MLPerf Results Validate CPUs for Deep Learning Training

Authors

  • Vice President, Core and Visual Computing Group, and General Manager, Machine Learning and Translation - Wei Li

BUILT IN - ARTICLE INTRO SECOND COMPONENT

I have worked on optimizing and benchmarking computer performance for more than two decades, on platforms ranging from supercomputers and database servers to mobile devices. It is always fun to highlight performance results for the product you are building and compare them with others in the industry. SPEC*, LINPACK*, and TPC* have become familiar names to many of us. Now, MLPerf* is filling in the void of benchmarking for Machine Learning.

I am excited to see the Intel® Xeon® Scalable processor MLPerf results submitted by our team because we work on both the user side and the computer system development side of deep learning. These results show that Intel® Xeon® Scalable processors have surpassed a performance threshold where they can be an effective option for data scientists looking to run multiple workloads on their infrastructure without investing in dedicated hardware.1 2 3

Back in 2015, I had a team working on mobile devices. We had to hire testers to manually play mobile games. It was fun initially for the testers, then it became boring and costly. One tester we hired quit on the same day. Our team created a robot to test mobile games and adopted deep learning. Our game testing robot played games automatically and found more bugs than human testers. We wanted to train neural networks on the machines we already had in the lab, but they were not fast enough. I had to allocate budget for the team to buy a GPU, an older version than the MLPerf reference GPU.4

Today CPUs are capable of deep learning training as well as inference. Our MLPerf Intel® Xeon® Scalable processor results compare well with the MLPerf reference GPU4 on a variety of MLPerf deep learning training workloads.1 2 3 For example, the single-system two-socket Intel® Xeon® Scalable processor results submitted by Intel achieved a score of 0.85 on the MLPerf Image Classification benchmark (Resnet-50)1; 1.6 on the Recommendation benchmark (Neural Collaborative Filtering NCF)2; and 6.3 on Reinforcement Learning benchmark (mini GO).3 In all these scores, 1.0 is defined as the score of the reference implementation on the reference GPU.4 For all the preceding results, we use FP32, the common numerical precision used in today’s market. From these MLPerf results, we can see that our game testing robot could easily train on Intel® Xeon® Scalable processors today.

The deep learning and machine learning world continues to evolve from image processing using Convolutional Neural Networks (CNN) and natural language processing using Recurrent Neural Networks (RNN) to recommendation systems using MLP layers and general matrix multiply, reinforcement learning (mixing CNN and simulation) and hybrid models mixing deep learning and classical machine learning. A general purpose CPU is very adaptable to this dynamically changing environment, in addition to running existing non-DL workloads.

Enterprises have adopted CPUs for deep learning training. For example, today, Datatonic* published a blog showing up to 11x cost and 57 percent performance improvement when running a neural network recommender system used in production by a top-5 UK retailer on a Google Cloud* VM powered by Intel® Xeon® Scalable processors.5 CPUs can also accommodate the large memory models required in many domains. The pharmaceutical company Novartis used Intel® Xeon® Scalable processors to accelerate training for a multiscale convolutional neural network (M-CNN) for 10,000 high-content cellular microscopic images, which are much larger in size than the typical ImageNet* images, reducing time to train from 11 hours to 31 minutes.6

High performance computing (HPC) customers use Intel® Xeon® processors for distributed training, as showcased at Supercomputing 2018. For instance, GENCI/CINES/INRIA trained a plant classification model for 300K species on a 1.5TByte dataset of 12 million images using 128 2S Intel® Xeon® processor-based systems.7 DELL EMC* and SURFSara used Intel® Xeon® processors to reduce training time to 11 minutes for a DenseNet-121 model.8 CERN* showcased distributed training using 128 nodes of the TACC Stampede 2 cluster (Intel® Xeon® Platinum 8160 processor, Intel® OPA) with a 3D Generative Adversarial Network (3D GAN) achieving 94% scaling efficiency.9 Additional examples can be found at https://software.intel.com/en-us/articles/intel-processors-for-deep-learning-training.

CPU hardware and software performance for deep learning has increased by a few orders of magnitude in the past few years. Training that used to take days or even weeks can now be done in hours or even minutes. This level of performance improvement was achieved through a combination of hardware and software. For example, current-generation Intel® Xeon® Scalable processors added both the Intel® Advanced Vector Extensions 512 (Intel® AVX-512) instruction set (longer vector extensions) to allow a large number of operations to be done in parallel, and with a larger number of cores, essentially becoming a mini-supercomputer. The next-generation Intel® Xeon® Scalable processor adds Intel® Deep Learning Boost (Intel® DL Boost): higher throughput, lower numerical precision instructions to boost deep learning inference. On the software side, the performance difference between the baseline open source deep learning software, and the Intel-optimized software can be up to 275X10 on the same Intel® Xeon® Scalable processor (as illustrated in a demo I showed at the Intel Architecture Day forum yesterday).

