HybridDev sandboxFine-tuningLLM inference

GPU Development Sandbox

A low-friction environment for developers to prototype on GPUs — notebooks, quick fine-tunes, and inference experiments — with just enough isolation and cost control to keep it from sprawling.

Customize in Designer
Business objective

Give developers quick, safe GPU access to experiment and validate before promoting workloads to production platforms.

Personas
ML engineerData scientistPlatform engineer
Key assumptions
  • ·Small, interactive, bursty jobs — not large training runs.
  • ·Speed of access and isolation matter more than peak performance.
  • ·Idle cleanup is needed or costs sprawl.
Logical architecture

Rendering diagram…

Design

Compute

Partition GPUs (MIG/time-slicing) so many developers share a few cards; whole-GPU access for short bursts via a queue.

Network

Standard networking; a path to promote artifacts to production registries.

Storage

Shared datasets and images plus modest per-user workspaces with quotas.

Kubernetes

Namespaced multi-tenant cluster; JupyterHub/notebook operator or dev-container flow; quotas and an idle reaper.

Security

Authenticated access, per-user isolation, no production data by default, and scanned base images.

Observability

Per-user utilization and idle detection; simple cost visibility to keep experimentation honest.

High availability
Not critical — best-effort. Losing the sandbox pauses experiments but not production.
Disaster recovery
Workspaces backed up; environment reproducible from IaC. Low RPO/RTO expectations.
Capacity
Sized to concurrent developers × a fraction of a GPU via partitioning. Small pool goes a long way.
Cost profile
Kept low by partitioning, quotas, and aggressive idle reclamation. The failure mode is silent sprawl.
Risks
  • mediumIdle notebooks hold GPUs indefinitely without a reaper.
  • lowSensitive data leaking into an under-controlled sandbox.
Implementation phases
  1. 1

    Access

    Notebook/IDE access on partitioned GPUs with per-user isolation.

  2. 2

    Guardrails

    Quotas, idle reaper, scanned images, cost visibility.

  3. 3

    Promotion path

    A clean route from sandbox artifact to production platform.

Validation checklist
  • A developer gets a GPU session in minutes.
  • Idle sessions are reclaimed automatically.
  • Users are isolated from each other's workspaces.
Recommended GPUs for this pattern
NVIDIA

L40S

48 GB · 350 W

NVIDIA

A30

24 GB · 165 W

NVIDIA

RTX 6000 Ada

48 GB · 300 W

AMD

Radeon Pro W7900

48 GB · 295 W