PORTLAND, Ore. — Intel revealed its roadmap for the future of technical computing, including customizing its high-end Xeon and Xeon Phi processors, at this week's Supercomputing Conference (SC13), Nov. 17-22 in Denver. The company promised to start housing memory chips inside the same package with its processors as well as integrating stacked memory dies onto future processors along with integrated high-speed switches and optical fabrics.
"We have the transistor budget to do customized innovation, and secondly we have a design methodology for system-on-chip and an architectural modularity that allows us the ability to work with our customers to customize products at various levels," said Rajeeb Hazra, Intel's vice president of the technical computing group and general manager of the datacenter group, in an interview with EE Times. "We are moving forward into workload-optimized architectures at a level of collaboration with our customers that we hadn't done previously."
At the last Intel Developer's Conference, the company discussed customizing its new Quark SoC, but SC13 is the first place that it revealed that it was working with technical computing customers to customize its high-end processors. Intel promised future Xeon and Xeon Phi models that integrate interconnection fabrics as well as stacked dies that add memory atop processors in 3D packages.
Intel has already integrated onto its CPUs math coprocessors, memory controllers, graphics, I/O controllers, and now in-package memory with Knights Landing. Next will be high-speed switches, optical fabrics, next-generation storage, and 3D stacked memory CPUs.
Intel will start by adding in-package memory dies alongside the processor, beginning with the next-generation Xeon Phi -- code-named Knights Landing. "I've talked about Knights Landing, and its in-package memory, and the train doesn't stop there. We are looking at various new classes of integrations, from integrating portions of the interconnect as well as next-generation storage and memory much more intimately onto the processor die," said Hazra.
Hazra said Intel would customize its in-package memory architectures for the customer's specific needs, with memory management units that enable customers to choose to implement caches, flat memory spaces, or hybrid combinations of the two.
"We have architected multiple memory usage models. So whether it’s a part of the flat memory space or it’s used in some form of a cache for applications that were not modified to make use of the new memory architecture -- we cater to all of those constraints and needs."
Intel also described an upcoming distribution of its Apache Hadoop -- the open-source software framework for processing large-scale unstructured big data on clusters -- that could access its Lustre parallel distributed file system in a manner transparent to application programs.
— R. Colin Johnson, Advanced Technology Editor, EE Times