Summit, the United States’ latest supercomputer, has ended China’s two-year reign as the world’s fastest supercomputer. Built by none other than global computer giant IBM, the Summit skyrocketed itself to the peak of the TOP500 ranking with a Linpack score of 122.3 PetaFLOPS to beat previous Chinese record-holder Sunway TaihuLight’s result of 93 PetaFLOPS.
The $200 million Summit supercomputer, stationed at the Department of Energy’s Oak Ridge National Laboratory, was conceived to research the fields of human diseases such as Alzheimer and cancer, astrophysics, fusion energy, and climate change. Summit occupies 5,600 square feet of space and weighs over 340 tons. To put those dimensions into perspective, the gargantuan supercomputer is equivalent to the size of two tennis courts with a weight that effectively surpasses that of a large commercial airliner. Summit is connected by 185 miles of fiber-optic cables, which is the distance that separates Knoxville from Nashville, Tennessee.
The Summit is based on IBM’s AC922 Power System that’s comprised of 4608 computer nodes, interconnected to each other through dual-rail Mellanox EDR 100 Gbps InfiniBand adapters. We've taken a deeper look at the inside of each node in our recent Regaining America's Supercomputing Supremacy With The Summit Supercomputer article.
Each node in the ecosystem is good for 42 TeraFLOPS of performance and comes equipped with two 22-core IBM Power9 processors with a maximum clock rate of 4.0 GHz, 512GB of DDR4 memory, 96GB of HBM memory, 1.6TB of NVMe storage, and six Nvidia Volta-based Tesla V100 graphics processing unit accelerators. Communication between the Power9 processors and Volta GPUs is made possible by NVLink 2.0, Nvidia’s energy-efficient and high-bandwidth interconnect technology. In total, the Summit supercomputer has at is disposal 9,216 Power9 processors, over 10PB of DDR4, HBM, and non-volatile memory, 250PB of storage, and 27,648 Tesla V100 accelerators — which was certainly not a bad way to spend $200 million dollars.
Although the Summit only needed to pump out a modest 122.3 PetaFLOPS to catapult its way past China’s Sunway TaihuLight supercomputer, the American supercomputer’s actual peak performance is much higher. The Summit is designed to deliver up to 200 petaFLOPs of computing power. It draws 13 megawatts of power and relies on a liquid-cooling system that pumps 4,000 gallons of water through the system every minute to prevent it from melting down.
| Ranking | System | Country | Rmax (TFlop/s) |
|---|---|---|---|
| 1 | Summit | United States | 122,300.0 |
| 2 | Sunway TaihuLight | China | 93,014.6 |
| 3 | Sierra | United States | 71,610.0 |
| 4 | Tianhe-2A | China | 61,444.5 |
| 5 | AI Bridging Cloud Infrastructure (ABCI) | Japan | 19,880.0 |
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Zhiye Liu is a news editor, memory reviewer, and SSD tester at Tom’s Hardware. Although he loves everything that’s hardware, he has a soft spot for CPUs, GPUs, and RAM.
-
DerekA_C So when they push it to the max is is more than double the performance of China's, sweet deal to bad Intel built China's super computer shame on Intel such damn sell outs of this great nation. Let China make their own CPU's GPU's Memory and other components and leave USA products in USA stop selling out it is ridiculous. MAGA!!!!Reply -
ElectrO_90 Dereka_c you are a douche.Reply
How many products do you own that are "Made in China" probably most of them -
maniac62 Reply
From Wikipedia:21091916 said:So when they push it to the max is is more than double the performance of China's, sweet deal to bad Intel built China's super computer shame on Intel such damn sell outs of this great nation. Let China make their own CPU's GPU's Memory and other components and leave USA products in USA stop selling out it is ridiculous. MAGA!!!!
The Sunway TaihuLight uses a total of 40,960 Chinese-designed SW26010 manycore 64-bit RISC processors based on the Sunway architecture... -
bit_user Reply27,648 Tesla V100 accelerators
Which should be good for 3.46 deep learning exaflops. I'm a little surprised they're not touting that, given how much Nvidia likes to talk about the deep learning performance of the V100.
And... exaflops? Okay, it's not exactly general-purpose, but still... has anything else yet reached exaflop territory?
Okay, speaking of general-purpose, the general-purpose fp16 performance should be around 824 petaflops. -
stdragon When the next one is named "Skynet", I highly recommend we pull the plug on it ASAP! But, well, we've all seen how that movie played out.Reply -
milkod2001 ''was conceived to research the fields of human diseases such as Alzheimer and cancer, astrophysics, fusion energy, and climate change''So many supercomputers and nothing groundbreaking came from it regarding that area of research for last 20 years. I bet they use it for something 'more interesting' like processing our personal data, global surveillance and stupid rush to get the best Ai advantage.Reply -
Brian_R170 Reply21091916 said:So when they push it to the max is is more than double the performance of China's, sweet deal to bad Intel built China's super computer shame on Intel such damn sell outs of this great nation. Let China make their own CPU's GPU's Memory and other components and leave USA products in USA stop selling out it is ridiculous. MAGA!!!!
You are thinking of the Tianhe-2 which used Intel chips and topped the supercomputer list in 2014. As a result, in 2015, the US Government banned Intel, NVidia, and AMD from selling chips to China for use in supercomputers. Less than two years later, China had created the Sunway TaihuLight that is much faster and uses only Chinese chips (this is the supercomputer referenced in this article). Exactly what you suggested came to pass: China made their own chips and it was a win-win... for China. The ban did absolutely nothing to stop China's supercomputer ambitions and US companies lost out on the sale of hundreds of millions of dollars in high-tech equipment.