PsiQuantum Adapts Silicon Photonics to Address AI Networking Bottlenecks

PsiQuantum has unveiled an eight-port optical switch prototype that leverages quantum-grade silicon photonics to reduce signal loss in high-density AI data center interconnects.

Chips & Climate Desk
3 min read
PsiQuantum Adapts Silicon Photonics to Address AI Networking Bottlenecks

As the scale of artificial intelligence training clusters moves from thousands to hundreds of thousands of GPUs, the industry is hitting a wall with traditional electrical switching. PsiQuantum, primarily known for its pursuit of a million-qubit quantum computer, is now pivoting its core silicon photonics expertise to solve this classical networking crisis. The company has demonstrated an eight-port optical switch prototype that manages light with unprecedented precision, maintaining a signal loss of just 1.2 dB. This performance is critical because every decibel lost in the network fabric necessitates higher laser power or more frequent signal amplification, both of which drive up the thermal and power envelopes of the data center.

The architectural shift proposed here moves away from the conventional Re-timer and Optical-Electrical-Optical (OEO) conversion cycles that define modern networking. In a standard data center environment, signals are constantly flipped between light and electricity to be processed by CMOS-based switches. This process is not only power-intensive but introduces micro-latencies that aggregate across large-scale distributed training jobs. PsiQuantum’s approach keeps the signal in the optical domain, using silicon photonics to route data at the speed of light without the overhead of heavy electrical processing, effectively turning the network into a transparent medium for GPU-to-GPU communication.

Technically, the breakthrough lies in the manufacturing process and the integration of optical components directly onto silicon wafers. By using standard high-volume CMOS foundries to produce these photonic integrated circuits, PsiQuantum is betting that it can achieve the reliability and scale required for the next generation of AI infrastructure. The 1.2 dB loss figure is particularly notable because it represents less than a quarter of the light being lost during the switching process, whereas the industry generally tolerates up to 3 dB. This headroom allows for more complex topologies and longer cable runs, which are becoming a necessity as physical space constraints force clusters to span across multiple rows or even separate halls.

This development places PsiQuantum in direct competition with established networking giants and emerging optical interconnect startups. While Nvidia’s InfiniBand and Broadcom’s Ethernet solutions currently dominate the market, they are increasingly reliant on pluggable optics that are reaching their physical limits in terms of density and power per bit. By integrating the switching fabric into the optical layer itself, PsiQuantum is targeting the 'middle of the rack' bottleneck. If successful, this technology could render traditional top-of-rack switches obsolete for high-bandwidth GPU clusters, replacing them with passive or low-power optical routing that generates a fraction of the heat.

The broader implications for the industry involve a fundamental rethinking of the data center backplane. We are moving toward a 'disaggregated' future where compute, memory, and storage are not confined to a single chassis but are pooled across a high-speed optical fabric. PsiQuantum’s ability to maintain signal integrity across an 8x8 matrix is merely the first step; the true test will be scaling this to 64x64 or 128x128 configurations. At those densities, the manufacturing tolerances for silicon photonics become incredibly tight, and any misalignment in the waveguides can lead to catastrophic signal degradation. However, the company’s experience in quantum computing, where every photon is precious, gives them a unique technical advantage in this precision engineering.

Looking ahead, the industry should watch for how PsiQuantum integrates these switches with existing CXL and NVLink protocols. For silicon photonics to become the standard, it must play well with the logic layer of the GPUs it connects. We expect to see the next phase of development focus on 'co-packaged optics,' where these photonic switches are moved even closer to the processor, perhaps even onto the same substrate. This would effectively eliminate the need for traditional copper traces on the PCB for high-speed I/O, solving the signal integrity issues that currently plague PCIe 6.0 and 7.0 implementations. The race to replace electrons with photons is no longer just a research project; it is a commercial necessity for the survival of the AI scaling laws.

Sources

  1. 01 PsiQuantum turns its photonics know-how into an optical switch for supercharging AI data center GPU networks — Tom's Hardware
#silicon-photonics #ai infrastructure #psiquantum #networking #gpu-clusters