|IBM And Nvidia Team Up To Build The World's Fastest Computer|
There is an arms race in computing, which has implications in international relations and geopolitics as significant as any conventional arms race. This is the race to build the world's fastest supercomputer. Having the fastest supercomputer is a matter of national pride, but it also confers real economic, political, and military benefits. Supercomputers are used for everything from long-range weather forecasting to the modeling and design of nuclear weapons.
Currently, the fastest supercomputer in the world is China's Sunway TaihuLight, capable of 93 PetaFLOPS, or 93 x 10^15 floating point operations per second. In 2018, the US plans to overtake Sunway with a new supercomputer now under construction at Oak Ridge National Laboratory (ORNL) in Tennessee. This computer, called Summit, will be capable of about 200 PetaFLOPS. A sister computer, Sierra, which is similar in design to Summit, is also being built at Lawrence Livermore National Laboratory in California.
Summit consists of about 4600 “nodes”, which are basically rack-mounted servers. Although Summit will be 5-10 times more powerful than its predecessor, it will have only ¼ of the nodes and use substantially less power.
It's what's inside these nodes that makes them so special. Each node consists of a specialized HPC server designed by IBM. The node contains two IBM Power9 processors and six Nvidia (NASDAQ: NVDA) Tesla V100 SXM2 GPU accelerators. The entire box is water cooled, as shown in the hardware photo at the top of the article. The hardware was on display at the recent Super Computing 17 Conference.
What is unique about the IBM system is that each Power9 processor communicates directly with the Nvidia GPU via Nvidia's proprietary high-speed NVLink digital interface. This built-in capability is unique to Power9 and provides it with 100 GB/sec of throughput via NVLink 2.0.
The system architecture makes extensive use of NVLink both to connect the GPUs to the Power9 processor and to each other, as shown in this diagram from Tom's Hardware.
In effect, the Power9 processors coordinate the computational flow in each node, while the Tesla V100s do the heavy lifting. IBM's decision to incorporate NVLink 2 interfaces on chip shows that the company understood where things were going in high performance computing. Rather than fight the trend towards the use of GPU computing, IBM wisely chose to co-opt it.
To read the entire article, select the following URL: