Research

The standing research programme at AI ThinkLab: low-bit quantisation, generative models at the edge, kernel autotuning, and evidence for optimised models.

Client work pays for the lab. The lab is what makes the client work worth buying.

AI ThinkLab keeps a standing internal programme rather than treating research as something that happens between projects. The rule we apply to it is narrow: we work on questions whose answers change what we would do on an engagement within about a year. Everything else is interesting reading, not a research line.

Current lines of enquiry

Low-bit quantisation past INT8

How far INT4 and mixed-precision schemes can be pushed on hardware whose compiler was designed for INT8 — where the accuracy actually goes, which layers refuse to cooperate, and when the dequantisation overhead quietly eats the gain.

Generative models at the edge

Small language and vision-language models on embedded SoCs: memory-bandwidth limits on decode, KV-cache strategies inside a fixed arena, speculative decoding when there is no second accelerator, and what a usable tokens-per-second actually feels like in a product.

Kernel and compiler autotuning

Search-based tuning of tiling, vector width and scheduling against real silicon rather than a cost model — and how much of the hand-written kernel work can honestly be automated before a human has to look at the disassembly again.

Evidence that an optimised model still behaves

Optimisation changes numerics. We are building the harness and the argument structure that lets a customer demonstrate to an auditor — or to themselves — that a quantised, pruned, re-scheduled model is still the model they validated.


How the programme is run

  • Reproduce before believing. A published result counts once it runs on hardware we hold, at the numbers claimed.
  • Measure on silicon, not in simulation. Cost models are a hypothesis; a board with a power meter on it is evidence.
  • Publish the negative results. Knowing which technique does not survive an NPU compiler is worth as much as knowing which does.
  • Feed it back into tooling. Anything learned twice becomes part of the harness, so the next engagement starts further along.

Publications & write-ups

Write-ups, reproductions and engineering notes are posted under Insights. Measured results go to Benchmarks. Where a piece of work is publishable in a venue rather than a blog post, it will be listed here.

Collaboration

We are open to joint work with university groups and with silicon vendors who want an independent evaluation of their toolchain — including one that says something they will not enjoy reading. Write to us if that describes you.

Sponsor a research line.

Some of what is on this page is directly useful to a company before it is useful to us. If one of them is on your critical path, it can be funded and accelerated.