Broadcom Tomahawk Ultra at Hot Chips 2025

A few weeks ago, we covered the Broadcom Tomahawk Ultra Launch for Scale-up Ethernet. At Hot Chips 2025, we get more.

Again, these are being done live during keynotes, so please excuse typos.

Broadcom Tomahawk Ultra at Hot Chips 2025

Here is a quick look at the HPC and scale-up AI fabric demands:

Broadcom says that Ethernet has often been seen as not suitable for these workloads, but it looks to change that by introducing Tomahawk Ultra. The 512 x 100G-PAM4 means that this can even handle 512 ports of 100GbE.

This is a new 51.2T switch designs for high-performance with in-network computing, low consisten latency, and high performance. The 51.2T at 64B is around 77B packets per second.

Here is a quick look at the packet forwarding piplien of Tomahawk Ultra.

Here are the key features again.

We covered this in our previous piece, but there is a link layer retry feature that happens at a low level on the switch.

Here is a bit more about how this works in detail.

There is also a credit based flow control features to keep buffers safe, here is a bit on that:

The header is also an optimized AI Fabric Header (AFH) which has a minimal set of fields in order to help the header to payload ratio.

Here is a bit on how this works in Scale-up Ethernet.

In many ways, like NVIDIA, there is in network computing support for collective operations.

Topology aware adaptive routing is important to keep networks running.

Here is a slide with a wall of words around programmable visibility on the switch and the observability.

Congestion control ensures certain links to not get overloaded.

Here is the line rate 64B packets across all ports. This is the big feature of this switch, it is not just a big packet throughput machine.

Here is the low latency side at 250ms.

We are going to be moving to a much higher-end network testing capability at STH with this big Keysight CyPerf box that is now in the lab.

Here is the switch lineup again with Tomahawk 6 as the big 102.4T throughput ASIC.

Final Words

I am not sure how much is new here versus our previous coverage, but this is a very cool chip. Broadcom is pushing the Scale up Ethernet project hard so we will see how it gets deployed. Or perhaps it is best said, we will see when we find it in one of our HPC or AI data center tours that we do.