Tesla K20 Encounter power supply bottleneck server difficult to surrender Barroth

Recently, Dell said it found it difficult to gpgpu the rack server because of power constraints.

Dell offers a variety of PowerEdge servers that support GPGPU acceleration, as well as a repository for optimal performance. However, Dell later found that many users are reflected in the 42U rack deployment of the PowerEdge c8220x server, due to power limitations of the problem, can not install two http://www.aliyun.com/zixun/aggregation/10867.html " >nvidia GPGPU Accelerator.

The PowerEdge c8220x Modular server supports CPU, GPU, and storage modularity, and can also expand modules that support up to 2 Tesla K20 Gpgpu accelerators. Although this design is highly computationally dense, it cannot support GPGPU because it does not provide sufficient power support.

For this reason, Dell has also done research to address the high energy consumption of the Nvidia GPGPU Accelerator, such as the impact on overall system performance when reduced to 225 watts under the default power consumption. The study found that in the case of 200 watt power, GPGPU can provide 4% performance per watt, and in the case of overclocking Gpgpu can not continue to improve performance per watt.

Dell also said that the first task of Nvidia Tesla K20 should be to find solutions to reduce the overall power consumption of the system. You also need to set up the CPU in the BIOS and modify the priority of the Tesla K20 through its software.

While Dell acknowledges that the Nvidia Tesla K20 GPGPU Accelerator delivers better per-watt performance than CPUs, users need to focus on overall power consumption. In addition, Dell research reports that if you want to further optimize per watt performance, users should adjust the power consumption of the NVIDIA TESLA K20 plate.