64B per cycle at 200MHz is 95.37 Gb/s. Unless there is other actively not due to the Homa core (?), then we will not fully saturate the link.
There is also dead bytes in the transaction between the Home core and the ethernet subsystem. This is because the AXI Stream interface is quantized in 64B chunks. These chunks are passed to the ethernet subsystem until a full packet has been sent, at which point the TLAST bit is toggled to tell the receiving core that it has an entire packet that can be sent. The problem is that the last 64B chunk is not necessarily completely full. We can specify to the ethernet subsystem via TKEEP bits to indicate how many bytes it should consider from the last 64B chunk, but this may require a higher clock frequency to saturate the link.
64B per cycle at 200MHz is 95.37 Gb/s. Unless there is other actively not due to the Homa core (?), then we will not fully saturate the link.
There is also dead bytes in the transaction between the Home core and the ethernet subsystem. This is because the AXI Stream interface is quantized in 64B chunks. These chunks are passed to the ethernet subsystem until a full packet has been sent, at which point the TLAST bit is toggled to tell the receiving core that it has an entire packet that can be sent. The problem is that the last 64B chunk is not necessarily completely full. We can specify to the ethernet subsystem via TKEEP bits to indicate how many bytes it should consider from the last 64B chunk, but this may require a higher clock frequency to saturate the link.