At SC15, Numascale announced the availability of NumaConnect-2, a scalable cache coherent memory technology that connects servers in a single high performance shared memory image. The NumaConnect-2 advances the successful Numascale technology with a new parallel microarchitecture that results in a significantly higher interconnect bandwidth, outperforming its predecessor by a factor of up to 5x. In capitalizing on Altera’s Field Programmable Gate Array (FPGA) technology, Numascale now offers an interconnect node capable of scaling with anticipated industry bandwidth demands, and featuring the flexibility to adapt to emerging standards and customer requirements.
Numascale’s top-selling R Analytics Appliance, which performs in-memory analytics on large data sets, can now receive frequent hardware performance upgrades as well as software improvements. The NumaConnect-2 will improve the overall performance of the R Appliance by 2x out of the gate, and the flexible FPGA architecture will allow Numascale to roll out frequent performance updates to the R Analytics Appliance over the coming years. This is especially appealing to a wide range of Numascale customers who are working on more and more complex problems, creating a need for ever more optimized hardware and software.
The NumaConnect-2 provides 30 GB/s link capacity (including switching) per node and global real-time clock synchronization for efficient OS scheduling on large systems. The interconnect also features a CPU memory coherent interface, an increased maximum node memory size of 768 GB, an increased fabric link bandwidth (x2), and support for directly connected moderately sized systems.
The new NumaConnect design allows shared memory systems with up to 256 TB of memory to be created from cost-effective commodity servers. Large NumaConnect-2 systems can be created using a Dragonfly or 3D Torus topology. Previous NumaConnect configurations were based on ring structures in 1, 2, or 3 dimensions, requiring anything beyond two nodes to be indirectly connected. The NumaConnect-2 allows more flexible topologies by providing two options for the link layer logic, depending on the desired system size.
Implemented in an Altera Stratix V FPGA, the NumaConnect-2 design uses a new parallel microarchitecture that supports more outstanding memory transactions, which in turn increases through-chip bandwidth. Support for the new HT-3 interface also increases the bandwidth between the processor and the NumaConnect-2 to 12.8 GB/sec.
Altera was chosen as a partner for this upgrade based on their FPGA technology and their integrated synthesis tool. According to Einar Rustad, CTO, Numascale, “The partnership with Altera is important because Altera can provide better performance with FPGA design than we were able to with our previous ASIC design. It also saved about a year of development time, and Numascale was able to spend that development time on architectural optimization of the current design instead.”
For their part, Altera viewed Numascale’s innovative technology as a key to opening up a new class of analytics computing appliances. Mike Strickland, Director of Strategic Marketing for Computer & Storage at Altera, said, “We are excited that Numascale chose Altera to deliver the flexibility and productivity of FPGA technology to the next generation of their products. Our FPGA allowed the NumaConnect-2 to deliver a compelling solution for speeding up in-memory data analytics.”
The NumaConnect-2 is currently available and can be found in Numascale’s Data Analytics clusters.