Prof. Radu Marculescu from Carnegie Mellon University in the USA gave an interesting keynote speech this morning on “Design and Optimization of Thousand-core Systems”.
He presented some of the existing work on intracommunication between a multi-core system. A 24-tile processor from Intel was highlighted and how the communication is performed inside there. Interestingly there are two options of communication and load sharing described: the fully-distributed approach and the fully-centralized. The former included a reference to a French paper where the cores are locally taking decisions on communication and load sharing and in the latter (Intel) the communcation and load sharing is done by a centralized agent.
Prof. Marculescu’s work described how an intermediate solution (use some long-distant links across the whole echilada) proved to be “good enough” in terms of efficiency. The Intel approach is in some sense optimum (as long as top-level agent has all the information it needs) and a minimum number buffers can be used and congestions can be avoided. With the French approach the congestions occur quite ‘early’ with short buffers. Using long-distant links the number of congestions could be significantly reduced and buffer length shortened.
Prof. Marculescu also highlighted some interesting ideas for future work: put the user into perspective. In order to utilize the results from above (long, intermediate links) one could tailor certain architectures to the user. Rather than going for a multi-purpose, fits everyone architecture, one could have a processor for different target groups. In my mind, you might already see some of this today, as different users will tend to buy different types of computers. But, that’s a bit different in some sense …
Off to Victoria Peak tonight.