Key takeaways
- Broadcom Tomahawk Ultra and NVIDIA's Spectrum-X SN5600 each offer 51.2 Tb/s of switching capacity, but that shared number hides a basic difference in purpose.
- Tomahawk Ultra is built for scale-up communication among tightly coupled accelerators. Broadcom specifies 250 ns switch latency at full throughput and hardware support for lossless transport and collective operations.
- The SN5600 is a 2U, 64-port 800 GbE switch system used inside Spectrum-X, an end-to-end platform that coordinates switches and SuperNICs for scale-out AI fabrics.
- This is not a clean chip-for-chip buying decision. Operators must compare complete switch systems, network adapters, software, topology and workload behavior.
- NVIDIA has moved Spectrum-X beyond the SN5600 generation. Its current Spectrum-6 systems reach 102.4 Tb/s per switch chip, so the 51.2 Tb/s comparison applies specifically to the Spectrum-4 SN5600.
The short answer
Broadcom Tomahawk Ultra and NVIDIA Spectrum-X solve adjacent networking problems.
Tomahawk Ultra brings Ethernet into the scale-up domain: the short, latency-sensitive links that let accelerators behave like one large computing system. Spectrum-X is an Ethernet platform for the scale-out fabric between GPU servers and racks. Its performance depends on coordination between the switch, the server-side SuperNIC and the platform software.
That distinction matters more than the identical 51.2 Tb/s headline. A 51.2 Tb/s Tomahawk Ultra-based system is optimized around switch transit latency and rack-scale accelerator traffic. A 51.2 Tb/s SN5600 provides port density for a leaf-spine network and participates in Spectrum-X congestion control and adaptive routing.
Specification snapshot
| Attribute | Broadcom Tomahawk Ultra | NVIDIA Spectrum-X SN5600 |
|---|---|---|
| Product scope | Ethernet switch silicon for AI and HPC scale-up | 2U Spectrum-4 Ethernet switch system within the Spectrum-X platform |
| Switching capacity | 51.2 Tb/s | 51.2 Tb/s |
| Published packet rate | Up to 77 billion packets/s at 64-byte line rate | 33.3 billion packets/s |
| 800 GbE-facing configuration | Depends on the system built around the chip | 64 OSFP ports at 800 GbE |
| Published switch latency | 250 ns at full throughput | Not listed in NVIDIA's public SN5600 product table |
| Network role | Tightly coupled XPU scale-up and HPC | AI back-end scale-out leaf/spine fabric |
| Host-side component | Endpoint-agnostic design | Spectrum-X SuperNIC, such as BlueField-3 for Hopper-era systems |
| Network operating system | Chosen by the system vendor | Cumulus Linux or Pure SONiC options |
The packet-rate figures are not an application benchmark. They describe different products under vendor-defined conditions and should not be used by themselves to predict training time.
What Tomahawk Ultra changes for scale-up Ethernet
Broadcom announced that Tomahawk Ultra was shipping on July 15, 2025. The company specifies 250 ns of switch latency while sustaining 51.2 Tb/s, plus line-rate forwarding of minimum-size 64-byte packets at up to 77 billion packets per second.
The latency target is the main story. Scale-up links carry synchronization and collective traffic among accelerators, so a small delay repeated across many communication steps can hold up the entire job. Broadcom says a Scale-Up Ethernet deployment using Tomahawk Ultra can keep XPU-to-XPU communication below 400 ns, including switch transit time.
Broadcom also changed how Ethernet behaves for this use case. The chip can reduce header overhead from 46 bytes to 10 bytes while remaining Ethernet-compliant. Link Layer Retry and credit-based flow control are intended to recover link errors and prevent buffer overflow without dropping packets. In-network collective support can execute operations such as AllReduce, Broadcast and AllGather inside the switch instead of sending all of that work back to the accelerators.
Those features make Tomahawk Ultra more than a faster conventional data-center switch. They are specifically aimed at the tightly synchronized traffic patterns that proprietary scale-up interconnects have traditionally handled.
Broadcom describes the collective functions as endpoint-agnostic and says Tomahawk Ultra is compliant with the Ultra Ethernet Consortium standard. The chip is also pin-compatible with Tomahawk 5, which can shorten the hardware redesign cycle for switch makers. Pin compatibility does not mean an operator can replace any installed switch in the field without qualification; it is a system-design advantage for vendors building products around the silicon.
