Broadcom Unveils Tomahawk 6: First 102.4Tbps Superchip, Capable of Driving 100,000 GPUs

In a move poised to reshape the landscape of AI infrastructure, Broadcom has unveiled its latest data center switch chip, the Tomahawk 6. This next-generation silicon boasts an unprecedented switching capacity of 102.4Tbps, making it the world’s first single-chip solution to surpass this threshold.

The Tomahawk 6 marks a significant leap forward, effectively doubling the performance of existing Ethernet switches. Designed with the demands of the AI era in mind, the chip is engineered to support the intricate interconnectivity required by massive GPU clusters. In a testament to its capabilities, a single Tomahawk 6 is capable of orchestrating the operations of up to 100,000 GPUs simultaneously – a feat that underscores the chip’s computational prowess.

The sheer bandwidth of the Tomahawk 6 is staggering. Its theoretical peak of 102Tbps translates to the processing power equivalent of over 25,000 4K movies per second. This is a remarkable six-fold increase in throughput compared to the previous Tomahawk 5 generation.

The architecture of the Tomahawk 6 is flexible, incorporating advanced features like 100G/200G SerDes interfaces and Co-Packaged Optics (CPO) support. This flexibility allows for customization, meeting the specific demands of hyperscale AI network operators.

According to Ram Velaga, Senior Vice President of Core Switching Business at Broadcom, a critical bottleneck in current AI computation stems from network inefficiency. The reliance on GPU clusters for AI necessitates high-speed data transfer, yet network limitations often result in utilization rates below 40% for these powerful processors.

By effectively dismantling data transfer bottlenecks, the Tomahawk 6 promises to revolutionize GPU cluster utilization. The chip is designed to boost the actual computational efficiency of GPU clusters to over 90%, providing a vital infrastructure foundation for training and inference of large-scale, trillion-parameter models.

The launch of the Tomahawk 6 promises more than just increased computational efficiency; it also has the potential to bring down the cost of AI training. While the manufacturing costs have doubled compared to the previous generation, Broadcom intends to maintain the pricing below $20,000 per chip, with volume discounts available.

Industry analysts are anticipating Tomahawk 6 could drive a reduction in AI training expenses by 30-40%. Furthermore, Broadcom has plans to introduce an upgraded Tomahawk 6, incorporating 3nm technology, in the second half of 2025, with a focus on further enhancing energy efficiency.