Multi-core and multi-core heterogeneous systems are increasingly favored by ultra-large scale computers, such as the recently released global high-performance computing Top500 Ranking first IBM Of Roadrunner Is to put your own Cell And AMD The haolong processors are mixed together. On the chip layer, including Intel, AMD And NVIDIA Are Developing heterogeneous multi-core or multi-core chips. For example AMD Proposed CPU + GPU Of Fusion , NVIDIA Of Tesla .
As we all know, the vast majority of current high-performance computing systems and applications are X86 With the continuous improvement in the performance of high-performance computing systems, multi-core and multi-core heterogeneous systems provide excellent performance when meeting certain special applications of high-performance computing. For example, for applications that pursue the performance of floating-point operations, GPU ( Graphics Processing Chip ) The speed is much higher than the traditional CPU ( Central processor ) Which is why AMD And NVIDIA The company believes that in the future of high-performance computing GPU Or CPU + GPU Will eventually replace CPU . However, from the current programming model of application software, most programming tools and applications are still based on X86 Architecture, so we need to make full use GPU Or CPU + GPU To change the traditional programming mode. In this regard, AMD Liu wenzhuo, Senior Technical Manager of the company in China CPU And GPU According to the keynote speech, the current programming model is the biggest challenge facing multi-core heterogeneous systems. So is multi-core isomorphic computing the development direction of High Performance Computing?
Intel High Performance Computing Software Manager Sanjaygoil The doctor told China Electronics daily that Intel will certainly make heterogeneous multi-core chips, " Heterogeneous kernels and heterogeneous Programming Environments " Different, Intel will stick to it " Heterogeneous kernels and the same programming environment " That is to say, no matter which kernel is used, the programming mode should be the same. Although there are already some heterogeneous systems in the industry Linpack The test has outstanding performance and is difficult to be tested in practical applications. Therefore, powerful software tools are required to reduce the difficulty of software programming in heterogeneous environments.
Dr. he Wanqing, Technical Manager of Intel China's Asia Pacific Customer Response Team, said that Intel's high-performance computing development tool can be used to analyze and optimize existing MPI Programming mode performance. At the same time, Intel has released its own Mpi2.0 , The interface remains unchanged, the programming mode remains unchanged, but it is more suitable for multi-core environments and can be better bound between processes and the kernel. In addition, Intel Openmpi In cooperation with industry partners.
From the interviews with participating vendors and the current status of high-performance computing, multi-core isomorphic architecture based on the same programming model should be the most practical path for high-performance computing systems in the future. And adoptCPU + GPUHeterogeneous systems, it is very difficult to change the industry's popular programming model. The future is full of uncertainty.
David Kirk, chief scientist at NVIDIA, said heterogeneous computing architecture is indeed the most effective solution for processing various types of data, however, he does not think that the knights corner accelerator developed by Intel can build a practical heterogeneous multi-core platform. The latter is a product launched by Intel for HPC applications.
Currently, GPU or special computing accelerators are used for data processing in many application fields, such as oil and natural gas exploration, earthquake and financial services, such as amd firestream and NVIDIA Tesla. As the number one processor manufacturer in the world, although Intel has encountered setbacks in the development of its own graphics chip product Larrabee, however, the company is still developing a chip product codenamed knights corner, which will integrate the multi-core architecture and x86. Because the current operating system needs to be supported by the x86 processor during operation, therefore, David Kirk believes that these products will not disappear.
The intel HPC platform will be equipped with a Knights corner chip consisting of an independent CPU and a multi-core HPC accelerator based on a PCI Express slot. Many HPC experts believe that the PCI Express bus will become the bottleneck of such accelerators, so the PCI Express bus bandwidth is too low. The solution to this problem is to launch a chip product that integrates the x86 core and a large number of parallel graphics cores. In this regard, AMD's solution is fusion. David Kirk thinks that PCI Express may not be a bottleneck.
