AMD Bets On APU Servers, Rolls Out The Developer Tools
“If you build it, they will come,” goes the thinking at AMD
AMD has released a set of developer tools for Accelerated Processing Units (APUs) ahead of the launch of ‘Berlin’ – the first ever server chip based on Heterogeneous System Architecture (HSA), expected in the first half of 2014.
The new software includes CodeXL 1.3, Project Sumatra, Project GCC/HSA, clMath, PGI Accelerator Compiler and ArrayFire 2.0.
The announcement was made at the APU13 event in San Jose, California.
“Servers must be efficient, scalable and adaptable to meet the compute characteristics of new and changing workloads. Software applications that leverage server APUs are designed to drive highly efficient, low-power, dense server solutions optimized for highly parallel and multimedia workloads,” said Suresh Gopalakrishnan, corporate vice president and general manager of the Server Business Unit at AMD.
Anticipation
In 2013, AMD introduced two product families aimed at the data centre – the more traditional Opteron 3300, and the Opteron X-Series – the world’s first server-class x86 APU, to be used in HP’s Moonshot software-defined servers.
APUs from AMD, first launched in 2011, act as a CPU and a GPU on a single die, improving data transfer rates between these components while reducing power consumption. “It’s not like the idea is new, but now we have the technology to put that powerful GPU back on the same piece of silicon as the CPU,” Margaret Lewis, director of Server ISVs and software planning at AMD, told TechWeekEurope.
The company has a hard task ahead of it: essentially, it needs to maintain two separate software ecosystems – one for APU servers, and one for its ARM-based servers.
In the first quarter of 2014, AMD will launch ‘Warsaw’ chips built to Open Compute specifications, offering better performance-per-watt than Opteron 6300. And in the second half of the year, we’ll see ‘Seattle’ – its first 64-bit ARM System-on-a-Chip.
However, the most interesting piece of silicon to hit the server market in 2014 has to be Berlin – a family of low power, cost-effective APUs for ultra-dense, fabric-based servers. These will be aimed at the enterprise customers that run Web applications, media content or gaming services. The easily scalable processors can also be used to provide hosted desktops, and could even have some High Performance Computing applications, for example in seismic analysis and simulations.
New tools
Over the last few years, AMD has worked with its partners and the open source community to develop software tools for these purposes, which it unleashed on Monday evening.
The developer toolkit for server-side APUs now includes CodeXL 1.3, AMD’s own suite for Windows and Linux that features remote debugging and profiling. Next is Project Sumatra – a joint open source project managed by Oracle and AMD that enables developers to code in Java and take advantage of GPU compute.
AMD has also cooperated with SUSE on GCC/HAS, enabling the popular open source Linux compiler, GCC, to support HSA. Meanwhile the OpenCL math libraries, contributed to open source in August under the name ‘clMath’, will enable developers to accelerate common scientific and engineering computations on AMD APUs and discrete GPUs.
Finally, there’s a new version of ArrayFire for OpenCL – a fast math library by AccelerEyes that uses clMath for GPU computing and offers an easy-to-use API for Windows or Linux developers.
AMD also announced a new unified SDK, tools and accelerated libraries for heterogeneous computing developers.
“You can get the tools in place, and then you give them to developers, and they do things with them that you never thought they would. These guys will figure out how to do things we didn’t think even have a use case,” said Lewis.
Do you know all about Green IT? Take our quiz!
