CHIP DESIGNER AMD said it needs to get ARM CPU cores accessing the same memory as its GPU cores before it can bring out ARM accelerated processing units (APU).
AMD's decision to design ARM chips for servers means that the firm can combine the ARMv8 architecture with its own GPU, something the firm already revealed to The INQUIRER. Now AMD has said that it needs to ensure both the CPU and GPU can access the same shared memory before it will launch an ARM based APU.
Suresh Gopalakrishnan, corporate VP and GM of AMD's Server Business Unit told The INQUIRER the firm could already design an ARM chip with a CPU and GPU that would each have their own memory, however that won't bring the advantages in both programmability and performance of having a shared memory structure between the two cores. He also talked up OpenMP support in future ARM APUs, an application programming interface that both Nvidia and Intel are working to support in their Tesla and Xeon Phi accellerators, respectively, for sharing memory among different cores.
Gopalakrishnan said, "You'll see the ARM APU first [before an Opteron APU]. You'll see the first ARM APU at a similar time as the first ARM chip. The value proposition [of an APU] is that we can go to the memory controller to access a common memory pool, that is what we have already done between our X86 and GPU cores. ARM has joined the HSA foundation and we have some work to do to make the ARM [core] and the GPU access the same memory. [...] Once that is done we can bring an ARM APU to the market."
AMD had previously said that Opteron APUs were in the pipeline, but Gopalakrishnan's comments suggest that Opteron APUs will not appear before 2014 at the earliest. It is a shame for the firm, which could have played on its GPU technology to win some server business from Intel and its Xeon chips. Nonetheless, AMD must think it has a better chance of gaining market share in the newer ARM server market. µ