Services

Engineering that moves inference onto more than one kind of hardware.

Four services that can run on their own or as one engagement. Each produces evidence and artefacts your team keeps, whether or not you move.

01

Workload benchmarking

Measure your models on alternative hardware against your current baseline, using your traffic shape and your quality bar.

We start from the workload you actually run, not a vendor demo. Together we agree the model, the request profile, the latency target and the quality checks, then measure it on your current hardware and on the alternatives.

What we do

  • Workload definition: model revision, precision, prompt and output lengths, concurrency
  • Baseline run on your current hardware and serving stack
  • Like-for-like runs on AMD GPUs and other candidate accelerators
  • Output quality comparison against the baseline

What you get

  • Benchmark report with distributions, not best runs
  • Scripts, configurations and pinned environments to re-run every result
  • A clear statement of what was and was not tested
02

Porting and compatibility

Get models, kernels and serving stacks running on ROCm and other non-CUDA software stacks.

Most of the work of moving off a single ecosystem is unglamorous: missing operators, custom kernels, framework version drift and behaviour that only appears under load. We find and resolve those barriers so the workload runs correctly first.

What we do

  • Compatibility audit of the model, custom operators and serving stack
  • Porting CUDA-specific kernels and dependencies to the target platform
  • Fixing correctness issues, numerical differences and stability under load
  • Upstreaming fixes where it helps the wider ecosystem

What you get

  • A working, tested build of your workload on the target hardware
  • Documented changes and the reasons for each
  • Known limitations and recommended follow-up work
03

Hardware-aware optimisation

Tune kernels, precision, memory layout and serving configuration for the accelerator you are running on.

Running is not the same as running well. Once a workload is correct on new hardware, we work through the layers that decide its efficiency, from kernel selection to batching policy, and measure each change against the agreed benchmark.

What we do

  • Profiling to find where time and memory actually go
  • Kernel selection, tuning and, where justified, custom kernels
  • Precision and quantisation choices, checked for quality parity
  • Batching, scheduling, parallelism and memory configuration

What you get

  • Optimised configuration with before-and-after measurements
  • A change log tying each gain to a specific change
  • Guidance your team can apply to future models
04

Compute economics and deployment

Turn measured performance into cost per token, capacity plans and a clear deployment decision.

Performance only matters at the price you pay for it. We convert benchmark results into the economics of serving your workload, state every assumption, and support the move into production when the numbers justify it.

What we do

  • Cost per million tokens at your latency target, per hardware option
  • Capacity and fleet sizing for your expected traffic
  • Sensitivity to hardware pricing, utilisation and model changes
  • Deployment support, monitoring and rollback planning

What you get

  • A cost model with explicit, editable assumptions
  • A recommendation, including when alternatives are not worth it yet
  • Runbooks for deploying and operating the chosen configuration

Hardware focus

AMD GPUs first. Other non-CUDA accelerators as their software matures.

We start where alternative hardware is already capable and the remaining barrier is mostly software. We work across the layers between your model and the silicon.

  • ROCm
  • HIP
  • PyTorch
  • Triton
  • vLLM
  • SGLang

Where to start

Most engagements begin with one workload.

A scoped benchmark of a single model is the quickest way to find out whether alternative hardware is worth your time. It produces evidence you keep either way.