The theoretical max numbers in your case will be limited with following:
1. PCIe limit: PCIe 3.0 is 8 GT/s per lane but effectively it delivers about 1.54% less data due to 128/130 encoding, so PCIe limit will be around 63Gb/s for 8 lanes
2. 40GE channel encoding – it uses 64/66 encoding, so about 3% of 40GE is lost here.
3. Ethernet per frame overhead for frame preamble (8 bytes) and interpacket gap (12 bytes but can be as small as 5 or even 1 byte on 10GE and 40GE networks)
4. And lastly need to take into account how iperf calculates BW: I believe it calculates only payload transmitted, i.e. it doesn’t account for TCP or UDP headers, and neither for IP and Ethernet/MAC headers and checksum.
NOTE: the overhead in 3 and 4 is per packet, so with bigger packets you will have smaller impact of these on BW numbers.