Cisco Systems, a prominent network supplier, has unveiled two new chips from its Silicon One series: Silicon One G200 and G202. These chips are designed for backbone switches utilized by large web scalers and are optimized for artificial intelligence (AI) applications, boasting low latencies and programmability. With a maximum throughput of 51.2 Tbit/s, Cisco is now on par with Broadcom’s Tomahawk 5 and NVIDIA’s Spectrum 4 chips, which already achieve this level of performance.
The Silicon One G200 is a 5 nanometer chip, delivering up to 51.2 Tbit/s based on 512 x 112 Gbit/s serializer/deserializer. This chip doubles the maximum throughput of its predecessor in the G100 series. On the other hand, the G202 achieves a throughput of 25.6 Tbps with 512 x 56 Gbps serializers/deserializers. Cisco envisions using these chips in clusters equipped with up to 32,000 GPUs, with a 400G connection, for AI applications. The manufacturer highlights the chips’ optimized energy efficiency when compared to competitors, although the significance of this feature remains uncertain for applications involving up to 32,000 GPUs. It is now a waiting game until the first switches based on these chips are released.
