With China threatening to construct the world’s first exascale supercomputer earlier than the US, the US Department of Energy has awarded a analysis grant to Hewlett Packard Enterprise to develop an exascale supercomputer reference design primarily based on expertise gleaned from the The Machine, a challenge that goals to “reinvent the fundamental architecture of computing.”
The DoE traditionally operated many of the world’s high supercomputers, however lately China has taken over in dramatic trend. China’s high supercomputer, Sunway TaihuLight, at present has 5 occasions the height efficiency (93 petaflops) of Oak Ridge’s Titan (18 petaflops). The US has gesticulated grandiosely about retaking the supercomputing crown with an exascale (1,000 petaflops, 1 exaflops) supercomputer that may be operational by 2021ish, however China is seemingly forging forward at a a lot quicker clip: in January, China’s nationwide supercomputer centre mentioned it could have a prototype exascale laptop constructed by the top of 2017 and operational by 2020.
To create an efficient exascale supercomputer from scratch, you could first
invent the universe clear up three issues: the inordinate energy utilization (gigawatts) and cooling necessities; growing the structure and interconnects to effectively weave collectively tons of of 1000’s of processors and reminiscence chips; and devising an working system and shopper software program that really scales to 1 quintillion calculations per second.
You can nonetheless bodily construct an exascale supercomputer with out fixing all three issues—simply strap collectively a bunch of CPUs till you hit the magic quantity—however it will not carry out a billion-billion calculations per second, or it’s going to be untenably costly to function. That appears to be China’s strategy: plunk down many of the in 2017, after which spend the subsequent few years attempting to make it work.
The DoE, alternatively, is wending its method down a extra sedate path by funding HPE (and supercomputer makers) to develop an exascale reference design. The funding is coming from a DoE programme referred to as PathForward, which is a part of its bigger Exascale Computing Project (ECP). The ECP, which was arrange beneath the Obama administration , has already awarded tens of hundreds of thousands of to numerous exascale analysis efforts across the US. It is not clear how a lot funding has been obtained by HPE.
Exascale computing within the USA
So, what’s HPE’s plan? And is there any hope that HPE can go via three rounds of the DoE funding programme and construct an exascale supercomputer earlier than China?
HPE is proposing to construct a supercomputer primarily based on an structure it calls Memory-Driven Computing, which is derived from elements of The Machine. Basically, HPE has developed numerous applied sciences that permit for a large quantity of addressable reminiscence—apparently as much as four,096 yottabytes, or roughly the identical variety of atoms within the universe—to be pooled collectively by a high-speed, low-power optical interconnect that is pushed by a brand new silicon photonics chip. For now this reminiscence is risky, however ultimately—if HP ever commercialises its memristor tech or embraces Intel’s 3D XPoint—it’s going to be persistent.
In addition, and maybe most significantly, HPE says it has developed software program instruments that may truly use this enormous pool of reminiscence, to derive intelligence or scientific perception from enormous information units—each publish on Facebook; the whole thing of the Web; the well being information of each human on Earth; that form of factor. Check out this quote from CTO Mark Potter, who apparently thinks HPE’s tech can save humankind: “We consider Memory-Driven Computing is the answer to maneuver the expertise trade ahead in a method that may allow developments throughout all points of society. The structure we’ve unveiled might be utilized to each computing class—from clever edge gadgets to supercomputers.”
In follow I feel we’re a way from realising Potter’s dream, however HPE’s tech is actually a superb first step in direction of exascale. If we examine HPE’s efforts to the three fundamental points I outlined above, you’d in all probability award a rating of about 1.5: they’ve made inroads on software program, energy consumption, and scaling, however there is a lengthy solution to go, particularly in relation to computational grunt.
After the US authorities banned the export of Intel, Nvidia, and AMD chips to China, China’s nationwide chip design centre created a 256-core RISC chip particularly for supercomputing. All that HPE can supply is the Gen-Z protocol for chip-to-chip communications, and hope that a logic chip maker steps ahead. Still, that is simply the primary stage of funding, the place HPE is anticipated to analysis and develop core applied sciences that can assist the US attain exascale; provided that it will get to section two and three will HPE must design after which construct an exascale machine.
Assuming the entire different elements fall into place, Intel’s newest 72-core/288-thread Xeon Phi would possibly simply be sufficient for the US to get there earlier than China—however with an RRP of $6,400, and roughly 300,000 chips required to hit 1 exaflops, it will not be low-cost.
Most of the DoE’s exascale funding has to this point been on software program. Just earlier than this story printed, we learnt that the DoE can also be asserting funding for AMD, Cray, IBM, Intel, and Nvidia beneath the identical PathForward programme. In complete, the DoE is handing out $258 million over three years, with the funding recipients additionally committing to spend not less than $172 million of their very own funds over the identical interval. What we do not but know is what these corporations are doing with that funding; hopefully we’ll discover out extra quickly.
Now examine how low-cost RAM adjustments computing…
This publish originated on Ars Technica UK
Listing picture by HPE