While the enterprise remains circumspect on prospects for Arm servers in the datacenter, the leadership HPC community is taking a bolder, brighter view of the x86 server CPU alternative. Amongst current and planned Arm HPC installations – i.e., the innovative Mont-Blanc project, led by Bull/Atos, the ‘Isambard’ Cray XC50 going into the University of Bristol, and commitments from both Japan and France among others — HPE is announcing that it will supply the United States National Nuclear Security Administration (NNSA) with a 2.3 petaflops peak Arm-based system, named Astra. On track for deployment at Sandia National Labs later this year as part of the NNSA’s Arm-centric Vanguard project, Astra will be the world’s largest Arm system ever built, according to HPE.
Based on the Apollo 70 servers announced by HPE at SC17, Astra will encompass more than 145,000 Cavium ThunderX2 cores in 2,592 dual-socket servers, with four compute nodes arranged in a dense 2U form factor. The system will employ a total of 5,184 Cavium 64-bit ARMv8-A ThunderX2 28-core, 2.0 Ghz processors, which have a TDP of 150 watts.
The ThunderX2 line, which became generally available in May, includes SKUs with as many as 32 cores, with clock speeds up to 2.5 MHz (plus a bump for turbo mode). In an interview with HPCwire, Mike Vildibill, vice president of HPE’s Advanced Technology Group, describes the 28-core, 2.0 Ghz CPU as “the Goldilocks part that provided the full memory performance that [Sandia] needed in a package that perfectly fit the price requirements of the system as well.”
The CPUs will be connected with Mellanox EDR InfiniBand in very bandwidth intensive configuration. With eight memory channels on each Cavium ThunderX2 Arm processor compared to six on today’s traditional CPUs, the Apollo servers offer 33 percent more memory bandwidth than Intel x86 servers. Those eight channels will support 8GB dual-rank DIMMs, providing 332 TB aggregate memory capacity and 885 TB per second of aggregate memory bandwidth. Sandia could have opted for higher-capacity DIMMs, but was able to satisfy its memory bandwidth requirements at a more optimum price point.
Vildibill, who is co-PI on HPE’s PathForward (exascale-focused) activities, emphasized the data-centric nature of the system, and drew a line from The Machine research program to Vanguard Astra to the company’s exascale targets. The prototype systems HPE built for The Machine were based on very early examples of the ThunderX2 processor, tracing back to when the technology was still part of Broadcom’s Vulcan brand (before Cavium acquired the Vulcan IP). Under HPE’s Comanche development program (which would become the Apollo 70 line), HPE delivered prototype systems to a number of DOE labs. Sandia has been running a rack of those systems for eight months, giving it lead time to begin the software porting and demonstration necessary to make Astra a usable system, according to Vildibill.
You can see how Astra (referenced as Vanguard-1) fits into Sandia’s Vanguard project– it’s ARM-centric co-design strategy with Los Alamos and Livermore Labs – in this slide from a March 2018 meeting.
The mission of the Vanguard project is to expand NNSA’s tri-lab HPC ecosystem by developing technologies viable for future ATS/CTS platforms supporting Advanced Simulation and Computing codes. It is a requirement of Vanguard systems that they be “large enough to serve as proof of concept for future Advanced Technology Systems, and sufficient scale to interest production multi-physics application teams.”
“We are committed to fielding the most advanced technologies as part of the Vanguard program,” said James Laros, Vanguard project lead at Sandia National Laboratories, in HPE’s official announcement. “We are collaborating with HPE to advance the Arm ecosystem and prove the viability of this architecture to support our national security mission.”
The Astra design is part of the larger longer term path to exascale for NNSA and for HPE. One of the biggest draws of Arm, according to Vildibill, isn’t necessarily the energy efficiency that has been associated with the platform, but rather the innovation cycle that is possible. “Innovation for Arm-based systems can be accelerated at a faster rate compared to the innovation cycle for processors that are being used for both HPC and enterprise markets at the same time,” Vildibill remarked. “It’s very difficult for some vendors to build a CPU that is aggressively adopting all the newest technology while also supporting a very long tail of legacy applications that go throughout the enterprise and so the [increased number of] memory channels is just one example. As we look out to exascale, we clearly see that today’s architectures cannot get us to exascale in a 20-30 MW envelope in the future; it just can’t, and that we absolutely see the need for innovation and some invention between now and exascale.”
The power footprint of Astra’s compute nodes will be about 1.2 MW, according to HPE, which compared to an equivalent top system of today, stands at about half the power consumption and three times more memory bandwidth.
Astra is also outfitted with an all-flash Lustre file system, essentially acting as a burst buffer tier, based on HPE’s Apollo 4520 platform, that Vildibill calls “Lustre in a box.” It’s a 4U box with a usable capacity of 366 TB, enabling nearly 250 Gigabytes per second of bandwidth.
Keeping Astra cool is the job of HPE’s new MCS-300 Cooling Unit, which company literature describes as “a liquid-to-air heat exchanger and fan system that removes heat generated by rack-installed IT equipment.”
The announcement from HPE and Sandia Labs is a prominent feather in Arm’s cap but Cavium’s implementation of Arm isn’t the only Intel-alternative on the field: AMD’s x86 Eypc server chip and IBM’s Power CPU are also in play, and RISC-V is gathering some steam as well. Vildibill affirmed there are many degrees of freedom in this discussion. “HPE is exploring a number of architectures as part of our exascale development activities,” he told us. “I would say absolutely exascale is not only an Arm play, but we do believe Arm is one of several emerging technologies that are relevant in exascale.”
This increased processor diversity plus above-average M&A churn have created disruption in the technology space that everyone is facing. ThunderX maker Cavium is currently being acquired by Marvell, which means three companies will have held the Vulcan/ThunderX2 technology in as many years, but given the healthy investments coming from the global HPC community, the processor’s prospects are looking good.