Cisco announced the 5672UP in January, 2014. It uses the same port ASIC as the Nexus 6001 introduced a year earlier. The port ASIC can support 3 x 40 Gb/s or 12 x 10 Gb/s ports. UP in the model name indicates that sixteen of the ports are Unified. This means they can run Ethernet, Fiber Channel, or Fiber Channel over Ethernet.

Nexus 5xxx switches are configured with an array of port ASICs connected by a cross-bar fabric. A major difference between the 55xx and 56xx is that in the former, buffer is dedicated per-port. The later has more buffer and it forms a shared pool for all the ports on the ASIC. A key point is in this slide:

Slide shows 14 MB out of 16 MB buffer is shared.

In big data wide area networking, burst absorption is more important than predictable performance. And the 5672 has an array of cool ways to find out when and where you are dropping frames and to see how full buffers are running. In common with many other Cisco data sheets, information about buffers is hard to find. It is exposed in deep dive sessions at Cisco-live, which is where I stole the slide. Note that buffer allocation is flexible. That means some configuration will be required to get it to do the right thing for SciDMZ applications. This is hardly unique.

Some of the input buffer packet memory is dedicated to each port. This is necssary to avoid head-of-line blocking. A small amount is used to hold packets sent to span ports. 14 MB is therefore the size of the shared pool. There is also 9 MB of output packet buffers. 3 MB of this is used for unicast packets. Under the most optimistic situation where only a single flow is claiming buffer space, 17 MB is available for burst absorption.

Both the Nexus 6001 and 5672UP have fine grained buffer monitoring built into their switch ASICs. You can read about it.