NVIDIA for HLS
Last reviewed: 2026-06-28 — see the freshness policy.
Learning objectives
After this chapter you will be able to:
- Place NVIDIA's healthcare stack (Clara, Parabricks, MONAI, BioNeMo, Holoscan, NIM) against HLS use cases.
- Explain why NVIDIA is an accelerated-compute layer that spans cloud and on-prem rather than a cloud itself.
- Decide when to reach for the NVIDIA stack and how it keeps PHI inside your boundary.
NVIDIA's HLS positioning
NVIDIA is not a cloud and not a clinical-data platform — it is the accelerated-compute layer beneath HLS AI, imaging, and genomics, plus a set of healthcare-specific SDKs and pre-trained models on top. Its GPUs run on every cloud (AWS/GCP/Azure all offer NVIDIA instances) and on-prem (DGX/HGX systems), unified by NVIDIA AI Enterprise — a supported software runtime. Think of it as the "picks and shovels": you compose it with the platforms in the previous chapters rather than instead of them.
This cross-cutting nature is why NVIDIA shows up in three different parts of this bootcamp: genomics (Part 7), medical imaging AI and drug discovery (Part 6), and on-prem/hybrid compute (Part 4).
The healthcare stack
flowchart TB
subgraph Hardware["Accelerated compute (cloud GPU or on-prem DGX/HGX)"]
AIE["NVIDIA AI Enterprise (supported runtime)"]
end
AIE --> Clara["Clara — healthcare platform umbrella"]
Clara --> Para["Parabricks<br/>genomics secondary analysis"]
Clara --> MONAI["MONAI<br/>medical imaging AI"]
Clara --> Holoscan["Holoscan<br/>real-time edge / device AI"]
AIE --> BioNeMo["BioNeMo<br/>drug discovery / biology"]
AIE --> NIM["NIM<br/>inference microservices"]
| Component | Domain | What it is |
|---|---|---|
| Parabricks | Genomics | GPU-accelerated secondary analysis (GATK, DeepVariant) — ~10 minutes for a 30× genome versus 15+ hours on CPU. See Sequencing pipelines and Genomics workflow orchestration for the detailed DRAGEN-vs-Parabricks trade-off. |
| MONAI | Medical imaging | Open-source PyTorch framework for medical-imaging deep learning (co-developed with King's College London); MONAI Deploy packages models for clinical use. See Medical imaging AI. |
| BioNeMo | Biotech / drug discovery | Generative-AI platform for biology — protein structure, molecular generation, and other biopharma models; widely adopted by life-science R&D. |
| Holoscan | Medical devices | Real-time, low-latency edge AI for streaming sensor/video data (surgical video, ultrasound, endoscopy). |
| Isaac for Healthcare | Robotics | Simulation and AI for medical robotics and surgical systems. |
| NIM | Deployment | NVIDIA Inference Microservices — containerized, optimized model-serving you self-host. |
| Clara | Umbrella | NVIDIA's healthcare platform brand spanning the above imaging/genomics/device SDKs. |
Cloud, on-prem, or both
Because the stack runs anywhere NVIDIA GPUs run, it sits naturally in hybrid HLS designs (see On-premises & hybrid):
- Cloud GPUs — burst Parabricks or imaging-AI jobs onto cloud GPU instances; release them after.
- On-prem DGX/HGX — keep large genomics or imaging workloads on owned accelerated compute, co-located with the data (sunk cost, data-local, data sovereignty).
- Edge — Holoscan runs inference next to the device, where latency and connectivity demand it.
PHI, NIM, and keeping data in your boundary
The most architecturally significant piece for an SA is NIM: self-hosted, containerized model inference. Because you run the model inside your own environment (your VPC or your on-prem cluster), PHI never leaves your boundary to reach a third-party model API. That sidesteps a major HIPAA and data-residency concern (no new business associate, no cross-border transfer) and is often the deciding factor for clinical GenAI in regulated or sovereign settings — compare with calling an external LLM API, which sends data out.
flowchart LR
PHI["PHI / clinical data"] --> NIM["NIM (self-hosted)<br/>in your VPC / on-prem"]
NIM --> Result["Inference result"]
NIM -. "no PHI egress" .-x Ext["External model API"]
When to reach for the NVIDIA stack
- Genomics at scale or under SLA — Parabricks when CPU pipelines are too slow or too costly (see HealthOmics for the managed alternative; Parabricks for raw acceleration on your own compute).
- Medical imaging AI — MONAI as the framework for training/deploying imaging models.
- Drug discovery / computational biology — BioNeMo for protein and molecule models.
- Medical-device / real-time edge AI — Holoscan.
- Self-hosted GenAI over PHI — NIM, to keep data in-boundary.
- Data sovereignty / on-prem mandate — the whole stack on DGX where data cannot leave.
Labs
RNASEQ— the nf-core genomics pipeline; Parabricks is the GPU-accelerated path for the secondary-analysis steps.- Medical-imaging and BioNeMo labs are good extension exercises (see the bootcamp format for the lab model).
Check yourself
- Why is NVIDIA described as a layer you compose with AWS/GCP/Azure rather than an alternative to them?
- A clinical team wants GenAI over patient notes but cannot send PHI to an external API. Which NVIDIA component addresses this, and how?
- A diagnostics lab needs to cut its whole-genome secondary-analysis time from hours to minutes on its own hardware. Which component fits, and which managed alternative would you compare it against?