VPU Deployment Playbook

Overview

Start Here

Architectures

Explore

Test

Migrate

Scale

Test

The purpose is to validate that Quadra VPUs deliver measurable performance and economic gains using your real workloads before any migration decisions are made. Testing is about evidence, not extrapolation.

This is not benchmarking hardware. It’s validating an operational decision.

What to test and why

Focus on the workloads that stress your system today by prioritizing what breaks first:

  • Codecs you run in production (H.264, HEVC, AV1)
  • Resolutions and frame rates that drive cost
  • Content types that reflect reality (motion, grain, live, UGC, surveillance)


Avoid synthetic content that flatters hardware and misleads decisions.

High Motion

Sports /
Live streams

Long Duration

Surveillance
/ Archives

Scale Bursts

UGC / 
Live Events

Compute Heavy

AV1 / HEVC

Test what breaks first.

How to run a valid test

Run VPUs side-by-side with your current compute.
Baseline requirements:

  • Same inputs
  • Same outputs
  • Same measurement window


Change one variable only: the compute layer. This is a compute substitution test, not a pipeline experiment.

What to measure

Focus on metrics that translate directly to operations and cost.
If a metric can’t be explained to finance or ops, it’s not primary.

Primary metrics:

  • Streams per server
  • Cost per stream
  • Watts per stream
  • Stability

Secondary metrics:

  • Startup time
  • Error rates
  • Resource utilization consistency

What “Good” looks like

A successful test shows:

  • Clear throughput gain at equivalent quality
  • Predictable utilization (no mystery saturation)
  • Stable operation across sustained runs
  • A measurable reduction in cost or power per stream


If results are ambiguous, adjust content or presets.
Don’t force conclusions.

The most common failures we see at this stage:

  • Changing multiple variables at once
 Invalid comparisons lead to false confidence.

  • Optimizing for peak results, not stability
 Short runs don’t predict production behavior.

  • Forcing conclusions from mixed data
 Ambiguity should trigger adjustment, not decisions.

Outputs of the Test stage

When testing is complete, you should have:

  • Documented benchmark results
  • A clear performance delta vs CPU/GPU
  • Confidence in operational behavior
  • A defensible recommendation to migrate, 
or not


If you can summarize results in one slide, 
you’re ready.

VPU-NATIVE INFRASTRUCTURE


Start testing VPU encoding in the cloud.

Start your pilot →

CUSTOMER deployment DE-RISKED

We shipped faster than your testing cycle.


Most vendors disappear between sales and deployment. We show up. We answer in hours. We ship while you’re still validating. That changes everything

“You guys added the features faster than we were able to verify them.”

customer Engineering team

Let's continue

If you can defend the evaluation rationale to a skeptical peer in two minutes, you’re ready to test.

< Go back
Next >