Server Roadmap – Server Design Summit
November 2012 – At the 2012 Server Design Summit, they held a panel on future directions for Servers amid the current industry discussion about power efficiency, throughput and performance. The panel was moderated by Sergis Mushell of Gartner and featured Ed Gasiorowski of Applied Micro, Jeff Underhill of ARM, David Borland of Calxeda, Devid Kassel of Flextronics, Jim Hughes of Huawei Tech and Stefanie Lemon of Intel.
The discussion was lead off by Intel which indicated its server platform is arranged by workload. Using a 3D chart with the axis of scaling of the # of cores, scaling of the physical nodes required by the server and the scaling of the “brawniness” or throughput of the cores. These server chips scale all the way from the Xeon PHI chips (an Itanium based chip) for HPC and HA applications to the Xeon E7, then Xeon E5, then Xeon E3 and finally Atoms.
The server space is defined at Intel as Server Cores/ Network / Storage / and Software Tech. These are addressed by their Xeon Cores (in Westmere to Sandybridge to IvyBridge to Haslett architecture), SSDs, power management software and network security with the McAfee products. These tools help manage the power/performance metrics for the servers. In the majority of the architectures that are NOT HPC compute applications, the core processor power is a very small portion of the total power. It is dominated by memory, storage, and network switching.
The next few speakers – ARM, Calxeda and Applied Micro all discussed the ARM Cortex A series processors. Currently Calxeda is the only major company shipping and ARM based server in 2012. Their single chip SOC uses a 32 bit A9 core and just needs DRAM to run Unix or Linux. The products are targeted at Web server, data storage, and edge of network analytics The key is not ARM core, but the low power switch fabric that includes the L2 cache. Their next generation server will support virtualization and then move to the ARM 64 platform.
Applied Micro is planing to introduce an ARM 64 based server in the first half of 2013. It will be initially targeted at video streaming nd Hadoop applications. To support the Integrated CPU with the communication I/O and integrated NIC at 10G/40G/100G they have created a switch fabric called Xgene. The key is to support lightweight A9 type 32b cores on the same bus and memory map as the big processing and power of the A15 64bit core. This will allow them theo address multiple product on the Xplatform line for Xmem, Xstorage and Xcompute applications.
Flextronics brought a unique perspective as thier are a contract manufacturer of servers rather than a primary supplier. Their input at the panel was manufacturing, DFM, parts availability and tradeoffs on processor, memory, storage, networking, poer supply and mechanicals. Their expertise from working with multiple customers allows them to help a client achieve their optimized performance/watt spec.
Huawei brought a new perspective that the server world was just rehashing the old CISC vs RISC architecture battle on the x86 vs Cortex processor debate. The bigger challenge is the return of server designs that have marketing margin. The cloud and scale out model is pretty low margin. The cloud model has ll but killed the high margin HA and HPC markets. The new servers need to address the technology area of persistent memory in the machines. This is the next breakthrough for power reduction n the servers, but like the current x86 vs Cortex A series battle it will be largely application driven on the method implementation.
