Broadcom unveils Tomahawk Ultra in AI rivalry with NVIDIA
Broadcom just dropped the Tomahawk Ultra, a powerful Ethernet switch chip built for massive AI and high-performance computing setups.
Thanks to TSMC's 5nm tech, it can connect up to 512 GPUs at super-fast speeds—territory where NVIDIA's NVLink used to rule.
By using standard Ethernet parts instead of custom hardware, Broadcom is aiming for more open, affordable AI networks.
NVLink vs Tomahawk Ultra
Unlike NVIDIA's tightly integrated NVLink (which is basically plug-and-play with their GPUs), Broadcom's chip uses Ethernet—making it easier and cheaper to scale but needing more technical know-how to get top performance for AI.
The real win here is flexibility: you're not locked into one vendor, and you can use regular cables and transceivers.
Will this chip really replace NVLink?
If your team has the skills to fine-tune networking for AI workloads, Broadcom offers a cost-effective way to build huge clusters without being tied down by proprietary tech.
But if you want something that "just works" out of the box for AI, NVIDIA's all-in-one approach is still the go-to choice.