Inter-cluster communication in VLIW architectures
ACM Transactions on Architecture and Code Optimization (TACO)
Proceedings of the 2008 ACM SIGPLAN-SIGBED conference on Languages, compilers, and tools for embedded systems
Proceedings of the 45th annual Design Automation Conference
SCRF: a hybrid register file architecture
PaCT'07 Proceedings of the 9th international conference on Parallel Computing Technologies
ACM Transactions on Embedded Computing Systems (TECS) - Special issue on application-specific processors
Hi-index | 0.00 |
Splitting a wide-issue VLIW processor in clusters decreases the clock period, area and power consumption. Previous studies of the physical benefits from clustering focused on the scalability of the register file, based on speculative analytical models. In contrast, we evaluate speed and area of the whole VLIW datapath, including the register files, FUs, and bypasses through realistic physical layout experiments. Our baseline is an optimized for speed 8-issue-slot VLIW pipeline, derived from a commercial media processor. Despite the frequent prior-art assumption that the register file defines the clock frequency of a clustered VLIW processor, we discovered it is the FU bypass network that limits the clock speed. After a drastic 1.75 clock frequency speedup from clustering the unicluster into two clusters, subsequent clustering brings modest 2.05 and 2.17 speedups for the 4- and 8-cluster VLIWs, respectively. Combined with cycle count increase trends due to clustering, this leads to the conclusion that excessive clustering does not speed up the processor. Furthermore, clustering reduces area of the VLIW datapath. In our experiments the area savings due to clustering reach 14.2%, 43.5% and 50.5% for the 2-, 4-, and 8-cluster VLIWs, respectively.