DNA 26-10: AMD Pensando DPU Powered Smart Switches

This week may be a little hardware heavy before we get into the use cases and capabilities of DPUs in future posts. I’ll be focused on the AMD DPUs and more specifically, the DPUs which enable the Cisco Nexus (N9324C-SE1U and N9348Y2C6D-SE1U) and HPE CX10K (CX10000 and CX10040) smart switches. Cisco also ship the Cisco 8102-28FH-DPU-O, which typically runs SONiC (Software for Open Networking in the Cloud) and is aimed at hyperscaler environments, so it deserves a mention but is not the focus of today’s post.

AMD currently provide three DPUs to the market – Elba, Giglio, and Selina. Selina is the newest and has not yet made its way into HPE or Cisco smart switches. Giglio improves on the second-generation Elba hardware with additional power and performance efficiency. As a result, both DPUs provide line rates of 400G but Giglio will provide better performance for features such as Layer 3 to 4 filtering or deep packet inspection.

In both Cisco and HPE, the DPUs are integrated directly with the switching silicon, which is Silicon One E100 for Cisco and the Broadcom Trident3 for HPE. The E100 supports up to 6.4T backplane switching capacity, while the Trident3 supports 3.2T. Three of the switches (Nexus 9324C-SE1U, Nexus 9348Y2C6D-SE1U, and CX10000) advertise 800Gb of network services, while the CX10040 provides 1.6Tbps of network services. Although switching throughput isn’t matched by DPU capability, it is a very cost-effective alternative to expensive network services appliances, such as next generation firewalls.  

The switch and PDU specifications breakdown as follows.

- Cisco N9324C-SE1U has four Elba DPUs, with combined 128G of memory and an estimated 800G of network services throughput.

- Cisco N9348Y2C6D-SE1U has two Giglio DPUs, with combined memory of 128G and again an estimate 800G throughput for network services.

- HPE CX10000 has two Elba DPUs and 64G of memory and offers 800G throughput on network services.

- HPE CX10040 doubles up with four Elba DPUs and 128G of memory for 1.6T network services throughput.

If you’ve taken the time to read the previous bullets carefully, you will notice that HPE CX10000 and Cisco N9324C-SE1U both have the same throughput (800G) but the HPE switch only has two DPUs. This is primarily due to how the DPUs are integrated into the switching chip, with Cisco using four 200G internal links, while HPE uses two 400G. However, there may also be difference in how Cisco and HPE measure performance.

Finally, a thought on adoption. Cisco offer price parity between the N9324C-SE1U and N9348Y2C6D-SE1U switches and the equivalent non-DPU top of rack switches. This raises the question, if you can provide a port-based Layer 4 to 7 firewall into your Cisco switches at no extra cost, why do you need to invest in expensive solutions for East to West segmentation? More to come on that.