ARM reveals more details on supercomputer architecture plans

Scalable Vector Extensions coming to the ARM instruction set will debut in Fujitsu's Post-K supercomputer

ARM has revealed more details about the Scalable Vector Extensions (SVE) technology it is developing for the ARMv8-A architecture that Fujitsu is planning to deploy in the Post-K supercomputer it is building for Japan's RIKEN Advanced Institute for Computational Science.

Detailed at the Hot Chips conference in Cupertino, California today, SVE is capable of handling vectors from 128 to 2,048 bits in length, and is intended for supercomputer makers like Fujitsu to help them adopt ARM products and deploy them in the world's biggest and most powerful computers.

SVE is a flexible extension to the ARM instruction set that can shift vector calculations from software into hardware, with the scheduler arranging calculations depending on the hardware available.

ARM engineers will, shortly, submit patches to the Gnu Compiler Collection (GCC) and LLVM compiler infrastructure project to support SVE auto-vectorisation. This will enable developers to take advantage of the technology without having to customise their applications accordingly. Compiled programs will also be portable and will be able to run on any microprocessor with SVE without needing to recompile the source code.

It follows the selection of ARM by Fujitsu in July for the Post-K supercomputer, which is scheduled to go live in 2020. Fujitsu is planning to shift from 2GHz Sun Sparc64 cores in the K supercomputer to ARM for Post-K. The current K supercomputer ranks fifth in the latest Top500 list of the world's most powerful supercomputers.

If everything goes according to plan, Fujitsu's Post-K machine will become the world's most powerful supercomputer, capable of 1,000 petaflops.

ARM is continuing to develop parts that may enable it to take on Intel in servers and high-end computing. The company points out that the cloud computing workloads, for example, require highly scalable configurations where the key is the interconnect, rather than the raw power of the microprocessors.

Its roadmap foresees ARM partners using the company's products first in storage and web-serving workloads, before shifting up the 'food chain' into data analytics, networking, collaborative platforms and systems management.

"A vast number of the workloads being deployed on servers are things like storage, web applications, and networking. Those kinds of things make up a significant proportion of the server market. They are also some of the fastest growing applications in terms of the workloads being deployed," Lakshmi Mandyam, director of server systems and ecosystems at ARM, told Computing.

These workloads are 'scale out' workloads, she added. "What that means is that they lend themselves to running across many different nodes. Each one of those nodes doesn't have to be very compute intensive. Those workloads are actually more about the input-output and memory performance than they are about the compute performance [of the microprocessor]."

The company therefore sees the cloud computing model becoming more like the high-performance computer market, where the CPUs of the world's most powerful supercomputers don't go for out-and-out raw speed. Instead, the machines in the top-10 typically weigh in at between 1.45GHz and 2.6GHz, but achieve their computational power via the number of cores, and the speed and scalability of the interconnects.

The Sunway TaihuLight, the world's most powerful supercomputer, for example, boasts 10,649,600 cores - three times its predecessor, with typical machines in the top 10 having anything from 115,984 to 1,572,864.