We write about silicon photonics, optical matrix multiplication, neural-network inference hardware, and the engineering problems that sit between the physics literature and production silicon.
How the GSK-1 SDK intercepts PyTorch Linear layers and ONNX Gemm nodes at runtime, what the PCIe DMA transfer overhead looks like in practice, and where the integration boundary sits in a multi-GPU inference server.
Read articleWhy laser coherence length sets a practical upper bound on waveguide mesh size, how the GSK-1 laser source is specified to maintain phase stability across a 512-element mesh, and what happens to matrix fidelity when coherence length is insufficient.
Read articleA tour of the memory bandwidth bottleneck in transformer inference, why weight fetches dominate energy consumption in autoregressive generation, and why the GSK-1 weight bank eliminates the repeated DRAM access cost entirely for static weight layers.
Read articleWhy TOPS and TOPS/W in isolation are insufficient metrics for evaluating photonic co-processors, and the case for joules-per-token, tokens-per-rack-unit, and tokens-per-dollar as the metrics that matter for production inference fleet operators.
Read articleHow thermal gradients across a silicon photonic die shift resonant wavelengths and phase offset, the thermoelectric stabilization approach used in the GSK-1, and the measurement protocol for verifying thermal stability under sustained inference load.
Read articleThe published literature on photonic neural networks demonstrates the physics clearly, but lab demonstrations differ from production hardware in packaging, thermal management, and host interface integration. This article describes the specific engineering problems Great Sky is solving in that gap.
Read articleHow WDM allows multiple matrix-vector products to propagate through the same physical waveguide mesh on different optical carrier wavelengths, the practical limits set by crosstalk and wavelength selectivity, and how this maps to batched inference workloads.
Read articleWhy a photonic hardware startup chose Boulder over the Bay Area or Boston, how proximity to NIST and CU Boulder affects the day-to-day pace of device development, and what the Colorado photonics ecosystem looks like in practice for a small team.
Read articleA systematic breakdown of optical insertion loss sources in a photonic inference chip, from fiber-to-chip coupling to waveguide propagation loss to modulator absorption, and the design choices the GSK-1 makes to keep the total loss budget compatible with a practical laser power level.
Read articleInP and GaAs offer higher electro-optic coefficients than silicon, but silicon photonics offers CMOS foundry compatibility, lower per-wafer cost, and established packaging ecosystems. This article compares the two material platforms specifically for inference co-processor applications.
Read articleAn accessible introduction to how coherent optical fields perform matrix-vector products in silicon waveguides: the Mach-Zehnder interferometer, unitary matrix decomposition, and the relationship between phase angles and matrix weights. No prior photonics knowledge required.
Read article
