Lattice-native silicon for sub-watt AI inference

Most of the arithmetic in a quantized network is really a nearest-neighbour decision. We project each activation block onto the 240 roots of the E8 lattice and resolve it by comparison, so the datapath does shift-and-add work instead of dense multiplication.

Layers that resist quantization fall back to a standard INT8 path on the same silicon, chosen per layer at compile time — no accuracy cliff and no second chip.

Why the E8 lattice

E8 is the densest sphere packing in eight dimensions. For a fixed number of representable points, no other arrangement keeps quantization error smaller — and its symmetry makes nearest-point search a bounded sequence of sign flips and shifts.

Structure only — fill each cell with your own audited or internally prepared figures before this page goes live. Nothing here is a projection we produced.

Arctic Flame was formed in 2024 around one question. When a quantized neural network runs, most of its work is deciding which of a small set of values each number is closest to — a comparison, not a multiplication. So why do accelerators still spend the energy to multiply? The company is founder-led and operates remotely in the United States, with FPGA bring-up in a shared hardware lab and all customer infrastructure on AWS in us-east-1.

Talk to us

Tell us what you are running and on what power budget. We answer evaluation requests within two business days.