Marvell unveils 102.4 Tbps AI switch
New Teralynx T100 switch targets AI data centres with lower power, higher radix scaling and flexible co-packaged optics architectures.
Marvell Technology has announced its Teralynx T100, the industry’s first 102.4 Tbps switch silicon purpose-built for AI and cloud data centre infrastructure, as hyperscale networks continue to scale to meet increasing demand from AI workloads.
Built on a 3nm process, the T100 is designed to address rising power and latency challenges in AI clusters, which are driving rack power levels toward 120 kW and placing increasing pressure on data centre networking efficiency.
Marvell claims the device delivers up to 25% lower power consumption compared to competing solutions, while enabling flatter, higher-radix network topologies.
A key feature of the platform is support for advanced deployment models, including co-packaged optics (CPO) and co-packaged copper (CPC), reflecting a broader industry shift toward tighter integration between compute, switching silicon and optical interconnects.
The switch supports up to 512 ports and is designed for both scale-out and scale-up AI fabrics, allowing operators to reduce network tiers and simplify large-scale AI cluster architectures.
It also incorporates AI-optimised traffic management, integrated telemetry and low-latency design features aimed at improving GPU utilisation in training and inference workloads.
Marvell said the Teralynx T100 will begin sampling to customers this quarter, marking a further step in its portfolio of high-performance switching solutions for data centre infrastructure.
The announcement highlights the growing importance of co-packaged and system-level integration approaches as AI workloads continue to reshape data centre design and power efficiency requirements.
