Broadcom Ships Tomahawk Ultra

[erek]

"Tomahawk Ultra is optimized for the tightly coupled, low-latency communication patterns found in both high-performance computing systems and AI clusters. With ultra-low latency switching and adaptable optimized Ethernet headers, it provides predictable, high-efficiency performance for large-scale simulations, scientific computing, and synchronized AI model training and inference.

When deployed with Scale-Up Ethernet (SUE specification available to the public here), Tomahawk Ultra enables sub-400ns XPU-to-XPU communication latency, including the switch transit time—setting a new benchmark for tightly synchronized AI compute at scale.

By reducing Ethernet header overhead from 46 bytes to just 10 bytes, while maintaining full Ethernet compliance, Tomahawk Ultra dramatically improves network efficiency. This optimized header is adaptable per application, offering both flexibility and performance gains across diverse HPC and AI workloads.

Tomahawk Ultra incorporates lossless fabric technology that eliminates packet drops during high-volume data transfer. Incorporating LLR, the switch detects link errors using Forward Error Correction and automatically retransmits packets, avoiding drops at the wire level. Simultaneously, CBFC prevents buffer overflows that traditionally caused packet loss. Together, these mechanisms create a truly lossless Ethernet fabric, delivering the level of reliability demanded by today's most data-intensive workloads.

Tomahawk Ultra also accelerates performance through In-Network Collectives solving one of the most persistent bottlenecks in AI and machine learning workloads. Rather than burdening XPUs with collective operations like AllReduce, Broadcast, or AllGather, Tomahawk Ultra executes these directly within the switch chip. This can reduce job completion time and improve utilization of expensive compute resources. Importantly, this capability is endpoint-agnostic, enabling immediate adoption across a wide range of system architectures and vendor ecosystems.

Designed with innovations in topology-aware routing to support advanced HPC topologies including Dragonfly, Mesh and Torus, Tomahawk Ultra is also compliant with the UEC standard and embraces the openness and rich ecosystem of Ethernet networking.

Introducing SUE-Lite
As part of Broadcom's Ethernet-forward strategy for AI scale-up, the company has introduced SUE-Lite—an optimized version of the SUE specification tailored for power and area-sensitive accelerator applications. SUE-Lite retains the key low-latency and lossless characteristics of full SUE, while further reducing the silicon footprint and power consumption of Ethernet interfaces on AI XPUs and CPUs.

This lightweight variant enables easier integration of standards-compliant Ethernet fabrics in AI platforms, promoting broader adoption of Ethernet as the interconnect of choice in scale-up architectures.

Platform for AI Scale-Up and HPC Scale-Out
Together with the 102.4 Tbps Tomahawk 6, Tomahawk Ultra forms the foundation of a unified Ethernet architecture: enabling scale-up Ethernet for AI, and scale-out Ethernet for HPC and distributed workloads.

Now Shipping
Tomahawk Ultra is 100% pin-compatible with Tomahawk 5, ensuring a very fast time-to-market. It is shipping now for deployment in rack-scale AI training clusters and supercomputing environments. To learn more about the Broadcom Tomahawk Ultra family click here."

Source: https://www.techpowerup.com/338944/...ultra-ethernet-switch-for-hpc-and-ai-scale-up

 

[Dan_D]

Great another feature on their shitty network controllers that's not going to work with Cisco or Brocade switches. God this company has been a pain in my ass for almost two and a half decades now.

 

[Balkroth]

Great another feature on their shitty network controllers that's not going to work with Cisco or Brocade switches. God this company has been a pain in my ass for almost two and a half decades now.

Huh ? I've never actually had a problem with them, and pretty much every white box vendor (and even Cisco, Nokia, etc) use their chips . I'd say probably 75% of the market is using them. These are their Router/Switch silicon, not really the NIC side of things.

 

[Dan_D]

Huh ? I've never actually had a problem with them, and pretty much every white box vendor (and even Cisco, Nokia, etc) use their chips . I'd say probably 75% of the market is using them. These are their Router/Switch silicon, not really the NIC side of things.

Yep. Google Catalyst switch and Broadcom issues. It's been going on for 25 years off and on with various models. I've had compatibility problems with Brocade switches too. I've seen SAN storage not play nice with them. I've seen all kinds of things. Being cheap and ubiquitous doesn't make them good.

They are the enterprise version of Realtek at best.

 

[Haswellbeast]

Great another feature on their shitty network controllers that's not going to work with Cisco or Brocade switches. God this company has been a pain in my ass for almost two and a half decades now.

Ugh, yeah, stupid proprietary bullshit Linux drivers, just another bump in the road when you install Linux on old macs, and I've seen that many old, low end laptops that didn't get Intel were stuck with broadcom

 

[Balkroth]

Yep. Google Catalyst switch and Broadcom issues. It's been going on for 25 years off and on with various models. I've had compatibility problems with Brocade switches too. I've seen SAN storage not play nice with them. I've seen all kinds of things. Being cheap and ubiquitous doesn't make them good.

They are the enterprise version of Realtek at best.

Umm, I'm not really sure who'd you use besides them?, I mean in your 800G/port routers and what not, they make very good Asics that every company uses because they really make the only good ones, until you get to the crazy high end proprietary stuff like Cisco Silicon one.

Maybe Marvel? But Broadcom is pretty much all for 100G above what not.

 

[sunruh]

what 36 bytes did they remove from the Ethernet header? (doesnt say in link i looked)
and
reducing those 36 bytes (on my 300 baud modem maybe) improved network efficiency by how much on a 100G+ network?

granted when you get into millions/billions of pkts yes...but on each?

and did you take those 36 bytes and increase the payload or just not fill the 1500/9000mtu?

 

[erek]

“The Massive Size of Broadcom Tomahawk 6 102.4T Switch Chips and Tomhawk Ultra Bonus​

It is always neat to see the new chips. Usually, these are buried under either large heatsinks or liquid coolers. Getting to see the Tomahawk packages pulled from a briefcase and placed onto a ledge in a Marriott during OCP Summit 2025 is always a great opportunity to bring our STH community these new chips. It also shows just how much goes into these different designs. Broadcom is on a mission to scale Ethernet everywhere, and these along with the Broadcom Thor Ultra 800GbE NICs are designed for next-gen high-performance compute clusters.“

https://www.servethehome.com/the-ma...-102-4t-switch-chips-and-tomhawk-ultra-bonus/

 

[DanNeely]

"It's a really big chip, and to make the point about how really big is is we're going to show a ton of photos with nothing providing any sense of scale of the sort needed to tell the difference between a part the size of a babies fingernail or a dinner plate."