US Dept. of Vitality Broadcasts Frontier Supercomputer: Cray and AMD to Construct 1.5 Exaflop Machine
The historical past of the pc business is a continuing progress. Processors are getting quicker, storage is getting cheaper, and storage is getting tighter. We see the affect of this progress in all sectors of society, and that extends to the highest the place nationwide governments proceed to spend money on larger and higher supercomputers. A part of the technological want and a part of the technological race, the exascale period of supercomputers, is beginning as orders for the primary exaFLOP-enabled techniques are about to run out. It's solely becoming that america Division of Vitality this morning broadcasts the contract for its quickest supercomputer, the Frontier System, constructed by Cray and AMD.
The supply of Frontier is scheduled for 2021, and when activated, would be the second and strongest of the 2 deliberate 2021-Exascale techniques of the US DOE. The efficiency is predicted to succeed in 1.5 exaFLOPS. The bold system, nonetheless, won’t be low cost. Priced at over $ 500 million for the system alone, and one other $ 100 million for analysis and improvement, Frontier is among the many most costly supercomputers the US Division of Vitality has ever ordered.
The brand new supercomputer is being constructed as a part of the Coral-2 supercomputer program for the US DOE. Frontier is to exchange the present supercomputer summit of the Oak Ridge Nationwide Laboratory. Summit is the present reigning champion on this planet of supercomputers with 200 petaFLOPS in efficiency. Accordingly, the US DOE and Oak Ridge wish to considerably enhance its efficiency for the brand new pc. All in all, Frontier ought to be capable of outperform Summit's efficiency by an element of seven, and it’s anticipated to be the quickest supercomputer on this planet after its activation.
Like Summit (and Titan earlier than), Frontier is an open science system, which means that it’s accessible to educational researchers to conduct simulations and experiments. Accordingly, the lab expects the supercomputer for use for quite a lot of initiatives in quite a lot of disciplines, together with not solely conventional modeling and simulation duties, but additionally extra data-driven methods for synthetic intelligence and knowledge evaluation. The latter is certainly a brand new floor for the laboratory and any customers of the system. As we’ve got seen within the Enterprise area lately, neural network-based AI is turning into an more and more widespread methodology of fixing issues and extracting evaluation from massive datasets. Researchers at the moment are exploring how one can optimize these applied sciences from the Web utilizing the newest technology techniques and their use on the exascale degree.
Supercomputer of the US Division of Vitality
Intel Scalable Xeon
~ 30 MW
N / A
N / A
Frontier: Powered by Cray & AMD
Formally, Cray would be the prime contractor for Frontier. However should you have a look at the specs, you may apologize for AMD. Cray, in flip, works with the system's chipmaker, so AMD gives a lot of the core for the brand new supercomputer. Designed because the next-generation CPU + accelerator system, AMD gives a mixture of CPUs and GPUs that do the heavy arithmetic. Each the CPUs and the GPUs for Frontier are equipped. As the first processor vendor, AMD can even be liable for growing the software program stack. The corporate is working with Cray to develop an improved model of its ROCm atmosphere to optimize efficiency from the large cluster of CPUs and GPUs.
On the CPU facet, AMD will provide a personalized next-generation EPYC CPU. AMD has confirmed that it’ll use a future technology of its Zen CPU cores, and given the timing of the undertaking, we’ll virtually actually be coping with a Zen three or Zen four design right here. It stays to be seen what the customized AMD CPU is, however their announcement has proven that the Frontier CPUs will embrace new directions for optimizing AI and supercomputing workloads.
On the GPU facet, AMD and Cray at the moment are nearer to their playing cards. Relatively than naming an structure or architectural technology, AMD merely says that the GPUs are "based mostly on the Radeon Intuition household" and "not but introduced". AMD's present public roadmap targets Subsequent Gen GPU improvement cycles of a mean of two years by 2020, which often is the structure we see. However with the particular wants of a supercomputer, AMD might have one thing extra particular.
What the corporate is confirming for now’s that it doesn’t hold its options. The HPC-focused GPU was developed for Frontier and can present help for Blended Precision Compute. The animal can be fed with HBM reminiscence and AMD can be tapping a model of Infinity Cloth to attach the CPUs and GPUs.
In reality, whereas AMD has retained the main points of the expertise gentle, it feels like this model of IF is probably the most superior model. AMD factors out that it’s an "extremely coherent material", the primary totally optimized CPU + GPU design for supercomputing. AMD's GPUs and CPUs are organized in a four to 1 ratio, with every EPYC CPU outfitted with four GPUs. It's price noting that AMD's slide has a mesh with every GPU hooked up to the CPU and two different GPUs, however I'm not studying this but as a result of AMD has not disclosed any additional particulars concerning the IF setup.
When AMD ascends to the blade degree, assembling all these nodes is the job of Cray. For Frontier, the supercomputer producer launches its new Slingshot connection, an equally bold connection that helps adaptive routing, congestion administration, and repair high quality options. Slingshot helps 200 Gbps per port, with particular person blades containing one port for every GPU within the blade in order that different nodes can learn and write knowledge on to the reminiscence of a GPU. Because of this, Frontier can have a big quantity of hyperlink bandwidth, which is nearly sufficient to scale the system to the ExaFLOP degree.
General, Frontier is organized in over 100 Cray Sashta cupboards. And whereas Cray has not introduced a selected energy consumption for Frontier, with every cupboard designed for 300 kW, this may deliver the general system to over 30 MW. What places issues in context is greater than double the ability consumption of the 13MW Summit. Whereas Frontier is a a lot quicker system than the supercomputer it replaces, Cray, AMD, and the US Division of Protection are seeing Moore's regulation sluggish as power effectivity turns into more and more tough. All in all, a preliminary remark within the press convention feels like Oak Ridge is putting in a complete of 40MW for Frontier, which, to say the least, is a substantial quantity of power.
Along with selling america' personal supercomputing management targets, securing the Frontier contract has additionally been an excellent success for Cray and AMD. Cray is now concerned in each exascale techniques from 2021 and strengthens its personal place on this planet of supercomputing. In the meantime, AMD, giving up this technology from the surface, has achieved a giant and prestigious victory for each the CPU and GPU departments.
In reality, it's attention-grabbing to notice that of the 2 2021 Exascale techniques ordered, each come from full-service processor distributors delivering each CPUs and GPUs. Techniques of the present technology similar to Summit use combined suppliers – eg. IBM + NVIDIA – Shifting to built-in distributors is a giant shift for these CPU + accelerator techniques. Utilizing a single vendor for all processors gives expertise and procurement advantages, which is helpful for each AMD and Intel. Whereas it's price noting that the CORAL-2 program obliges the DOE to purchase techniques with two totally different architectures, so if the long run is built-in techniques, then AMD and Intel are the logical alternative.
In any case, the work with the contract concluded for Frontier is simply half carried out. AMD and Cray might want to evolve their and software program for the system, to not point out particular specs for the completed supercomputer. One can assume that over the subsequent few years there’ll at all times be information about Frontier resulting in its set up in 2021.