HN Showcase
Back to browse
VGPU Silo Capacity Calculator (identify siloed capacity in mixed mode)

VGPU Silo Capacity Calculator (identify siloed capacity in mixed mode)

by frankdenneman·Feb 25, 2026·3 points·0 comments
Visit ProjectView on HN

AI Analysis

●●SolidSolve My ProblemBig BrainNiche Gem

Reveals silent GPU capacity loss in vGPU mixed-mode—niche infra pain made visible.

Strengths
  • •Non-obvious insight: fragmentation from profile compatibility rules wastes capacity even with free memory—the visualization makes this concrete.
  • •Interactive placement simulation (manual or Monte Carlo) lets operators reason about tradeoffs before committing to a profile catalog.
  • •Tied to real NVIDIA hardware SKUs and vGPU profiles with cited sources—concrete, not theoretical hand-waving.
Weaknesses
  • •Extreme niche: only relevant to NVIDIA vGPU deployments in mixed-mode; excludes hyperscalers using native MIG or full-GPU allocation.
  • •Educational tool, not a solver—doesn't recommend optimal profile catalogs or automate placement, just helps visualize the problem.
Target Audience

GPU infrastructure architects, Kubernetes/virtualization operators, datacenter capacity planners

Similar To

NVIDIA vGPU documentation tools · Capacity planning simulators

Post Description

I built a small online tool to help visualize a problem that shows up in mixed-size vGPU environments: GPU capacity can become siloed even when memory is still free.

In these setups, deployable capacity is limited by placement rules and profile compatibility, not just by available memory. Two clusters with the same free memory can behave very differently depending on how profiles are mixed and placed over time.

This tool helps identify that siloed capacity.

You can:

- choose a GPU type and vGPU profile catalog - model mixed-size placement behavior - see how profile choices influence deployable capacity over time

It’s not a placement simulator and doesn’t try to predict exact VM startup order. The goal is to make placement constraints visible so architects and operators can reason about profile catalogs before deployment.

Tool: https://frankdenneman.nl/tools/vgpu-silo-capacity-calculator...

Background article explaining the model: https://frankdenneman.nl/posts/2026-02-24-mixed-size-vgpu-mo...

Feedback welcome, especially from people running virtualized GPU or shared inference environments.

Similar Projects