infrastructure.
For ComputerVault’s AI customers, this means:
• Lower TCO over 5 years
• Rapid deployment without sacrificing performance or security
• Full compliance with regulatory data handling requirements
• Future-proof scalability to adopt more powerful GPUs like NVIDIA L40 for larger models
Conclusion
The benchmark results confirm that ComputerVault can deliver scalable AI capabilities — from lightweight
inference to generative AI and limited training — entirely within your controlled infrastructure. For
pharmaceutical R&D teams, this balance of performance, cost-efficiency, and security creates a powerful
platform for innovation without compromising sensitive data.
Suggested Next Step: A ComputerVault pilot deployment in customer’s datacenter or a chosen colocation
facility can validate these results in a production-like environment — mapping workloads to the most cost-
effective GPU profile while ensuring IP and data security remain fully protected.
