Machine for Parallel Graph Reduction [George89] is another shared memory implementation based on the sequential G-machine. A prototype implementation has been constructed for the BBN Butterfly multiprocessor, which consists of a number (15) of processing elements (MC68020 + 4Mbyte memory) interconnected through a delta network. Each processor can access transparently all memory in the whole machine, but local references are at least five times faster than references to remote memory. Therefore locality considerations have been taken into account in the AMPGR design. The AMPGR machine, unlike the !;G?-machine, is driven by compiler derived parallelism based on strictness analysis. The compiler inserts explicit spark instructions in the G-code whenever it can derive that a particular expression is needed; corresponding wait instructions are generated prior to the actual usage of the expression. To overcome the synchronisation overheads associated with large numbers of fine-grain tasks, m...