Thanks for providing all the codes of the hardware design which are very helpful. I have a small question about the configuration of the GIN workload used in Table V and the GCN workload in Table VIII.
The latency of GIN shown in Table V is 0.1799 ms, which is 0.1799*10^-3*330*10^6 = 53970 cycles using 1741 DSPs as shown in Table III. Table VIII shows that the latency for processing the GCN workload, such as Cora, is 6.912 us, which is 6.912 * 10^-6 * 330*10^-6 = 2281 cycles with 747 DSPs. Given that the size of HEP is much smaller than Cora's, and even accounting for different numbers of layers, the disparity in latency (approximately 50x difference, calculated as (53970*1741)/(2281*747) seems unusual. This suggests that FlowGNN processes heavier workloads (e.g., Cora) with significantly lower latency.
I guess the only reason is that I might make some mistakes in the configuration of GIN workloads, such as graph size, feature size, or layer number. Could you please provide the detailed configurations for the GIN and GCN workloads used in Tables V and VIII?
Thanks for providing all the codes of the hardware design which are very helpful. I have a small question about the configuration of the GIN workload used in Table V and the GCN workload in Table VIII.
The latency of GIN shown in Table V is 0.1799 ms, which is 0.1799*10^-3*330*10^6 = 53970 cycles using 1741 DSPs as shown in Table III. Table VIII shows that the latency for processing the GCN workload, such as Cora, is 6.912 us, which is 6.912 * 10^-6 * 330*10^-6 = 2281 cycles with 747 DSPs. Given that the size of HEP is much smaller than Cora's, and even accounting for different numbers of layers, the disparity in latency (approximately 50x difference, calculated as (53970*1741)/(2281*747) seems unusual. This suggests that FlowGNN processes heavier workloads (e.g., Cora) with significantly lower latency.
I guess the only reason is that I might make some mistakes in the configuration of GIN workloads, such as graph size, feature size, or layer number. Could you please provide the detailed configurations for the GIN and GCN workloads used in Tables V and VIII?
Thank you very much.