Over the past few years, Intel has worked with DL framework developers to optimize many popular open source frameworks such as TensorFlow*, Caffe*, MXNet*, PyTorch*/Caffe2*, PaddlePaddle* and Chainer*, for Intel® processors. Intel has also designed a framework, BigDL for SPARK*, and the Intel® Deep Learning Deployment Toolkit (DLDT) for inference. Since the core computation is linear algebra, we have created a new math library, Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN), specifically for deep learning, based on many years of experience with the Intel® Math Kernel Library (MKL) for high performance computing (HPC). The integration of Intel MKL-DNN into the frameworks, and the additional optimizations contributed to the frameworks to fully utilize the underlying hardware capabilities, are the key reason for the huge software performance improvement.

I’ve often been asked whether CPUs are faster or slower than accelerators. Of course, accelerators have certain advantages. For a specific domain, if an accelerator is not generally faster than a CPU, then it is not much of an accelerator. Even so, given the increasing variety of deep learning workloads, in some cases, a CPU may be as fast or faster while retaining that flexibility that is core to the CPU value proposition. Thus, the more pertinent question is whether CPUs can run deep learning well enough to be an effective option for customers that don’t wish to invest in accelerators. These initial MLPerf results1 2 3, as well as our customer examples, show that CPUs can indeed be effectively used for training. Intel’s strategy is to offer both general purpose CPUs and accelerators to meet the machine learning needs of a wide range of customers.

Looking forward, we are continuing to add new AI and deep learning features to our future generations of CPUs, like Intel® Deep Learning Boost (Intel® DL Boost), plus bfloat16 for training, as well as additional software optimizations. Please stay tuned. For more information on Intel® software optimizations, see ai.intel.com/framework-optimizations. For more information on Intel® Xeon® Scalable processors, see intel.it/xeonscalable.

Informazioni su prodotti e prestazioni

1

Punteggio di 0,85 con il benchmark MLPerf Image Classification (Resnet-50) 0,85 volte rispetto al riferimento MLPerf(+) utilizzando un processore Platinum Intel® Xeon® 8180 a 2 chip. Addestramento MLPerf v0.5 divisione chiusa; sistema impiegato Intel® Optimization for Caffe* 1.1.2a con Intel® Math Kernel Library for Deep Neural Networks (Intel® MKL-DNN) v0.16. Recuperato su www.mlperf.org il 12 dicembre 2018, entry 0.5.6.1. Il nome e il logo MLPerf sono marchi commerciali. Visitare www.spec.org per ulteriori informazioni.

2

Punteggio di 1,6 con il benchmark consigliato (Neural Collaborative Filtering NCF) 1,6 volte rispetto al riferimento MLPerf(+) utilizzando un processore Platinum Intel® Xeon® 8180 a 2 chip. Addestramento MLPerf v0.5 divisione chiusa; sistema impiegato Framework BigDL 0.7.0. Recuperato su www.mlperf.org il 12 dicembre 2018, entry 0.5.9.6. Il nome e il logo MLPerf sono marchi commerciali. Visitare www.mlperf.org per ulteriori informazioni.

3

Punteggio di 6,3 con il benchmark di apprendimento rinforzato (mini GO) 6,3 volte rispetto al riferimento MLPerf(+) utilizzando un processore Platinum Intel® Xeon® 8180 a 2 chip. Addestramento MLPerf v0.5 divisione chiusa; sistema impiegato TensorFlow 1.10.1 con Intel® Math Kernel Library per libreria Deep Neural Networks (Intel® MKL-DNN) v0.14. Recuperato su www.mlperf.org il 12 dicembre 2018, entry 0.5.10.7. Il nome e il logo MLPerf sono marchi commerciali. Visitare www.mlperf.org per ulteriori informazioni.

(+) Riferimento MLPerf  (adottato da MLPerf v0.5 comunicato stampa della comunità): MLPerf Training v0.5 è una suite di benchmark per misurare la velocità di sistemi ML. Ogni benchmark MLPerf Training è definito da un set di dati e un obiettivo di qualità. MLPerf Training fornisce anche un’implementazione di riferimento per ogni benchmark che utilizza un modello specifico. La seguente tabella riassume i sette benchmark nella versione v0.5 della suite.

