Compute / memory / movement

Make room
for the
next compute.

AI systems are built around the work they need to do. ChipQuay explores the architecture between a model and the machine.

Independent accelerator concept. No fabricated chip, benchmark or production claim.

Macro photograph of an integrated circuit and surrounding board traces
REFERENCE IMAGE / Integrated circuit detail

Architecture is
the whole system.

A proposed floorplan is a way to ask better questions about compute, memory and the work moving between them.

CQ / 01CQ / 02CQ / 03CQ / 04CQ / 05CQ / 06CQ / 07CQ / 08CQ / 09CQ / 10CQ / 11CQ / 12CQ / 13CQ / 14CQ / 15CQ / 16
Design study / 01

Follow the data.

Sixteen illustrative tiles. A shared interconnect. A memory boundary. Select a layer in the explorer to understand its role.

Open the layer explorer ↗

Concept illustration, not a tape-out or a specification.

Engineering perspective

No single metric
tells the story.

01

Compute with a purpose

Explore a proposed accelerator building block through workloads, data movement and scheduling.

02

Bring memory closer

Memory placement is an architecture question. Examine the tradeoffs before quoting throughput.

03

Make the system legible

A developer view should explain how work reaches the hardware and where a bottleneck begins.

Explore a different perspective

Get closer to the architecture.