01 / Statement AMD GPUs and other non-CUDA accelerators Fig. A

Independent AI compute engineering

The fastest models in the world all run on the same handful of chips.

That’s a bottleneck. Seidia exists to break it.

We take the models you already run and make them work on accelerators the industry forgot: porting, optimising, benchmarking, and proving the numbers on real hardware.

  • 01

    Same model quality.

  • 02

    Same reliability.

  • 03

    Dramatically lower cost.

02 How we measure

Steps / 07

Replace assumptions with measurements.

Moving a workload to new hardware should be an engineering decision backed by evidence. We measure each option against your current baseline, so the choice is yours to make on the numbers.

  1. 01

    Define the workload

    We write down exactly what is being measured: model and weights revision, precision, the distribution of prompt and output lengths, concurrency, and the latency target the service must hold.

  2. 02

    Establish the baseline

    The same workload runs on your current hardware and serving stack. Every alternative is compared against this reference, not against a published figure.

  3. 03

    Pin the environment

    Drivers, firmware, compute platform versions, framework versions and container digests are recorded alongside every result, so a run can be reproduced later.

  4. 04

    Measure under load

    Runs include warm-up, repeat several times and hold load long enough to expose instability. We report distributions and percentiles rather than the best run.

  5. 05

    Check quality parity

    Outputs are compared with the baseline using evaluations agreed up front. Numerical divergence and behaviour changes are reported, not hidden.

  6. 06

    Price the result

    Throughput at the latency target becomes cost per million tokens, using hardware prices you provide or public pricing, with every assumption stated.

  7. 07

    Publish the evidence

    You receive the report together with the scripts, configurations and raw measurements, so your team can re-run any number in it.

03 The bottleneck

C / 01 - 04

When every workload depends on the same chips, everyone pays the same premium.

C / 01

Software lock-in

Models, kernels and serving stacks are written and tuned for CUDA first, so moving them is treated as a research project rather than an engineering task.

C / 02

Supply sets the price

When capacity is scarce, the cost of inference is decided by availability, not by what the workload actually needs.

C / 03

Compatibility gaps

Alternative accelerators often have capable hardware, but missing kernels, immature tooling and unfamiliar failure modes stop teams from trying.

C / 04

Claims nobody can reproduce

Published benchmarks rarely match your model, your traffic or your latency target, so they are hard to act on.

04 Specimen

Fig. B / Illustrative

Default toolchain versus Seidia-optimised, on one normalised benchmark

Every figure below is a placeholder. It shows the shape of the artefact, not a result.

Illustrative example, not a client result model-a / model-b Specimen / B

Normalised benchmark index, default toolchain = 1.00

  • Default toolchain
  • Seidia-optimised
  • Prefill throughput: default toolchain 1.00, Seidia-optimised 1.26. Higher is better. Placeholder value, not a measured result.
  • Decode throughput: default toolchain 1.00, Seidia-optimised 1.18. Higher is better. Placeholder value, not a measured result.
  • Throughput at batch size 64: default toolchain 1.00, Seidia-optimised 1.21. Higher is better. Placeholder value, not a measured result.
  • Time to first token, p95: default toolchain 1.00, Seidia-optimised 0.72. Lower is better. Placeholder value, not a measured result.
  • Cost per million output tokens: default toolchain 1.00, Seidia-optimised 0.61. Lower is better. Placeholder value, not a measured result.

Placeholder index values, illustrative only. Bars show the shape of the artefact, not a measured saving.

What was tested

model-a and model-b, at the agreed concurrency and latency target, on recorded drivers, framework versions and container digests.

What was not tested

Other model revisions, longer context windows, multi-node serving, and traffic above the agreed concurrency.

What you receive
  • A cost model with explicit, editable assumptions
  • Scripts, configurations and pinned environments to re-run every result
Illustrative example, not a client result. Figures are placeholders, not measured savings.
05 Independence

Dissent / 01

A recommendation, including when alternatives are not worth it yet

Our recommendation follows the measurements, including when the answer is to stay where you are.

07 Process

Steps / 05

Five steps, each with something you can check.

  1. 01

    Profile

    Agree the workload, the baseline, the metrics that matter and the quality bar it must meet.

  2. 02

    Port

    Bring the model and serving stack up on the target hardware and resolve compatibility barriers.

  3. 03

    Optimise

    Work through kernels, precision, batching and memory to close the gap to the baseline.

  4. 04

    Validate

    Re-run the agreed benchmark under pinned conditions and check output quality against the baseline.

  5. 05

    Deploy

    Hand over reproducible configurations, results and a cost model your team can act on.

08 Principles

Statements / 04

Claims are earned by measurement.

Independent by design.

We do not sell hardware or capacity. Our recommendation follows the measurements, including when the answer is to stay where you are.

Reproducible, or it does not count.

Every result ships with the environment, configuration and scripts needed to run it again.

Quality parity is non-negotiable.

A cheaper deployment that changes model behaviour is not a saving. Output quality is checked against the baseline every time.

Economics, not peak numbers.

We report cost per useful token at your latency target, not headline throughput under ideal conditions.

09 Terms

How we work

No rip-and-replace. No retraining. No vendor lock-in.

Your workload stays yours; we just make it run where it’s cheaper.

10 Today

We work directly with AI laboratories, inference providers and AI-driven businesses to benchmark their workloads, clear the software barriers, and put hard cost savings on the table.

  • AI laboratories

    Serve research and production models on more than one hardware ecosystem without rewriting your stack or compromising model quality.

  • Inference providers

    Add capacity on alternative accelerators with measured throughput, latency and cost per token before you commit to a fleet.

  • AI-driven businesses

    Understand what your inference bill would look like on different hardware, and what it would take to get there.

11 Tomorrow

The independent optimisation layer between AI and all of the world’s compute.

Every accelerator competing for your workload, every workload free to move, and no one held hostage by a single chipmaker.

12 / Next step

Bring a workload.We will measure it.

Tell us which models you serve, on what hardware, and what you need to improve. We will tell you honestly whether alternative compute is worth testing.