"Computing beast" Inspur nf5468a5 GPU server open trial free application

In the near future , Technology media vs. tide NF5468A5 The server has conducted a series of professional evaluations , The report shows that this GPU The server is in a typical AI Computing scenarios have superior performance than expected , stay MLPerf Training、MLPerf Inference、Alphafold2、NAMD、HPL、Stream And other typical applications have shown amazing leading performance , Therefore, it is praised by the media as “ Calculating beast ”. The official website of Inspur information shows ,NF5468A5 Underway “ Value model Limited free trial ” Activities , Users who have a strong demand for computing power can apply for free .

 

NF5468A5 It is a product launched by Inspur information for AI Training 、AI Reasoning 、HPC、 Video processing and other application scenarios GPU The server , stay 4U Space carrying 2 star AMD EPYC processor , Support up to 8 Zhang shuangkuan acceleration card , The ingenious partition heat dissipation design effectively realizes CPU And GPU Diversion of modules , At the same time through PCIE 4.0 Direct connection effectively reduces CPU and GPU Communication delay between . The server supports up to 8T Of DDR4 Memory 、409.6 GB/s Total memory bandwidth , And it provides 8 Full height, full length and double width PCIe x16 Physical slot for . Its powerful processor performance 、 Huge memory capacity and bandwidth 、 rich IO Expand , Perfect for AI Calculation 、 Cloud computing 、HPC And the workload of various businesses of the enterprise .

Media right NF5468A5 Conducted a series of evaluations . among HPL The test results show that ,NF5468A5 carrying 2 star AMD EPYC 7543 processor , Floating point calculation speed is 2.69 TFLOPS, according to AMD Platform theoretical floating-point calculation speed , The computing efficiency of the processor reaches 93.74%. stay STREAM In the test , Due to the use of multithreading parallel , Measured result memory bandwidth 373 GB/s, Compare the theoretical bandwidth of the platform memory , The measured memory bandwidth efficiency is also amazing 91.1%.

 

NF5468A5 HPL test result

 

NF5468A5 Memory bandwidth test results

stay AI Training performance test , wave NF5468A5 collocation 8 Zhang NVIDIA A100 PCIE 40GB GPU, Use MLPerf Training V1.0 Code training convolutional neural network ResNet50, The number of pictures processed per second can reach 21486 Zhang , A single machine 35 Minutes to complete Resnet50 model training . Refer to recent issues MLPerf Training list , carrying 8 Zhang NVIDIA A100 40G GPU The best result of the card server is 36.2 minute . so to speak , In the same way GPU In the configured server , wave NF5468A5 Of ResNet50 Training performance is the best .

 

ResNet50 Training test results

stay AI Reasoning performance test , carrying 1 Zhang NVIDIA Tesla T4 GPU Of NF5468A5, Use MLPerf Inference V1.0 Code ,ResNet50 The test result is processed per second 5671.9 A picture , This achievement is also very excellent . meanwhile ,NF5468A5 It can well support the Cambrian MLU270-S4 Reasoning accelerator card ,Caffe Under the framework of ResNet18 The computing performance exceeds 7000 A picture .

 

ResNet50 Reasoning test results

meanwhile , The media also developed a special accelerator for Inspur information M10A Performance tests were carried out , It turns out that , wave NF5468A5 collocation 1 Zhang M10A, Can be realized 480fps 1080P Smooth transcoding of video , a sheet M10A The video processing capacity of is equivalent to the performance of a two-way server . Besides ,NF5468A5 carrying 1 Zhang RTX3090 The graphics card ,ETHASH Algorithm performance breakthrough 100MH/s.

 

M10A Video transcoding performance test results

wave NF5468A5+ Single card RTX3090 HASH Algorithm test results

 

Algorithm	ETHASH	ETCHASH	AUTOLYKOS2	BLAKE3	MTP	MTP-TCR	OCTOPUS
performance	108MH/s	108MH/s	232MH/s	2.44GH/s	7.23MH/s	28.78MH/s	103.07MH/s
Algorithm	KAWPOW	PROGPOW	PROGPOW-VEIL	PROGPOW-VERIBLOCK	PROGPOWZ	FIROPOW	/
performance	55MH/s	54.4MH/s	54.85MH/s	27.31MH/s	54.37MH/s	54.91MH/s	/

 

 

 

NF5468A5 stay HPC It also has excellent performance in application performance . Media in NF5468A5 The platform is equipped with 2 star AMD Milan-X 7773X Run common Meteorological Applications WRF And Computational Fluid Dynamics Applications OpenFOAM Conduct performance benchmarking . Test data shows ,WRF Test its performance and compare it with two on the same platform Rome 7742 The computing performance of the processor is improved 23%~34%; And in the OpenFOAM In the test , Its performance is compared with that of the same platform Rome 7742 Processor computing performance improved 34%~80%.

 

WRF In different AMD Performance comparison on processors

 

OpenFOAM motorbike Examples are different AMD Performance comparison on processors

In the latest evaluation , The media is also right NF5468A5 The server AI+Science The performance of the application scenario has been comprehensively evaluated . The test selected two recent hot applications AlphaFold2 and NAMD. The evaluation results show that , For the length in 1000 Within the protein sequence , The complete time of structure prediction is basically less than half an hour , It means a NF5468A5 The server can complete at least 384 individual Alphafold2 Protein sequence prediction task ; For molecular dynamics simulation ,STMV Example in NF5468A5 Can be realized on 90.6ns/day The calculation speed of , One server can be realized in one day 100 Ten thousand atoms are close 100ns The simulation . wave NF5468A5 GPU The server can meet the needs of most scientific research teams in AlphaFold2、NAMD And other fields of scientific application AI Accelerating computing needs .

NF5468A5+ Single sheet A100 Predicted AlphaFold2 top1 Model calculation performance

 

 

NAMD stay NF5468A5 Test results of the platform

Through many different configurations 、 In depth evaluation of different scenes , The media believes that the tide NF5468A5 It is a powerful 、 It has a wide range of application scenarios GPU The server . The hardware design of the server is reasonable , Maximize the performance advantages of core components , And ensure the stable operation of the server through the partition cooling design . meanwhile ,NF5468A5 Widely compatible with mainstream accelerator cards , With a more flexible computing architecture, users can meet the needs of image recognition to the greatest extent 、 natural language processing 、 Voice recognition and other multi scenario application requirements .

at present , According to the official website of Inspur NF5468A5 Is launching “ Value model Limited free trial ” Activities , Interested users may wish to apply , Try it out . Click on “ Read the original ” You can sign up for .