Micron unveils new parallel computing architecture

November 19, 2013 | 09:59

Tags: #big-data #coprocessor #co-processor #dimm #fpga

Companies: #micron

Micron has announced what it claims is a fundamentally new computing architecture designed with heavy parallelism in mind, using the paradigm of dynamic random access memory (DRAM) to speed the search and analysis of complex and unstructured data streams.

Dubbed Automata, the technology looks to all the world like a standard dual in-line memory module (DIMM) but where one would normally find DRAM chips the board instead plays host to a series of Automata co-processors. Based around a mesh fabric design, which sees tens of thousands - scaling to millions, the company claims - of simple processing elements being enabled or disabled to create a task-specific processing engine, the technology has much in common with field-programmable gate arrays (FPGAs) but with a focus on mining big data in record time.

'Micron's Automata Processor offers a refreshingly new way of solving problems that is very different from all other accelerator technologies,' claimed Srinivas Aluru, professor of computational science at the Georgia Institute of Technology and one of the testers of pre-release Automata hardware, at the unveiling. 'By deploying this in interesting ways, we have been able to solve a much larger instance of the NP-hard biological motif-finding problem than was previously reported, using the resources within a single Automata Processor board.'

The system achieves high performance by tackling computation from the opposite direction to a standard central processing unit (CPU): where the CPU applies a single instruction to as many chunks of data as possible before moving onto the next instruction, an Automata processor runs thousands of instructions simultaneously with a view to solving a given problem in the least amount of time.

Micron, traditionally a memory specialist, is keen to point out that its technology is no vapourware - starting with the admission that similar technologies, from tile-based highly-parallel many-core co-processor architectures to the aforementioned FPGAs, already exist. The company is to publish a technical paper describing the Automata processor design and development in the IEEE Transactions on Parallel and Distributed Systems journal - with a pre-release copy available here (PDF warning) - and claims to have taped out the first silicon with prototypes in-hand. The first engineering samples, the company claims, will be ready next year - although the company isn't yet sharing a date for mass availability.

Based on a DDR3-like memory interface, Micron has stated it plans to make the Automata processors available as either individual components for use in embedded systems or as DIMM-packaged boards. A PCI Express card with several Automata DIMMs on-board is also in the works, allowing developers to get started ahead of the technology's integration into future motherboards. A software development kit is expected to be released in the first half of 2014.

Micron has already listed a swathe of markets where the Automata could find use, ranging from bioinformatics and other data-heavy scientific fields to video and image analysis and even network security scanning. To help drive adoption of the technology, the company has also announced a deal with the University of Virginia on the first Centre for Automata Computing.

It'll be a while before mainstream users are clamouring for Automata support, however: the co-processors are designed specifically for data analysis and pattern-matching, and are unlikely to find much use in a gaming rig or office desktop.
Discuss this in the forums
YouTube logo
MSI MPG Velox 100R Chassis Review

October 14 2021 | 15:04