NEC Corp explains that each node in its new 1,024-way capable Cenju-4 parallel supercomputer (CI No 3,215) can be divided for shared memory (ccNUMA) or local memory (message-passing) at the same time. Users could also configure all physical memory (512Gb) as the shared memory or local memory. It says Cenju-4’s memory interface has been designed to show shared memory as a kind of shared set of devices such as files. It wants to take the model further. Although it says it won’t disclose actual numbers until some time in the future, NEC claims Cenju-4 has smaller latencies than conventional shared memory systems and believes that the communications latency of its previous Cenju-3 system is the same as IBM Corp’s current SP-2 parallel processor are almost the same, even though SP-2’s processor is five times faster than the previous generation Mips RISC used in Cenju-3. NEC says the total network throughput of Cenju-4 increases as the number of nodes increases. It performs at 200Mbps. Although Silicon Graphics Inc’s Origin 2000 system achieves 800Mbps, NEC claims SGI’s real throughput is at most 200Mbps per node (two processors). It needs 800Mbps because more than two nodes may share the same line in hyper-cube topology. รก
