GIKSN Research is an independent lab. Our active bench is in AI. The Hardware sector is where the compute substrate for that work gets written about honestly. Some of that writing is ours because AI writing without hardware honesty is a marketing document. Most of it is expected to come from contributors whose day-to-day work is here.

Why Hardware matters to our AI work

AI is a hardware problem before it is a model problem. NPUs, mobile CPUs, memory bandwidth, cache hierarchies, thermal envelopes and interconnect all decide what a given model can actually do at runtime. We work assuming honest hardware numbers. Where honest numbers do not exist, we ask for them. This sector is where they land.

What Hardware writing here looks like

A hardware paper on GIKSN treats specifications as claims. Claims are arguments. TOPS is not enough. We want to know at what precision, at what batch size, at what sequence length and with what memory bandwidth utilisation. If a number is estimated the author says so. If a design axis loses on a workload it was not built for the author says that too. The point is that the paper survives its own numbers a year later.

Surveys of parts and subfields are welcome. Original research is welcome when the author has actually run the benchmark. Vendor decks are not.

Community weight

The bulk of hardware writing on GIKSN is expected to come from vetted contributors. The lab does not tape out chips. What we do is publish honest evaluations of publicly available parts when they touch our own AI work, plus reference tooling and characterisation notes when useful. Contribution is gated. Applications go through the About page. Accepted contributors join the private Telegram working channels alongside the public read-only one.