AX-1 · Neural ProcessorAxiom Compute

Intelligence,
at the edge.

Axiom builds neural processors that run real AI on-device — fast, private, and efficient enough to disappear.

AX-1 · Die 182mm²
AX-1
1,024TOPS
4 nmProcess
12 WEnvelope
0.98Sparsity
Datasheet

Every number, on the record.

Process node4 nm FinFET
Die area182 mm²
Transistors41 billion
Memory bandwidth1.2 TB/s
On-package memory96 GB SRAM
Supported precisionsFP16 · BF16 · INT8 · INT4
Operating range−40 °C to 85 °C
InterfacePCIe Gen5 · LPDDR5X
Performance

Measured, not marketed.

Transformer inference8.4×
Vision model throughput6.2×
Speech-to-text latency4.7×
Energy per inference−81%
Architecture

Around the model, not the board.

01 / SENSE

Ingest

Sensor fusion and tokenization at the edge of the die.

02 / COMPUTE

Matrix engine

1,024 cores, in-memory compute, sparse routing.

03 / OUTPUT

Decode

On-chip decode with a hardware root of trust.

Built for

The machines that need to think.

01Autonomous machines
02Wearables
03Robotics
04Smart cameras
Engineering samples · Q1 2026

Build on the edge.

AX-1 engineering samples are available to select partners now.

Request samples →
partners@axiom.ai