Add synchronization between cores forthat memory access, and the problemquickly becomes untenable. Itstates that the memory bottleneck is afundamental problem that no one hassolved and that Patterson’s Power Wall,which describes the trade-offs betweenmemory-bus bandwidth, power, andparallelism, still applies, at least withrespect to current multicore-processingarchitectures ( Reference 2). In that e-mail,Fish referred to a paper he co-authored,on behalf of Venray ( Reference 1). Thesubject line was, “ARM X-Gene missedthe Sandia Labs multicore report”-an irresistible headline.
In response to my blog about theApplied Micro X-Gene announcement,I received an e-mail from Russell Fish,co-inventor of the Sh-Boom processorand now with Venray Technology. Still,the tools and scalability are lacking.Ī more fundamental problem exists,though. Past a certain point, the benefits ofmulticore do not necessarily scale linearlyaccording to the number of cores.That we know, and many programmersand developers have spent many daysand nights working on the tools andalgorithms to make maximum use ofthe parallelism potential that multicorearchitectures theoretically provide. Those promises sound good, butthe hardware part bears more scrutiny,especially in the context of unstructureddata. Applied Micro is toutingit as the first “server on a chip” andpromises it will reduce the total cost ofownership of a server farm by as muchas 30% through energy savings alone. Operating at frequencies as high as 3GHz and touting as many as 128 quad-issue,out-of-order cores, the X-Geneconnects through a coherent terabitfabric and a memory throughput of 80Gbytes/sec. I’m justpicking on X-Gene because it’s one ofthe most recent. So why am I bringing this up now?Last month, at ARM TechCon, ARMannounced the ARM Version 8 ISA(instruction-set architecture) hourslater, Applied Micro unveiled X-Gene,the first processor employing the architecture.Oddly, parallels may existbetween those 14-bit Korean radiosand X-Gene-and all other multicoreprocessors, for that matter. The original four-transistor designgoes back to the pioneer Regency TR-1, which Texas Instruments andIndustrial Development Engineering Associates designed. This is a test do not adjust your age: Do you recall thetransistor-radio wars of the 1960s? Japan then dominatedthe maturing transistor-radio market, and, in an effort to differentiatethemselves, Korean manufacturers started addingfunctionless transistors to their radios just so that they couldsay they had eight, 10, 14, and, eventually, 16 transistorsinstead of the necessary four to eight.
0 kommentar(er)