What Spectrum-X and the SN5600 provide
Spectrum-X is broader than a switch ASIC. NVIDIA defines it as an end-to-end, standards-based Ethernet platform combining switches with SuperNICs. In the Hopper-era generation, the SN5600 switch works with BlueField-3 SuperNICs. Telemetry-based congestion control and adaptive routing coordinate the two ends of the network to spread traffic and isolate competing workloads.
The SN5600 itself is a 2U system with 64 OSFP ports running at 800 GbE. NVIDIA lists 51.2 Tb/s of maximum throughput and 33.3 billion packets per second. Those specifications make it a dense leaf or spine building block for connecting GPU servers across racks.
NVIDIA markets Spectrum-X for networks ranging from enterprise clusters to cloud AI factories with hundreds of thousands of GPU connections. Its Multiplane design can reach 128,000 GPUs in two tiers by dividing each GPU's SuperNIC traffic across independent planes. That scale claim describes a network architecture, not the number of GPUs attached to one SN5600.
NVIDIA also says Spectrum-X can improve AI network performance by 1.6 times compared with traditional Ethernet. That is a vendor platform claim, not a neutral head-to-head benchmark against Tomahawk Ultra. The result depends on the workload, topology, endpoints and comparison network.
Why “open versus closed” is the wrong shorthand
Both offerings use Ethernet, and both vendors emphasize standards. Broadcom says Tomahawk Ultra maintains Ethernet compliance and supports the UEC standard. NVIDIA describes Spectrum-X as standards-based Ethernet and offers open network-operating-system choices, including Pure SONiC.
There is still an integration difference. Broadcom sells switch silicon that system vendors can incorporate into their own platforms. NVIDIA tunes Spectrum-X as a coordinated stack spanning its switches, SuperNICs, telemetry and software. That can simplify validation for an NVIDIA-centered AI factory, while a merchant-silicon design may offer a different mix of system vendors and operational tooling.
Calling one option “open” and the other “closed” erases those practical choices. Buyers should instead ask which network operating systems are supported, whether switch and NIC combinations are validated, what telemetry is exposed, and how much freedom they retain to change individual layers later.
Do not compare Tomahawk Ultra with all of Spectrum-X
The SN5600 is the useful 51.2 Tb/s reference point, but it is no longer the top of NVIDIA's range. NVIDIA's current Spectrum-6 generation reaches 102.4 Tb/s per switch chip. The SN6600 exposes 128 ports of 800 GbE, while larger systems use multiple chips.
That product progression makes the scope of any comparison important. Tomahawk Ultra versus SN5600 is a comparison of a scale-up switch chip with a scale-out switch system at the same aggregate bandwidth. Tomahawk Ultra versus the entire Spectrum-X family would mix network roles and hardware generations.
Questions to ask before choosing a fabric
- Where is the bottleneck? Accelerator-to-accelerator communication inside a rack points toward scale-up latency. Congestion across many racks points toward scale-out routing and traffic control.
- What is the complete bill of materials? Compare shipping switch systems, NICs or SuperNICs, optics, cables, software licenses and support—not just ASIC throughput.
- Which claims have been tested on your workload? Vendor packet rates and platform uplift figures are not substitutes for a job-completion-time test using the intended model and topology.
- How will the fabric be operated? Network OS choice, telemetry, automation and failure recovery can matter as much as peak bandwidth.
- Which generation is being quoted? Confirm whether an NVIDIA proposal uses Spectrum-4 SN5600 hardware or the newer Spectrum-6 family.
Bottom line
The identical 51.2 Tb/s rating makes Tomahawk Ultra and the Spectrum-X SN5600 look like direct rivals. They are better understood as products optimized for different layers of an AI network.
Tomahawk Ultra focuses on low-latency Ethernet scale-up, with 250 ns switch latency, lossless link mechanisms and in-switch collectives. The SN5600 supplies 64 ports of 800 GbE for the Spectrum-X scale-out platform, where coordinated switches and SuperNICs manage traffic across large GPU fabrics.
For infrastructure planners, the useful decision is not which vendor owns the larger headline number. It is whether the immediate design problem is tightly coupled accelerator scale-up or multi-rack scale-out—and which complete, shipping system can meet that requirement under the operator's own workload.
Sources
This article was researched and fact-checked against the following sources: