LFW: What advantages does optical compute offer that currently available compute infrastructure doesn’t?
Steinman: Optical compute is well suited to accelerate mathematical operations central to AI and scientific workloads, particularly linear operations such as vector-matrix multiplication. Our PACE 2 platform, for example, uses a 128 × 128 configurable optical matrix multiplier and a heterogeneous optoelectronic architecture to perform matrix operations with high performance, programmability, and versatility. Because the photonic compute engine can execute these operations without conventional pipeline stages, it can deliver extremely low latency and power-efficient performance. We offer the platform as a PCIe add-in card accelerator with support for developer tools including ONNX, OpenCL, PyTorch, and TVM to make it easier to explore real-world workloads on optical hardware.
LFW: Which markets are best served with optical compute technology?
Steinman: The most immediate opportunities are markets where workloads require high throughput, low latency, and strong energy efficiency. These include AI data centers, high-performance computing, scientific computing, financial engineering, medical imaging, and other compute-intensive industrial applications. At the data center level, hyperscalers are particularly well positioned to benefit from optical interconnect and CXL-enabled resource pooling because they operate at a scale where memory, compute utilization, power, and infrastructure efficiency have large economic impact.
LFW: Why is photonics attractive for chip interconnects?
Steinman: Photonics is attractive for chip interconnects because it can deliver high bandwidth, low latency, extended reach, and lower power over distances where copper becomes increasingly constrained. Optical fabric technologies such as ours are designed to support native compute protocols for composability and efficient resource sharing to help modern systems scale vertically and horizontally across nodes. Our optical communications hardware is compatible with PCIe 5.0/6.0 and CXL 2.0/3.0, which enables disaggregated AI memory and resource scaling within or across server racks.
LFW: Any challenges encountered while designing your photonics portfolio?
Steinman: Building production-ready photonic systems requires innovation across the full stack: Photonic devices, complementary metal-oxide semiconductor (CMOS)-compatible silicon photonics, electronic-photonic codesign, advanced packaging, software, reliability, and manufacturability. On the compute side, one challenge is continuing to identify and optimize algorithms and workloads that can fully exploit optical acceleration. On the interconnect side, the industry must balance the advantages of optics against the cost, reliability, and power perceptions of incumbent copper solutions. This is why Lightelligence has taken a hardware/software codesign approach and developed multiple deployment paths, including pluggable optics, near-package optics, and longer-term copackaged optics.
LFW: What’s next for Lightelligence?
Steinman: Lightelligence is focused on scaling optical computing and optical interconnect from advanced technology into deployable infrastructure. We’re continuing to expand PACE 2 as a developer-friendly optical compute platform, advance Photowave for PCIe/CXL-based disaggregated memory, and develop LightSphere X and distributed optical circuit switching for more flexible GPU superclusters. More broadly, we’re working with customers and manufacturing partners to bring larger photonic systems into data centers and continuing to develop the technologies to support the transition to near-package and copackaged optical architectures.
More information can be found on our website and follow us on LinkedIn for updates.