The ARM approach to Big Data & High Performance Computing (HPC) or Supercomputers is all about building a balanced approach, not taking any one direction to an extreme - in other words, we maximize efficiency by not over-designing any particular component. I believe Big Data & HPC is all about taking a more systemic/holistic approach by balancing the overall parallelism (=performance), locality (=efficiency) and density (=cost).
We have seen a significant shift in the industry. It isn’t a new shift it’s been happening for a while. The industry can no longer rely solely on process reduction as a method to improve performance. PPA (Performance Power Area) will increasingly become reliant on silicon integration.
ARM focuses on designing & licensing fundamental IP building blocks - which is all about integration. Integration fosters an ecosystem of standard pieces, effectively acting as COTS-on-Silicon (my invented term). COTS-on-Silicon encourages multi-suppliers to give the ecosystem the ability to deliver, through collaboration, continuous innovative solutions for Big Data & HPC type problems.
These solutions can be localized and optimized for an end design. This effectively enables the circuit board to be miniaturized onto a single chip - and retain energy efficiency and control cost.
The next era of systems will follow a-build-what-you-want style. By allowing the targeting of a SoC (System-on-Chip) to solve specific problems – one solution doesn’t fit both Big Data and compute. A SoC can be targeted to optimize the power-performance for Big Data or HPC utilizing common infrastructure and components and leveraging the vast software ecosystem.
The ARM business model allows a SoC designer to take standard parts and add specific value to solve a problem. Specific value is in the form of specialized IP. And everything else comes off-the-shelf. This creates a diverse and competitive environment which can be leveraged to bring rapid innovation.
The balanced approach is going to affect the industry. There is no point providing a single attribute when it is detrimental to the other attributes. A single direction doesn’t scale. The technology has to be balanced. By selecting a careful technology balance (such as ARM’s big.LITTLE™ processing, Mali® GPGPU, network I/O, external memory, CoreLinkTM on-chip interconnect, off-chip interconnect etc.) companies can select the right PPA fit-for-purpose.
NOTE: area is effectively translated to cost since the larger the silicon area used the higher the cost.
It means that supercomputer procurement has a much broader choice of systems. Where the balanced approach allows them to take the data feed of the system’s capability and pair-it precisely with the compute-node performance. This opens up the ability to match a large-scale facility with a problem domain. Previously such a versatile approach had been quite limited.
We can no longer solely rely on process reduction to improve performance. Advancing the PPA will increasingly become reliant on integration. The IP business model fosters a healthy ecosystem of tested components, which makes integration easier. This allows tailored solutions to be created for Big Data & HPC problems. COTS-on-Silicon encourages multi-suppliers throughout the ecosystem. The ecosystem can deliver cost-effective solutions in sensitive markets.
The collaborative balanced approach is changing the way that Big Data & HPC systems are going to be designed. This approach allows the architects to match the solution with the problem domain by pairing network fabric characteristics with the right level of compute-node performance.