Skip to content

AI InfrastructureCase study

Spyre CI/CD Modernization

Rebuilt how IBM tests its Spyre AI accelerator platform: from sequential runs on persistent Jenkins agents to a parallel Tekton pipeline on ephemeral Kubernetes pods running on bare-metal IBM Power.

Company
IBM
Role
Software Engineer Intern
Timeline
May – Aug 2026

Impact

6x

Faster AI/ML test runtime

9

Parallel Tekton pods

POC

Validated for wider Spyre rollout

The problem

AI/ML test suites for the Spyre accelerator ran one after another on persistent Jenkins agents, so every change waited on a long, serial feedback loop.

Persistent agents also held on to bare-metal IBM Power machines with Spyre cards attached, keeping expensive accelerator hardware tied up even when nothing was running.

What I built

  1. 1

    Kubernetes on bare metal

    Architected Kubernetes-based CI/CD directly on bare-metal IBM Power hardware with Spyre AI accelerators, so test workloads could be scheduled onto the accelerator nodes.

  2. 2

    Ephemeral pods instead of persistent agents

    Migrated persistent Jenkins agents to ephemeral pods that are created for a run and released afterwards, so the hardware is only claimed when it's actually needed.

  3. 3

    Fan out with Tekton

    Engineered a Tekton pipeline that splits the test workload across 9 pods running in parallel, replacing the sequential Jenkins execution.

  4. 4

    Prove it, then scale it

    Ran it as a proof of concept against the existing pipeline and validated it as the model for broader CI/CD across Spyre infrastructure.

How it flows

  1. 01

    Code change

    Triggers a Tekton PipelineRun

  2. 02

    Schedule

    Ephemeral pods land on Spyre-equipped Power nodes

  3. 03

    Fan out

    9 pods run AI/ML test shards in parallel

  4. 04

    Release

    Results reported and pods torn down

The results

  • AI/ML test runtime dropped 6x compared with sequential Jenkins execution.
  • Accelerator hardware is used on demand instead of being held by always-on agents.
  • The proof of concept was validated for broader CI/CD implementation across Spyre infrastructure.

Stack

  • Kubernetes
  • Tekton
  • Jenkins
  • IBM Power
  • Spyre AI Accelerator
  • CI/CD

Next case study

Bobbotics Hardware Test Automation Tool