Several months ago, Hewlett-Packard demonstrated memristor technology, a new system architecture that can be dynamically changed between logic operations and memory storage. Memristor innovator Stan Williams, a senior HP fellow and director of its Information and Quantum Systems Lab,claimed that the new computing paradigm could enable calculations to be performed “in the same chips where data is stored, rather than in a specialized CPU.”
The company also revealed that it had designed the architecture to allow multiple layers of memristive logic to be stacked in a 3D fashion, resulting in a tenfold increase in memory density. Among many authors and analysts who read the news, some eventually reminisced on the innovative history of Moore’s Law and marveled at its progress over both prosperous times in the IT space and rough economic downturns that have stifled chip logic advancement. Now, HP researchers are exploring ways to make their memristor architecture useful in future server and data center designs where data has become an exponentially growing asset and management issue simultaneously.
“Re-thinking the balance of computer, storage and communications will happen, and it will have big implications,” said Partha Ranganathan, a principal investigator and distinguished technologist in the exascale data center project at HP Labs.
Researchers working with memristor technology intend to rebalance these fundamental core system components with a new chip called a “nanostore.” From an architectural perspective, a nanostore is just a 3D stack of processor cores connected to non-volatile memristor cores (NVRAM). The new processor-memory design will essentially place data at the heart of the computing transaction, rather than the CPU itself.
With a new “stateful logic” paradigm that moves from observing the CPU as the “brain” of a computer system to the data itself as the center of a system (based on the nanostore concept), HP Labs has found the new design approach to have a ten times greater performance factor for the same cost of energy. “This is early work in [3D stacks and memristors], and we definitely think we can get better [performance] factors,” said Ranganathan.
He also mentioned that it could take roughly five years before nanostore devices are ready for commercial use. He and other researchers at HP Labs plan to publish papers later this year on the initial ideas for nanostores and for a new low-power processor called a “microblade.” HP Labs has identified three unique kinds of server designs that could be optimized for different types of processing workloads. In an energy proportional design, server performance is dynamically scaled up or down based on an application’s particular needs. In a consolidated design, multiple jobs are packed into a system. In a microblade design, jobs are broken down into highly parallelized tasks than can be handled by multiple low-power processors (think ARM and Intel Atom chips). This last type of design is known in the HPC space as “physicalization,” or in other words, the concept of building high-density compute nodes out of clusters of very cheap, low-power processors.
Of course, microblades using low-power processors are limited by the extent to which underlying algorithms can be split into separate tasks. In a recent article by Jon Stokes from ArsTechnica, he notes that a combination of high margins on server chips and the overall organization of the hardware on the die have left an opening for simpler, cheaper solutions like those from ARM and Intel’s Atom lineup.
In addition, there is also newly envisioned potential for Intel to reassess its stakes in the HPC market with the introduction of its MIC (Many Integrated Core) server architecture. But whether or not MIC will be the last remaining highly parallel, low-power solution to power the data centers of tomorrow before HP mass produces nanostores with memristor technology remains unclear.