Consulting

03 Optimise

Inference Performance & Efficiency

How can we get more from the hardware and workloads we already run?

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

Who it is for

Organisations with existing inference workloads and concerns about latency, throughput, concurrency, utilisation, memory use or runtime configuration.

What's included

  • Workload baseline and performance profiling.
  • Diagnosis of runtime, batching, concurrency, model precision, GPU memory and topology bottlenecks.
  • Targeted tuning of vLLM, SGLang, llama.cpp or other appropriate inference stacks.
  • Relevant Linux, driver and GPU configuration advice and agreed changes.
  • Controlled before and after tests and documented recommendations.
  • Change record, measured findings and technical handover.

What's not included

  • A committed performance uplift.
  • Major application rewrites.
  • Unrestricted administration.
  • Hardware purchase.
  • 24/7 support.
  • Continuing infrastructure operations.

Price and timing

From $12,000 AUD ex GST

Typically 4-6 days of engineering effort over 1-3 weeks.

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

Which inference stacks do you tune?

vLLM, SGLang, llama.cpp or other appropriate inference stacks, together with the relevant Linux, driver and GPU configuration.

How do we know the changes helped?

Changes are measured with controlled before and after tests, and you receive a change record with the measured findings and documented recommendations.

Do you commit to a specific improvement?

No. Results depend on your workload, hardware and current configuration, so we measure rather than promise. You receive the measured findings either way.

Do we need a Validate assessment first?

No. Each stage is independently purchasable. If you would like both, Validate and Optimise can be scoped together as one programme.

Is ongoing support part of the service?

No. Optimise ends with a change record, measured findings and a technical handover to your team. Continuing operations stay with you.

Inference slower or costlier than it should be?

Find where latency, throughput and memory are being lost, and finish with measured before-and-after results and a documented handover.

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.