Supercomputer Summit will allow the United States to return to the first line of the Top 500 rating
China is gradually taking a dominant position in the use of supercomputers. Most systems from the Top 500 ranking are located in this country. Also the most productive supercomputer called Sunway TaihuLight is in China. It has a capacity of 93 petaflops, which is 3 times the capacity of the Tianhe-2 system, which is on the second line of the rating and is also located in China.
But the US is developing a new supercomputer called Summit, which will allow the country to return to the first line of the Top 500 ranking. Its productivity more than 2 times exceeds the capabilities of Sunway TaihuLight and will amount to more than 200 petaflops. The performance of the new product will be 11 times higher than that of the predecessor Titan, but it will require a much smaller area for the system. The Titan used 18688 processor blocks, while Summit would require only about 4,600 units, each of which will provide about 40 teraflops performance. The new supercomputer will use IBM and NVIDIA hardware, supplemented by a liquid cooling system. Each unit will receive a server system IBM Power Systems AC922 with two IBM POWER9 processors and six high-performance NVIDIA Volta GV100 accelerators.
Each IBM POWER9 processor supports 8-channel RAM, so each unit will be able to work with 16 channels of RAM with a bandwidth of 340 GB / s. Each chip contains 8 billion transistors, up to 24 computing cores, capable of processing up to 96 instruction flows simultaneously (thanks to the SMT4 function). Detailed characteristics of the GPU NVIDIA Volta GV100 can be found in one of the previous materials.
Each processor unit is supplemented with associated DDR4 and HBM2 (High Bandwidth Memory) 512 GB and additional nonvolatile RAM with a capacity of 1600 GB (maximum is supported up to 2000 GB of such memory). These 1600 GB of memory serve as a packet buffer. Here, large data sets are accumulated before they are transferred to the main memory node.
To transfer data between the computational units, 96 PCIe 4.0 lines and the Mellanox EDR InfiniBand two-port adapter are used. It provides a throughput of 392 GB / s, which is only 8 GB / s less than the theoretical maximum of 400 GB / s. The NVIDIA NVLink 2.0 native interface provides 100 GB / s bandwidth on the CPU-to-GPU and GPU-to-GPU connections.
The power consumption of the Summit supercomputer will be 15 MW, which is comparable to the appetite of the current Chinese leader. But it should be taken into account that Summit's performance is more than 2 times higher than that of Sunway TaihuLight (200 petaflops against 93 petaflops). If to compare with Titan (energy consumption 9 MW), then the increase in energy consumption at 35% leads to an 11-fold increase in productivity. This indicates a higher efficiency of the new solution. The site where Summit will be located, is capable of maintaining the operation of systems with power consumption of up to 20 MW.
Among other features of the supercomputer Summit can be called the following. To accommodate all computing power and the necessary maintenance infrastructure, a space roughly equivalent to two basketball courts is required (the size of the playground is 28 meters in length and 15 meters in width from the inner edge of the confining lines). For the computer complex, about 220 km of cables will be required. Implementation Summit is scheduled for 2018.
Source: techpowerup