The Power Of Optical Compute

The core of Lumai’s unique AI processor is an Optical Matrix Multiplier. We use the properties of beams of light moving in 3D space to perform the calculations. By using very wide vectors and high optical clock speeds, data is processed at least 50x faster and 90% more efficiently than today’s solutions.

Photons stream through advanced optical components in millions of parallel beams moving beyond silicon and electrons to deliver orders-of-magnitude performance gains, providing massive scalability for AI.

Hardware Precision

Proprietary data encoding/ decoding
algorithms ensures optimal precision

Large Matrix

Patented designs of precision optical systems enable unique large-scale computing in 3D
volume

High Speed Data Conversion

Extremely wide interface provides high data throughput and amortizes data conversion
overhead

Memory

Three-tier memory architecture enables efficient & high-capacity memory access for KV Cache

Quantization

Hardware-aware quantization schemes  for robust inference

The Lumai Iris System Architecture

Lumai Iris is a complete AI inference server designed for datacenter deployment. Lumai Iris uses a hybrid digital-
optical architecture. Optical handles the compute bottleneck, digital delivers on everything else.
Lumai Iris hybrid architecture: Optical handles the compute bottleneck, digital delivers on everything else

Scaling That Improves Efficiency

In conventional digital systems, performance gains often come with disproportionate power costs. Our architecture
flips this - efficiency increases as performance increases.  This increased efficiency is because performance is proportional to the vector size squared (whereas power is linear to the vector size). Larger systems are not just more powerful, they are inherently more efficient.

Optical Clock Speeds

Optical components can easily operate at higher-frequencies

Matrix Dimensions

Throughput grows quadratically with matrix size, while power scales linearly. Larger matrices mean higher efficiency

Component-Level Gains

Advances in optical and optoelectronic efficiency compound system-level performance over time

Scalable Memory

Memory scales independently from compute, bypassing the limits of on-chip digital memory

Scalable optical architecture for

the next 20 years of AI

Contact Us

Arrow