ARM and chip partners Applied Micro and Cavium are expanding efforts to bring the ARM architecture into the high-performance computing world.
At the International Supercomputing Conference (ISC) July 13, both chip makers announced moves that they hope will enable their 64-bit ARM-based products to gain more traction in a high-performance computing (HPC) market that is dominated by the likes of Intel, IBM, Advanced Micro Devices and Fujitsu.
Applied Micro announced platform partnerships with HPC system makers Cirrascale and E4 Computing Engineering based on its X-Gene chip technology, while Cavium officials unveiled support in its ThunderX systems-on-a-chip (SoCs) for such technologies as flash memory and InfiniBand connectivity as well as enhanced partnerships with companies like Penguin Computing and E4.
"Our ThunderX ARMv8 CPUs were designed specifically for high-performance: Our unparalleled 48 cores, high-memory bandwidth and capacity, and integrated workload accelerators are a great match for HPC and hyperscale workloads," Steve Cumings, director of market development for Cavium, said in a statement when talking about initial orders E4 is now taking for its ThunderX-based Arka servers. "The value of accelerated performance and reduced [total cost of ownership] is highlighted by the orders E4 customers are already placing."
Officials with ARM and partners—not only Applied Micro and Cavium, but also AMD and Marvell Technology—have been talking for several years about taking its low-power chip architecture, which is found in most smartphones and tablets, into the server space. Initially the focus was on highly dense, low-end servers. However, HPC has become an area of interest, given the demand for high performance and energy efficiency.
They're making strides. Supercomputer maker Cray said in November 2014 that it will evaluate using ARM-based chips in its systems and is working with Cavium to create compute clusters powered by the chip maker's 48-core ThunderX SoCs. In March, Cavium announced it was adding support for Nvidia's Tesla GPU accelerators to its processors, a move that will help Cavium make ThunderX more attractive to organizations in such areas as HPC, scientific computing and cloud computing, according to the company.
Companies running HPC environments increasingly are turning to GPU accelerators like those from Nvidia and AMD, or x86-based coprocessors from Intel, to boost the performance of their systems while holding down the power consumption.
At the ISC, Applied Micro officials said the company is partnering with Cirrascale on the RM2916 HPC Platform, which is powered by the X-Gene SoC, Nvidia's Tesla K40 GPU accelerator and Mellanox's ConnectX-3 high-speed InfiniBand adapter.
"The combination of the power-efficient X-Gene processor and a high-performance GPU compute engine, with a robust Linux-based software environment is very well suited for a range of scale-out computing deployments," John Williams, vice president of marketing at Applied Micro, said in a statement.
Applied Micro also announced the latest generation of E4's Arka RK003 HPC platform that also includes the Nvidia and Mellanox technologies. Company officials said they will be demonstrating the Cirrascale and E4 systems at the ISC.
For its part, Cavium—along with announcing the initial orders for E4's ThunderX-based Arka servers—said Penguin Computing is introducing its Tundra server family based on Cavium's 48-core SoCs and compliant with the Open Compute Project specifications. Penguin officials said the Tundra servers will focus on hyperscale and HPC workloads, such as big data, large-scale graph analytics, molecular dynamics and cloud storage. The systems will be ready for customer evaluation starting in September, with general availability coming in the fourth quarter.
Cavium also announced support for SanDisk flash memory products and a partnership with Mellanox to bring InfiniBand connectivity to the ThunderX chips.
"High-performance compute clusters leveraging high-performance scale-out compute technologies based on ThunderX require a very high bandwidth and ultra-low-latency cluster interconnects," Gopal Hegde, vice president and general manager of Cavium's Data Center Processor Group, said in a statement, noting that the ThunderX SoCs already offer high-memory bandwidth and integrated accelerators. "This collaboration with Mellanox will enable an InfiniBand interconnect option to our customers for high-performance and ultra-low latency interconnects for their HPC platforms based on ThunderX processors."
Cavium also is rolling out the beta release of PathScale's EKOPath HPC compiler and library software optimized for ThunderX.
In addition, company officials announced that researchers for the Mont Blanc project at the Barcelona Supercomputing Center had successfully tested a ThunderX- based server platform that was integrated by E4 engineers.