Route selection is a major factor which affects the performance of the HBM subsystem. Consider the following point when designing the system:
- In any condition, PTP would offer the best performance per pseudo channel for any traffic pattern because no lateral bandwidth is used across the HNoC.
- The performance of the PTPL routing case would almost be identical to the performance of PTP, because again no lateral bandwidth is used across the HNoC.
- The PTPG routing case would see a performance drop compared to PTP or PTPL. This is due to the HNoC lateral bandwidth limitation. In a few cases, the PTPG performance might be commensurate with PTP when the traffic is bottlenecked by DRAM efficiency rather than the HNoC bandwidth.
- The UNR traffic case would see a performance drop compared to PTP, PTPL, and PTPG in most cases. In UNR, each master is transacting with multiple slaves. In certain traffic patterns, the performance is affected due to the variations of loads handled by a pseudo channel. The effectiveness of the HBM_NMU read reorder buffer is also reduced due to the incoming interleaved read responses from multiple destinations.