[Gems-users] What are the reasonable values for on-chip, off-chip bandwidths?

["Lide Duan" <leaderduan@xxxxxxxxx>]

Hello there,

I ran SPLASH2 Raytrace with 2 configurations. The only difference is the off-chip bandwidth. Config1 has 64 Byte/cycle (bw_unit: 1000, bw_multiplier: 64) for both on-chip and off-chip links, while Config2 has 64 Byte/cycle for on-chip links but 8 Byte/cycle for off-chip links. Here are the details:

# processors: 16
Processors per chip: 8 (thus totally 2 chips)
# L2 banks: 32 (2 for each processor)
# memories: 16 (all are off-chip)
I use FILE_SPECIFIED topology and GARNET network. A 4x4 Torus network is used on each chip to connect the L1 and L2 caches. There is an additional off-chip switch connected to the two chips and the memories. As stated above, the only difference between the two configurations is that Config2 has much lower bandwidth on the links in/out the off-chip switch (i.e. the off-chip links). 

I expected that Config1 should perform better than Config2 given that it has higher off-chip bandwidths. However, for Raytrace (between two magic points), the simulation results are exactly the same. Ruby_cycles, Total_misses, CPI, etc are the same. Even the Average Link Utilization are both 0.0120178 flits/cycle. My initial thought is that the extremely low interconnect activities lead to the same performance because even 8 Byte/cycle is enough for off-chip communication. Does it make sense or do I miss something here?

So I am wondering what values are reasonable/suitable for on-chip, off-chip bandwidths? (Basically I want to observe some difference in situations like above)

Thanks,
Lide