The AMD-powered Frontier supercomputer is now the primary formally acknowledged exascale supercomputer on the planet, topping 1.102 ExaFlop/s throughout a sustained Linpack run, rating first on the newly-released Top500 record of the world’s quickest supercomputers. The system not solely overtakes the earlier chief, Japan’s Fugaku, however blows it out of the water — the truth is, Frontier is quicker than the subsequent seven supercomputers on the record, mixed. Notably, whereas Frontier hit 1.1 ExaFlops throughout a sustained Linpack FP64 benchmark, the system delivers as much as 1.69 ExaFlops in peak efficiency.
Frontier additionally now ranks because the quickest AI system on the planet, doling out 6.88 ExaFlops of mixed-precision efficiency within the HPL-AI benchmark. That equates to 68 million directions per second for every neuron within the mind.
Moreover, the Frontier Take a look at and Improvement (Crusher) system additionally positioned first on the Green500, denoting that Frontier’s structure is now additionally probably the most power-efficient supercomputing structure on the planet (the first Frontier system ranks second on the Top500). The system delivered 52.23 GFlops per watt whereas consuming 21.1 mW of energy throughout the run. At full utilization, Frontier consumes 29 mW.Â
The Frontier supercomputer’s win is however certainly one of many accomplishments for AMD on this yr’s Top500 record — AMD EPYC-powered programs now comprise 5 of the highest ten supercomputers on the planet, and ten of the highest twenty. In actual fact, AMD’s EPYC is now in 94 of the Top500 supercomputers on the planet, marking a gentle improve over the 73 programs listed in November 2021, and the 49 listed in June 2021. AMD additionally seems in additional than half of the brand new programs on the record this yr. Nevertheless, as you may see within the above album, Intel CPUs nonetheless populate most programs on the Top500, whereas Nvidia GPUs additionally proceed because the dominant accelerator.Â
Nevertheless, by way of energy effectivity, AMD reigns supreme within the newest Green500 record — the corporate powers the 4 most effective programs on the planet, eight of the highest ten, and 17 of the highest 20.Â
The Frontier supercomputer is constructed by HPE and is put in on the Division of Power’s (DOE) Oak Ridge Nationwide Laboratory (ORNL) in Tennessee. The system options 9,408 compute nodes, every with one 64-core AMD “Trento” CPU paired with 512 GB of DDR4 reminiscence and 4 AMD Radeon Intuition MI250X GPUs. All instructed, the system has 602,112 CPU cores tied to 4.6 petabytes of DDR4 reminiscence.
Moreover, the 37,888 AMD MI250X GPUs function 8,138,240 cores and have 4.6 petabytes of HBM reminiscence (128GB per GPU). The CPUs and GPUs are tied collectively utilizing the HPE Cray Slingshot networking cloth.
Your entire system is related to an insanely performant storage subsystem with 700 petabytes of capability, 75 TB/s of throughput, and 15 billion IOPS of efficiency. A metadata tier is unfold out over 480 NVMe SSDs that present 10PB of the general capability, whereas 5,400 NVMe SSDs present 11.5PB of capability for the first high-speed storage tier. In the meantime, 47,700 PMR laborious drives present 679PB of capability.Â
Assembling Frontier was a problem unto itself, as ORNL needed to supply 60 million elements with 685 completely different half numbers to construct the system. The chip scarcity hit throughout development, impacting 167 of these half numbers, so ORNL discovered itself brief two million elements. AMD additionally bumped into points as 15 half numbers for its MI200 GPUs encountered shortages. To assist circumvent the shortages, ORNL labored with the ASCR to get Protection Priorities and Allocation System (DPAS) rankings for these elements, that means the US authorities invoked the Protection Act to acquire the elements because of their significance to nationwide protection.
Whereas Frontier will get the nod for the primary officially-recognized Exascale supercomputer on the planet, China is essentially thought to have two Exacscale supercomputers that broke the barrier a yr in the past. Sadly, these programs have not been submitted to the Top500 committee because of political tensions between the US and China. Nevertheless, the shortage of official submissions to the Top500 — a Gordon Bell submission was tendered as a proxy — has led to some doubt that these are true exascale programs, no less than as measured with an FP64 workload.
For now, Frontier is formally the quickest supercomputer on the planet and the primary to formally break the exascale barrier. The nearly-mythical, oft-delayed Intel-powered Aurora is predicted to come back on-line later this yr, or early subsequent yr, with as much as 2 ExaFlops of efficiency, rivaling Frontier for the highest spot within the supercomputing rankings.
Up subsequent for AMD? El Capitan, a 2 ExaFlop machine final recognized to be coming on-line in 2023. Upon completion, it can vie with Aurora for the title of the quickest supercomputer within the Top500. Â
