Broadcom earlierAnnounceIt launched the Tomahawk 6, a switch processing chip designed for data centers, which is claimed to be able to process up to 102.4Tbps of data exchange volume on a single chip, thereby meeting the interconnection needs of ultra-large-scale GPU cluster computing.
單組Tomahawk 6約可驅動多達10萬組GPU協同運算，其峰值可達102.4Tbps的資料交換量，約等同每秒可處理2.5萬部4K畫質電影內容，相比前一代Tomahawk 5可處理資料吞吐量約提升6倍，更比傳統乙太網路交換系統對應資料吞吐量提升6倍性能表現。
To increase data throughput, Broadcom uses a 6G/100G SerDes interface design in the Tomahawk 200, and a CPO co-packaged optical design to meet the current large-scale data processing computing needs of artificial intelligence and other applications.
Due to the increasing demand for collaborative computing across large numbers of GPUs in data centers, traditional Ethernet-connected architectures often result in GPU cluster utilization rates below 40%. The Tomahawk 6's superior data throughput performance will boost GPU cluster computing efficiency to 90%, enabling the training and inference of extremely large AI models with over a terabyte of parameters.
In addition, even though Broadcom emphasizes that it has kept the price of a single Tomahawk 6 unit below US$2 through a technology premium strategy, the manufacturing cost is still more than double that of the previous generation product. However, based on actual benefit estimates, it is expected that the overall AI training cost can be reduced by 30%-40%.
Currently, Broadcom's major partners include Google, OpenAI, and Meta. It also expects to launch an upgraded version of the Tomahawk 3 switch processing chip built on a 6nm process in the second half of this year.
