6904 buffers

A lot of detailed information here. 4 x 40G oversubscribes the switch fabric bandwidth by 2X. The line card can run in performance mode where two of the interfaces are shut matching the port and fabric bandwidths. Or it can run in oversubsribed mode where all four ports are active. Cisco has a nice white paper on this line card:

http://www.cisco.com/c/en/us/products/collateral/switches/catalyst-6500-series-switches/white_paper_c11-696669.html

Each 40G slot has four virtual ports (VP) which are each allocated their own 21 MB slice of the output buffer pie. Since all the packets in any single flow will hash to a single VP, the max buffer available to a single flow is 21 MB. Each 40G CFP can run in 4 x 10G mode in which case the output buffer per 10G port is [still] 21 MB. The input buffer is similarly divided into four virtual chunks. Snippet from the white paper:

• Port Group 1& 2 has combined Ingress packet buffer of 10 MB same as the Port Group 3 & 4
• Port Group 1 & 2 has combined Egress packet buffer of 176 MB same as Port Group 3 & 4
• Distribution of packet buffers among the ports in Performance or Oversubscribed mode is explained in more detail below
• Ingress buffer for Port Group (1&2) or Port Group (3&4):
• In Performance mode 1ms of storage = 10 MB or 10 MB/4 for each of the 4 x 10 Gigabit Ethernet
• In Oversubscribed mode = 10MB/2 for each of the 2 x 40GE ports and/or (10MB/4)/2) for each of the 8 x 10GE
• Egress buffer for Port Group (1&2) or Port Group (3&4):
• In Performance mode, 10 ms of storage equals 176 MB or 176 MB/4 for each of the 4 x 10 Gigabit Ethernet
• In Oversubscribed mode equals 176 MB/2 for each of 2 x 40 Gigabit Ethernet ports and/or (176/4)/2) for each of 8 x 10 Gigabit Ethernet