Olix OTPU: How Photonic Inference is solving the AI Energy crisis

Light-Speed Tensor Operations

London-based Olix has secured $220M to bring its Optical Tensor Processing Unit (OTPU) to the global inference market...

The How: Architecture & Implementation:

Unlike traditional GPUs that use electrical logic gates, the OTPU uses photonic interferometers to perform matrix multiplications. * Zero-Heat Multiplication: Since photons don't generate resistance heat like electrons, the OTPU can perform trillions of operations at near-zero thermal output. * Analog-to-Digital Interconnects: High-speed DAC/ADC arrays that bridge the photonic core with traditional DDR5/HBM4 memory. * Linear Scaling: Bandwidth scales linearly with the number of wavelengths used, allowing for massive parallel processing on a single fiber core.

Performance Benchmarks:

• Efficiency: 10 Watts per PetaFLOP (compared to 300W for current top-tier GPUs).
• Throughput: Sustained 500 TOPS (Tera Operations Per Second) at the edge.
• Reliability: 50,000+ hour MTBF (Mean Time Between Failures) for the integrated laser sources.

Strategic Industry Impact:

Olix is targeting the Edge AI Inference market, where power and cooling are the primary constraints. This technology will enable GPT-4 level intelligence in hardware as small as a smart camera or a handheld medical scanner.