In the traditional server chip market, Intel is a giant, whether it is IBM Power or a ARM camp, the share of the share is inexpensive. But the front of the front of the artificial intelligence, IBM seems to be more advantageous.
According to foreign media reports, recently IBM and NVIDIA teamed up with new server IBM Power Systems S822LC for High Performance Computing (there are two products for IBM Power Systems S821LC and IBM Power Systems S822lc for Big Data), from this string name It is not an ordinary server, which is specially introduced for artificial intelligence, machine learning, and advanced analysis application scenarios.
The IBM official claims that the digital data processing speed is 5 times faster than other platforms, and 80% of the average performance per dollar compared to Intel X86 servers.
Where is this server stronger than Intel X86?
It is understood that the server uses two IBM Power8 CPUs and 4 NVIDIA Tesla P100 GPUs. Power8 is currently the strongest CPU of IBM. From previous media evaluation data, its performance is better than Intel E7 V3, and Tesla P100 is a high-performance computing (HPC) graphics card that NVIDIA released this year, such a configuration combination It is naturally not weak in processing performance.
There are two reasons:
First, compared to the CISC instruction set, the POWER processor used by the RISC instruction set can perform multiple instructions at the same time, and can divide a command into multiple processes or threads, and is executed simultaneously with multiple processors, so parallel processing performance. It is better than the Intel X86 chip based on the CISC architecture.
In addition, this server is ingenious in the "cooperation" between Power8 and Tesla P100.
Another major feature of the Power architecture is to give full play to the advantage of GPU performance.
In fact, Tesla P100 has two versions, one is NVIDIA launched NVLink version in April this year, and the other is the PCI-E version released in June. In simple, the former is the enhancement version of the latter, paired with IBM Power8 It is Tesla P100 NVLink version.
The Tesla P100 uses the Pascal architecture that enables page migration between the CPU and the GPU, but 4 each NVLink version is also configured with 40 GB NVIDIA NVLINK ports per second, and the division enters the GPU cluster. NVLINK is the unique high-speed interconnect technology of OpenPower Foundation. Its effective bandwidth up to 40GB / s, which is an upgraded version of PCIe, enough to meet the multi-chip parallel computing needs. However, the number of CPU refers to this standard is supported, and Power8 is one of them (Intel is not here).
This means that Power8 CPUs can complete communication with Tesla P100 GPUs, which allows the connection speed between the CPU and GPUs in IBM Power Systems S822LC for high Performance Computing. The connection speed is far from ordinary in PCIe Exchange data on the bus.
IBM said, "This feature means that the database application, high-performance analysis application and high performance computing application run can be much more data set on the X86 system in the GPU in the PCI-E interface. Run. "
In addition, the semi-precision floating point calculation performance of Tesla P100 reached 21 trillion per second - approximately 14% higher than the GPU inserted into a modern PCI-E slot, so that the importance of training depth neural network is not Combo.
IBM also made a longitudinal contrast, compared with the Tesla K80 GPU accelerator of the old Power S822LC server, the acceleration capacity of the new server has increased more than twice.
It is expected that IBM Power9 in the world will continue to optimize the CPU GPU group.
Why is it "CPU GPU"?
It is well known that CPUs have long been unbearable on computational tasks in artificial intelligence and intelligence and deep learning. Therefore, many companies have introduced artificial intelligent chip concepts, such as Google's TPU (Tensor Processing Unit); there is also an industry in providing FPGA more suitable for depth learning algorithms, which is the main reason for the acquisition of Altera at high prices.
However, the above two alternative CPUs are not mature, and most companies are still a combination of "CPU GPU", or called heterogeneous servers. THE INVEN, in this isomerism mode, the serial portion of the application runs on the CPU, and the GPU is used as a coprocessor, which is mainly responsible for calculating the heavy task.
Because of the CPU, the advantage of GPU is very obvious:
1. CPU is mainly optimized for serial instructions, while GPUs are optimized for large-scale parallel operations. Therefore, the latter is faster in large-scale parallel operations;
2. Under the same area, there is more operational units (integer, floating-point multiplying unit, special arithmetic unit, etc.);
3. In general, GPU has a larger bandwidth Memory, so there is also good performance in the application of large throughput.
4.GPU's demand for energy is much lower than the CPU.
Of course, this does not mean that the artificial intelligence server does not have a demand for the CPU. The CPU is still an indispensable part of the calculation task. It also requires high performance CPU in the depth learning algorithm processing task to perform instructions and the GPU is transmitted, while playing The universality of the CPU and the complex task processing capability of the GPU can achieve the best results. The popular point is to implement the collaborative calculation of the CPU and GPU.
Although NVIDIA and Intel and other chip companies are fascinated by GPU and CPUs, it is innecting to be weak, but in fact, these companies have begun to increase research and development in heterogeneous calculations, at least in the near future, CPU and GPU combination Continue to become the most effective solution in the field of artificial intelligence.
Original link ·: https://www.eeboard.com/news/intel-75/
Search for the panel network, pay attention, daily update development board, intelligent hardware, open source hardware, activity and other information can make you master. Recommended attention!
[WeChat scanning picture can be directly paid]
Technology early know:
Foxconn United States Construction Factory employee wages stunned
Huang Zhang's backup, really can save a Meizu who has been in the "death" road?
Google acquires HTC, HTC Pixel mobile phone R & D team joined Google Team
Want! The world's first AMD Ryzen notebook is about to sell: 100,000 yuan
The price of memory is booming, DIY buyers say: can't afford it, reveal the mystery of memory prices skyrocket
Our other product:
