Optical SRAM and the Photonic Latch
Photonic chips compute at the speed of light but lose most of that speed on electrical↔optical conversions at memory boundaries. Optical SRAM (the photonic latch) — on-chip memory that stores data in the optical domain via resonant cavities or microscopic light loops — is the research frontier that would let photonic chips bypass the memory wall. Multiple university labs working on it; density still far below electrical SRAM.
**Optical SRAM** — also called **photonic latch** or **on-chip optical memory** — is the research direction aimed at letting Q.ANT Photonic AI Processor (NPU 2, 2026) keep their data in the optical domain across many compute steps, bypassing the electrical↔laser conversions that dominate their latency and power today. ## Why it matters Current photonic chips compute at the speed of light but their inputs live in conventional electrical memory (DRAM, SRAM, HBM). Every memory access requires an electrical-to-optical conversion and back. These conversions: - Consume significant energy (defeats the main advantage of photonics). - Introduce latency (defeats the compute-speed advantage). - Require high-quality photodetectors and modulators (hardware cost). See Memory Wall for the broader context. Photonic computing without optical memory moves the bottleneck but does not remove it. Optical SRAM is the prospect of compute + memory both in the optical domain. ## Approaches ### Resonant cavities Light bounces inside a microscopic optical cavity (ring resonator, photonic crystal cavity). The cavity's resonant wavelengths encode stored bits. Advantages: compact, can interface cleanly with waveguide-based compute. Disadvantages: storage is lossy (photons leak out over time, unlike electrons in SRAM), capacity per cavity is limited. ### Microscopic fiber loops / delay lines Light circulates continuously through a closed loop. State is held as the propagating waveform. Advantages: well-understood physics. Disadvantages: physically large for meaningful storage duration (light travels 30 cm in 1 ns), loop losses accumulate. ### Phase-change materials GST (Ge-Sb-Te) and similar materials switch between amorphous and crystalline phases with different optical properties; states are non-volatile. Advantages: true non-volatile memory. Disadvantages: slow write, limited endurance (similar to NAND flash), read currently requires active probing. ### Optical bistability Nonlinear optical elements (Kerr cavities, semiconductor optical amplifiers) can be bistable — two stable states set by input history. Advantages: fast. Disadvantages: requires optical power continuously to maintain state (volatile, not truly static). ### Spin-photon hybrids Light encodes state into atomic or solid-state spin systems (NV centers, quantum dots) which retain information at room temperature for microseconds to seconds. Advantages: long storage times. Disadvantages: not yet at useful densities. ## Fundamental challenges 1. **Photons don't stop easily.** They have to keep moving or be absorbed. Storage requires recirculation or non-optical intermediate states — neither is as clean as electron trapping in electrical SRAM. 2. **Density.** Electrical SRAM achieves tens of megabytes per mm². Best optical approaches are orders of magnitude behind. 3. **Coupling efficiency.** Getting light into and out of storage elements cleanly, without insertion loss dominating, is hard. 4. **Thermal management.** Optical elements are temperature-sensitive (cavity resonances shift with temperature). 5. **Integration.** Building both optical compute and optical memory on the same die at useful densities requires processes that don't exist at production scale yet. ## Who's working on it - **MIT Photonics Group** - **University of Southampton** (UK) - **TU/Eindhoven** (Netherlands, major photonic packaging center) - **Stanford** - **IBM Research** - Multiple DARPA-funded efforts under the **Electronic Resurgence Initiative** family - Commercial labs at Lightmatter, Q.ANT, Celestial AI ## If it works If useful-density optical SRAM scales, GPU monopoly becomes vulnerable for specific workload classes — specifically those where weights + activations can stay resident on-chip for many inference steps (transformer inference with KV-cache reuse, certain vision models). Training is still memory-hungry enough that it would probably stay electrical. If it doesn't scale, Q.ANT Photonic AI Processor (NPU 2, 2026) remain specialized coprocessors for workloads with high compute-to-memory-access ratios — a real but bounded niche. ## Status as of 2026 No commercial optical SRAM at deployable density. Multiple lab demonstrations showing individual components work. A coherent end-to-end system with optical compute + optical memory at competitive density is **the multi-billion-dollar open question** in photonic computing.