Today, AMD announced their new strategic initiative to produce ARM architecture based Opteron processors by 2014. This new initiative will be complementary in nature and will not cannibalize nor affect their existing processor variety such as traditional x86 processors and Fusion concept based APUs. The main goal for AMD’s upcoming ARM-based processors is to address a growing need for a higher compute power per unit of cost and electricity consumed. In essence, more efficient processors to cater to more serialized workloads.
While the news of a major CPU manufacturer opting for an ARM-based server processor architecture might be surprising to some, it’s actually not far fetched when you consider that NVIDIA actually first committed to develop ARM-based processors in a big way when they first mentioned of their project Boulder. Having mastered the mobile computing space to deliver powerful Tegra mobile processors for both ‘super’ phones and tablets, NVIDIA’s CEO revealed their intentions to scale their ARM processor design know-how to build a processor containing several ARM based cores to deliver a server class processor to challenge AMD and Intel in the traditional server market.
With AMD’s upcoming expanded portfolio of server class CPUs as announced today, ARM-based processors in traditional server class computing environments are no longer just an idea nor a concept.
As AMD puts it, the strategic partnership with ARM represents the next phase of AMD’s strategy to drive ambidextrous solutions in emerging mega data center solutions. The explosion of data created and the demand for compute power is growing exponentially and data centers are where all of this are managed. These data centers themselves see a growth of 30% a year and to capitalize on it, AMD requires new solutions to tackle the growing workloads that are driving new business models.
These workloads are however changing and no longer fit the traditional one size fits all model with traditional compute power. AMD has identified that the fastest growing workloads currently are the small and highly parallel workloads which are best handled by ARM’s architecture. As a result, it will help deliver higher compute per dollar and it also more power efficient.
Of course AMD admits that ARM-processors alone aren’t the best solution for all needs and it only address some aspects efficiently. For example, ARM based processors cannot populate network links as well as x86 CPUs do. Since most ARM processing cores are designed to handle small workloads and require many cores to scale up their compute power, it’s inefficient to link them each direct to the network. To solve this, AMD will leverage on its SeaMicro Freedom supercomputer fabric (an interconnect technology that provides low latency, high bandwidth transactions with built-in redundancy) to efficiently link of processors in a cluster which will in turn link up to the network for a more flexible and complete processing solution for a data center. It also the only interconnect fabric that supports both Ethernet and storage traffic and that really plays into AMD’s deployment advantage for simplicity.
Another strong advantage of AMD’s SeaMicro Freedom supercomputer fabric is that it’s compatible with difference processor types from CPUs, APUs to different architectures like x86 and ARM. This means AMD has the capability of interfacing with different computing nodes of various processing capabilities and even to the extent of creating a processor with various types of processing cores in one chip. As such, AMD’s Fusion computing ecosystem is extended to well beyond the mobile and desktop computing space to even server and data center levels. In some ways, it can be seen as a form of an evolution from AMD’s Torrenza initiative to leverage on various ‘accelerators’ that are best suited for the job.
AMD has multiple challenges ahead, from delivering the actual processor to being a proponent of this new initiative and get developers behind it to fulfill the software ecosystem that can take advantage of this new processor.
First of all, one needs to understand that this entire initiative isn’t just solely AMD’s choice to differentiate themselves, but to also deliver a solution that customers like Amazon, Dell, Facebook and Red Hat that are interested to invest in more efficient compute farms suitable for their workloads. As sighted above, there are workloads that favor the traditional servers, and there are workloads that would fare better if it's processed on an ARM based processing architecture.
Secondly, AMD mentioned that the production of these multi-core 64-bit ARM based Opteron processors will commence sometime in 2014. This would mean actual delivery and solutions based on these chips can be expected in the later half of 2014, or at the latest, sometime in 2015 should there be setbacks. That’s still a very long time away and market dynamics may yet change in the next few years. While they are not abandoning their current processor line-up, the added research and effort into this new processing solution is still a big undertaking and risk.
Last but not least, as summarized above, AMD would need to help develop an appropriate software ecosystem and that will take time to develop as there aren’t suitable processing solutions to emulate on at the moment. This would mean even if the hardware is ready by 2014 or 2015, the actual usability of this platform to its full capabilities could take a little longer than that.
It is certainly interesting times ahead and we wish AMD all the luck to drive the change in the data center landscape, just as it did with 64-bit x86 processors back in 2003 for the server and desktop space.