Supercomputer maker Cray announced what it calls its last supercomputer architecture before entering the era of exascale computing. It is code-named “Shasta,” and the Department of Energy, already a regular customer of supercomputing, said it will be the first to deploy it, in 2020.
The Shasta architecture is unique in that it will be the first server (unless someone beats Cray to it) to support multiple processor types. Users will be able to deploy a mix of x86, GPU, ARM and FPGA processors in a single system.
Up to now, servers either came with x86 or, in a few select cases, ARM processors, with GPUs and FPGAs as add-in cards plugged into PCI Express slots. This will be the first case of fully native onboard processors, and I hardly expect Cray to be alone in using this design.
Also beefing up the system is the use of three distinct interconnects. Shasta will feature a new Cray-designed interconnect technology called Slingshot, which the company claims is both faster and more flexible than other protocols for interconnecting, along with Intel’s Omni-Path technology and Mellanox’s Infiniband.
There has been an effort to improve interconnect technology, since communication between processors and memory is often the source of slowdown. Processors, while not growing at the rate of Moore’s Law anymore, are still left waiting to hear from other processors and memory, so expanding the interconnects has been a growing effort.
Slingshot is a high-speed, purpose-built supercomputing interconnect that Cray claims will offer up to five times more bandwidth per node than existing interconnects and is designed for data-centric computing.
Slingshot will feature Ethernet compatibility, advanced adaptive routing, first-of-a-kind congestion control, and sophisticated quality-of-service capabilities. Support for both IP-routed and remote memory operations will broaden the range of applications beyond traditional modeling and simulation. Reduction in the network diameter from five hops in the current Cray XC generation of supercomputers to three will reduce latency and power while improving sustained bandwidth and reliability.
Cray is looking beyond just the HPC market with Shasta, though. It’s targeting modeling, simulation, AI and analytics workloads — all data-centric enterprise workloads — and says the design of Shasta allows it to run diverse workloads and workflows all on one system, all at the same time. Shasta’s hardware and software designs are meant to tackle the bottlenecks and other manageability issues that emerge as systems scale up.
Slingshot’s architecture is designed for applications that deal with massive amounts of data and need to run across large numbers of processors, like AI, big data and analytics to provide synchronization across all processors.
One sign that Cray is targeting the enterprise is Shasta has the option of industry-standard 19-inch cabinets instead of Cray’s custom supercomputer cabinets, and it supports Ethernet, the data center standard for interconnectivity, along with the standard supercomputer interconnects.
A supercomputer company pushing down into the enterprise will certainly force HPE, Dell, Cisco and the white-box vendors to up their game quite a bit.
https://www.networkworld.com
The Shasta architecture is unique in that it will be the first server (unless someone beats Cray to it) to support multiple processor types. Users will be able to deploy a mix of x86, GPU, ARM and FPGA processors in a single system.
Up to now, servers either came with x86 or, in a few select cases, ARM processors, with GPUs and FPGAs as add-in cards plugged into PCI Express slots. This will be the first case of fully native onboard processors, and I hardly expect Cray to be alone in using this design.
Also beefing up the system is the use of three distinct interconnects. Shasta will feature a new Cray-designed interconnect technology called Slingshot, which the company claims is both faster and more flexible than other protocols for interconnecting, along with Intel’s Omni-Path technology and Mellanox’s Infiniband.
There has been an effort to improve interconnect technology, since communication between processors and memory is often the source of slowdown. Processors, while not growing at the rate of Moore’s Law anymore, are still left waiting to hear from other processors and memory, so expanding the interconnects has been a growing effort.
Slingshot is a high-speed, purpose-built supercomputing interconnect that Cray claims will offer up to five times more bandwidth per node than existing interconnects and is designed for data-centric computing.
Slingshot will feature Ethernet compatibility, advanced adaptive routing, first-of-a-kind congestion control, and sophisticated quality-of-service capabilities. Support for both IP-routed and remote memory operations will broaden the range of applications beyond traditional modeling and simulation. Reduction in the network diameter from five hops in the current Cray XC generation of supercomputers to three will reduce latency and power while improving sustained bandwidth and reliability.
Cray is looking beyond just the HPC market with Shasta, though. It’s targeting modeling, simulation, AI and analytics workloads — all data-centric enterprise workloads — and says the design of Shasta allows it to run diverse workloads and workflows all on one system, all at the same time. Shasta’s hardware and software designs are meant to tackle the bottlenecks and other manageability issues that emerge as systems scale up.
Slingshot’s architecture is designed for applications that deal with massive amounts of data and need to run across large numbers of processors, like AI, big data and analytics to provide synchronization across all processors.
One sign that Cray is targeting the enterprise is Shasta has the option of industry-standard 19-inch cabinets instead of Cray’s custom supercomputer cabinets, and it supports Ethernet, the data center standard for interconnectivity, along with the standard supercomputer interconnects.
A supercomputer company pushing down into the enterprise will certainly force HPE, Dell, Cisco and the white-box vendors to up their game quite a bit.
https://www.networkworld.com
No comments:
Post a Comment