Consulting

02 Validate Flagship

GPU Workload Assessment & Validation

Does our hardware actually deliver what our workloads require?

  1. Discover Planning & Architecture Know what you need before you buy.
  2. Validate GPU Workload Assessment Prove what your hardware can deliver.
  3. Optimise Performance & Efficiency Get more from the infrastructure you own.
02 Validate. Three stages, each independently purchasable. Start where you are, or scope two or more together as one programme.

Who it is for

Clients with existing or newly acquired GPU infrastructure, including deployments delivered by other suppliers, who want independent evidence of performance and capacity.

What's included

  • Workload and infrastructure baseline.
  • Benchmark setup on your own hardware, in your own environment.
  • Defined latency, throughput and concurrency sweeps, plus agreed inference engine and precision comparisons.
  • Power and thermal analysis: power-limit sweep, performance per watt and sustained-load thermal behaviour, because these decide what a rollout can actually sustain.
  • Bottleneck analysis, capacity estimates and hardware and expansion recommendations.
  • A professionally typeset written report with numbered findings, an evidence register, agreed raw measurements and checksums.
  • A client-specific test harness that reproduces the agreed measurements on your hardware using open-source tools, licensed for your internal use. TensorFoundry's own benchmarking toolkit is not delivered.
  • Findings walkthrough.

Scoped on request, not included by default

  • Resilience and failover testing.
  • Multi-node scaling.
  • Extensive architecture comparisons.
  • Additional workloads.

Reference scope

One site, one defined GPU configuration and one representative workload, with a test matrix agreed before commencement.

What's not included

  • Unlimited tests.
  • Remediating every issue identified.
  • Production ownership.
  • Procurement.
  • Ongoing monitoring.
  • Optimisation implementation, which is scoped separately.

Price and timing

From $15,000 AUD ex GST

Typically 8-10 days of engineering effort over 3-5 weeks, depending on your readiness and access.

The starting fee assumes a bounded workload, site and platform scope. Additional complexity, architectures, workloads, sites or testing are quoted separately. Elapsed time is not engineering effort.

Discuss your workload

Frequently Asked Questions

Is the testing done on our own hardware?

Yes. Measurements are taken on your own hardware, in your own environment, against a test matrix agreed before commencement.

Can you assess infrastructure another supplier delivered?

Yes. Validate is for existing or newly acquired GPU infrastructure, including deployments delivered by other suppliers. The assessment is independent of whoever supplied the hardware.

What do we keep at the end?

A typeset written report with numbered findings, an evidence register, agreed raw measurements and checksums, and a client-specific test harness licensed for your internal use, so you can reproduce the agreed measurements yourself. TensorFoundry's own benchmarking toolkit is not delivered.

Why does elapsed time run longer than the engineering effort?

Elapsed time depends on your readiness and on access to the hardware, so the engineering effort is usually spread across several calendar weeks.

Is resilience or multi-node testing included?

Not by default. Resilience and failover testing, multi-node scaling, extensive architecture comparisons and additional workloads are scoped on request, where they are feasible.

Hardware installed, or about to be?

Get independent evidence of latency, throughput, capacity and power on your own infrastructure before you scale the rollout.

Discuss your workload Explore our services

Other stages

Three stages, each independently purchasable. Start where you are, or scope two or more together as one programme.