Users and Vendors Consider Server Processors

October 29, 2012, AMD ??, San Francisco—A panel looked at the new issues associated with the latest AMD announcement to start a new ARM-based server family. Nathan Brookwood moderated the panel, and panel members were: Tim Burke from Red Hat, Jay Parikh from Facebook, Jimmy Pila from Dell, Simon Seegars from ARM, and Lisa Su from AMD.

Amazon videos, cell phone technologies all need more servers. Is the role of micro- servers overblown?

Seegers suggested a change in server design to go towards more small CPUs and I/O. These components can be integrated into an SoC to get much more power efficiency, potentially in the 10’s of percent ranges.

Doesn’t the role of CPUs relative to power get dwarfed by the DRAM, network, and storage?
Su contemplated changing from the traditional view of the data center where the workload was relatively homogeneous. Now the workloads have changed and CAPX and OPX are the most important. These changes will need a reduction of networks and their topologies to reduce the number of nodes from the system to the network. The changed economics are the drivers.

Data centers are built as plug-in scalable. Could moving to a small CPU server could change the economics?
Pila responded that the power is the biggest cost factor. The CPU is the single largest power consumer. The data growth rate means that the data centers need to reduce power and increase compute power watt.

Software conversion from x86 to ARM, is it worth it?
Parikh offered that power, cost, reliability and modularity are not the same as an enterprise and high performance compute configuration. As people move to mega-data centers, the big workloads will still take hundreds to thousands of servers to write out a web page. Modularity will occur as users change out hardware and move to the right scale and hardware. Flexibility needs modularity.

Some tasks need a big CPU, others only need a light processor. What is a good ratio?
Parikh responded that these are orthogonal issues. As people change from a powerful CPU and clock down, the latency will increase.

Software time to transition from x86 to ARM, or GPU?
Burke considered that this change is moving more quickly than originally thought. The drivers are bottom-up from the users. The enterprises are developing 64-bit components for cloud workloads. This conversion is very fast paced, as open source and Java runtime components play a big part in the developments. The open source community is agile and fuels change and innovation. There is a lot of open source code for 32-bit ARM, and the 64-bit code will follow.

How to convert blocks of code for different instruction sets?
Parikh stated that developers just need to put in a team and build their own tool chain. For example, Apache has been changed by moving to a new compiler. It is x86 based, but can be cross-compiled for any platform.
Burke declared that roll-your-own tools is a challenge. developers need a common standard platform. The collaboration for a standard ARM platform is beginning.
Seegars added that the ARM partners are driving the adoption.

AMD’s contributions?
Su answered that they are getting partners and working with them. Ecosystems evolve, and the fabric is the key. Expect growth to start in the next 3-7 years. The x86 is not going away, the ARM will be incremental increases. They have been using the A5 for security functions, so ARM is already an HSA partner.