Numascale has announced the launch of its NumaConnect SMP adapter. This represents the beginning of a new era for HPC users, allowing high-performance SMP systems to be built using standard servers and hardware adapters at cluster prices.

Numascale is showcasing a multi-node system where the NumaConnect SMP adapters are plugged into the HyperTransport connector in each node, and are interconnected through a switchless 3D torus to form a cache coherent shared-memory NUMA system. With no OS or software changes, the system runs one single OS image with the total memory shared and available to all processors.

At the heart of the NumaConnect technology is the NumaChip, which combines a high-speed, low-latency 3D torus interconnect and a cache coherent memory controller. The unique, directory-based caching architecture is designed with 3 levels of cache and onboard cache memory, helping to reduce the amount of memory transfers needed between nodes, thus effectively increasing performance and scalability. The NumaChip design allows for creating single systems with up to 256TB of shared memory.

As clusters have become the dominant system architecture in HPC today, the NumaConnect SMP Adapters provides a way to build scalable SMP systems at more affordable cluster prices. "Systems based on standard high volume servers interconnected with the full-blown Numascale technology with distributed cache coherent memory will be extremely interesting for a wide range of compute-intensive applications that today can only be efficiently executed on systems costing significantly more," says Petter Bjørstad, founder of Parallab and professor and head of the Department of Informatics, University of Bergen.

There is also a potential 40 percent operational cost savings over cluster alternatives due to a very high utilization rate, low cost for power and space, high programmer productivity, and simplified system management. "It is a very exciting time to bring this breakthrough technology to market," says Kåre Løchsen, CEO of Numascale. "For years the HPC market has been forced to pay a very high premium for large scale shared memory SMP systems. With NumaConnect, we will enable the building of systems that have similar specs as competing systems, but at a fraction of the cost of its enterprise alternatives."

In addition, as users can run MPI code over NumaConnect, this architecture will be an interesting alternative to both large enterprise SMP as well as cluster technology. "Our sales model should be very compelling to end-users, integrators, and OEMs as we intend to ship NumaConnect at a component level for utilization within a large range of systems and application areas," continues Mr. Løchsen. "Our plan is to engage across the HPC community and allow our end-users to work with a choice of vendors or integrators to solve their most complex problems with the lowest possible total cost of ownership."