Author of this article: Wu Mingxin, Deputy Secretary-General of China Simulation Society
In recent years, cutting-edge technologies such as artificial intelligence, digital twins, quantum technology, Internet of Things, network computing, digital design and simulation, augmented reality/virtual reality/mixed reality (AR/VR/MR) are constantly developing and widely used in film and television entertainment , education, design, medical care, aerospace, industry and many other fields, providing important support for the digital economy and digital transformation of enterprises. Metaverse technology with simulation as its soul and computing power as its engine is pushing the Internet into the next year. The rapid development of Metaverse has a strong path dependence on simulation computing. Therefore, in the field of simulation computing, the pursuit of computing power is endless.
AMD and Intel have been in love with each other in the X86 server market for many years. Can EPYC with Su Ma's blessing really achieve corner overtaking? This time we were lucky to get a server equipped with a dual-socket AMD third-generation EPYC (Python) 7763 processor. Let us see how this processor performs in the field of simulation computing.
server configuration:
First introduce the server configuration used in this test.
For AMD, we got the AMD EPYC third-generation 7763 processor, which has 64 physical cores, 256MB L3 cache, main frequency 2.45GHz, maximum overclocking frequency 3.5GHz, and 256 threads can be realized on two-way servers!
For Intel, we use Xeon Platinum 8358 , which is the platinum series in Ice lake. It has 32 physical cores, the main frequency is 2.6GHz, and the turbo frequency is 3.3GHz.
For the sake of fairness, we created a 16-core virtual machine on two servers and tested them in the virtual machine. At the same time, in order to test the high core concurrent performance of AMD processors, we created a 58-core virtual machine on the AMD server alone. Core virtual machine for testing. The processor and memory ratio of the test virtual machine is uniformly 1:8.
Test 1: 3D reconstruction test
test environment:
This test uses colmap+OpenMVS to form a 3D reconstruction pipeline, uses colmap to calculate the pose of the image, and then sends it to OpenMVS for model reconstruction.
COLMAP is a generic structurefrom-Motion, SfM and Multi-View Stereo, MVS pipeline with graphical and command-line interfaces. It provides a wide range of features for the reconstruction of ordered and unordered image collections.
OpenMVS is a relatively classic MVS (Multi-View Stereo) open source library, which integrates the entire technical solution of 3D reconstruction (camera model, multi-view stereo geometry, dense reconstruction, surface reconstruction, point cloud fusion, texture mapping). We use very classic and effective algorithms in each reconstruction process, such as SGM and PatchMatch, which are very classic depth map reconstruction algorithms for 3D reconstruction.
The test results are as follows:
The comparison in the above figure is the running time. The shorter the time, the better. It can be seen that under the same number of cores, AMD processors have certain advantages over Intel in 3D reconstruction. When we are in a 58-core virtual machine When testing, the performance improvement is basically linear, which meets our expectations.
Test 2: Colmap (pure CPU calculation)
On the basis of the previous test, test the feature extraction, feature matching and sparse reconstruction of colmap in turn (pure cpu calculation), the test results are as follows:
In high-performance computing scenarios, the higher-frequency AMD EPYC 7763 performs better than the Intel Xeon Platinum 8358.
In the feature extraction stage, the AMD processor with the same core consumes only 73% of the Intel processor; the feature matching stage takes 93%; the sparse reconstruction stage takes 91%.
After increasing the number of AMD processor cores from 16 cores to 58 cores, the time-consuming feature extraction stage is reduced by 68%, and the feature matching stage is reduced by 49%.
Test 3: Nerve Radiation Field Reconstruction
This test is due to the pytorch version engineering pipeline of a paper "NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis". The specific libraries and versions required are as follows:
The specific process of Nerf neural radiation field technology is: by using a sparse input view set to optimize the underlying continuum density function, so as to achieve the effect of synthesizing new views of complex scenes.
The specific pipeline is shown in the figure below:
The network structure used in the experiment is a fully connected network.
The data used is the official low-res-fern data, and all experiments are iterated 200,000 times.
Test Results:
Referring to the above official data, AMD and Intel servers, the running time of the same task has been reduced.
Compared with the official data, the AMD server has reduced the running time by 1 hour, 22 minutes and 41 seconds, which is about 17.2%; the Intel server has reduced the running time by 51 minutes and 55 seconds, which is about 10.8%.
From the results, the AMD server can complete iterations in a shorter time and run more efficiently.
Summarize:
From the above three test results, it can be seen that the single-core performance of the current mainstream server processors is not very different, and the AMD EPYC third-generation processor can integrate more physical cores through advanced packaging technology and process technology. Intel has a certain advantage. I hope that the two companies can continue to compete in the future and provide consumers with more favorable prices and stronger computing power.
There are various requirements for computing power in the field of simulation computing. This test can only try to get a glimpse of the leopard, and it cannot fully measure the level of the processor. We hope that through this test, we can have a more comprehensive understanding of the development of server hardware and play a role in the simulation industry. The impetus obtained, and provide some references when selecting the server.