We replace the datasheet with
real workloads.
So your strategic decisions are based on solid data, not guesswork.
Mission
aistack is a public and open benchmarking initiative, launched by imec, that aims to make sure that model & hardware builders, infrastructure & application providers can make informed decisions on the future of their AI stack. We need to make sure the right silicon is built for the right models and vice versa. This starts with getting a clear and unbiased view on how a set of AI applications perform on a diverse set of hardware.
True co-design and co-optimization across the different layers of this complex stack can only be done through solid datapoints generated where real AI workloads meet real systems & silicon.
As the worldwide leading R&D chip lab, we intend to unite the hardware and software ecosystem. If you're interested in joining us on this adventure, reach out!
Who we work with.
Whatever you build — the models, the infrastructure, the application on top, or the hardware itself — you cannot design your part without being influenced by the others. That's why all parts of the stack are brought together at aistack. We're here to feed the community with open benchmarking results that bridge the gap between AI model builders and chip providers.
AI model builders
Your checkpoints, running on silicon the market can't buy yet.
Architecture decisions outlive their hardware. We run your checkpoint on pre-release silicon and hand back a real trace — so you design for the chip you'll ship on, not the one on your desk.
- Frontier checkpoints on pre-release silicon
- Hardware traces, never estimates
- Architectures compared head-to-head
Infrastructure providers
How your stack holds up under workloads that don't exist yet.
Capacity planning is a bet on demand you can't see. We put tomorrow's load on today's stack — showing where it fits, saturates and stalls — while the purchase order is still a draft.
- What fits, what saturates, what stalls
- The load you'll face in 18 months
- Capacity bets de-risked before the racks ship
Hardware builders
Independent numbers on your chip — the kind buyers actually trust.
We don't sell anything and we don't take sides, which means the benchmark results we publish for your silicon are numbers a customer (and your investors) can check for themselves.
- Vendor-neutral, hardware-instrumented runs
- Results a buyer can reproduce
- The bench as a sales tool
Application builders
What your app really costs per request, before you commit to a stack.
Cost per request decides whether a product works, and it only shows when real workload meets real hardware. We measure yours end to end — so the stack you pick is one your margins survive.
- Cost and latency per request, measured
- Self-hosted versus API, like for like
- The trade-off priced before you commit
Independence
Loyal to the data.
No favourites, no vendors, no exceptions.
Reproducibility
Verify the run.
Every number comes from real workload data you could reproduce yourself.
Transparency
The data speaks.
We learn the most when we're honest.
Collaborating minds.
These are the people leading the work — and the ones you'll actually talk to. Behind them stand imec's top-notch engineering teams & research groups and imec's decades of experience in bringing the semicon industry together. Measuring the whole stack is not something anyone does alone, which is why we do it with a great team and great partners.
Shared allergies: guesswork