Benchmark

Set di dati

Obiettivo di qualità

Modello di implementazioni di riferimento

Classificazione dell’immagine

ImageNet

Classificazione del 74,90%

Resnet-50 v1.5

Rilevamento di oggetti (leggero)

COCO 2017

21,2% mAP

SSD (backbone Resnet-34)

Rilevamento di oggetti (pesante)

COCO 2017

0.377 Box min AP, 0.339 Mask min AP

Mask R-CNN

Traduzione (ricorrente)

WMT inglese-tedesco

21,8 BLEU

Machine translation neurale

Traduzione (non ricorrente)

WMT inglese-tedesco

25,0 BLEU

Trasformazione

Consigli

MovieLens-20M

0,635 HR@10

Filtraggio collaborativo neurale

Apprendimento di rinforzo

Videogame professionali

Previsione di movimenti 40,00%

Mini Go


Norme di formazione MLPerf: https://github.com/mlperf/training_policies/blob/master/training_rules.adoc

4

Sistema di riferimento MLPerf*: Configurazione della piattaforma Google Cloud : 16 vCPU, Intel Skylake o successive, 60 GB di RAM (n1­standard­16), 1 GPU NVIDIA* Tesla* P100, CUDA* 9.1 (9,0 per TensorFlow*), nvidia­docker2, Ubuntu* 16.04 LTS, Pre­emtibility: disattiva, Riavvio automatico: disattivo, Disco di avvio da 30 GB + 1 disco SSD persistente da 500 GB, docker* image: 9.1­cudnn7­runtime­ubuntu16.04 (9.0­cudnn7­devel­ubuntu16.04 per TensorFlow*).

6

Novartis: misurazioni effettuate il 25 maggio 2018. In base ad accelerazione con 8 nodi rispetto a un singolo nodo. Configurazione del nodo: CPU: processore Gold Intel® Xeon® 6148 a 2,4 GHz, 192 GB di memoria, Hyper-Threading: abilitata. Scheda di rete: Intel® Omni-Path Host Fabric Interface (Intel® OP HFI), TensorFlow: v1.7.0, Horovod: 0.12.1, OpenMPI: 3.0.0. OS: CentOS* 7.3, OpenMPU 23.0.0, Python 2.7.5. Tempo di addestramento per ottenere accuratezza fino al 99% nel modello. Fonte: https://newsroom.intel.com/news/using-deep-neural-network-acceleration-image-analysis-drug-discovery.

7

GENCI: Occigen: 3306 nodi x 2 processori Intel® Xeon® (12-14 core). Nodi di elaborazione: processore Intel® Xeon® a 2 socket con 12 core ciascuno a 2,70 GHz per un totale di 24 core per nodo, 2 thread per core, 96 GB di DDR4, Mellanox InfiniBand Fabric Interface, doppio binario. Software: Intel® MPI Library 2017 Update 4 Intel® MPI Library 2019 Anteprima tecnica OFI 1.5.0PSM2 con Multi-EP, 10 GbitEthernet, SSD locale da 200 GB, Red Hat* Enterprise Linux 6.7. Caffe*: Intel® Optimization for Caffe*: https://github.com/intel/caffe Libreria Intel® Machine Learning Scaling (Intel® MLSL): https://github.com/intel/MLSL Set di dati: Pl@ntNet: I risultati prestazionali dei set di dati interni CINES/GENCI sono basati su test effettuati in data 15/10/2018.

8

Collaborazione tra Intel, Dell e Surfsara: misurazioni effettuate in data 17/05/2018 con 256x nodi di processori Gold Intel® Xeon® 6148 a 2 socket. Nodi di elaborazione: processore Gold Intel® Xeon® 6148F a 2 socket con 20 core ciascuno a 2,40 GHz per un totale di 40 core per nodo, 2 thread per core, L1d 32K; cache L1i 32K; cache L2 1024K; cache L3 33792K, 96 GB di DDR4, Intel® Omni-Path Host Fabric Interface (Intel® OP HFI), a doppio binario. Software: Intel® MPI Library 2017 Update 4 Intel® MPI Library 2019 Anteprima tecnica OFI 1.5.0PSM2 con Multi-EP, 10 Gbit Ethernet, SSD locale da 200 GB, Red Hat* Enterprise Linux 6.7. TensorFlow* 1.6: realizzato e installato dalla sorgente: https://www.tensorflow.org/install/install_sources Modello ResNet-50: specifiche di topologia https://github.com/tensorflow/tpu/tree/master/models/official/resnet. DenseNet-121Model: specifiche di tipologia https://github.com/liuzhuang13/DenseNet. Convergenza e modello prestazionale: https://surfdrive.surf.nl/files/index.php/s/xrEFLPvo7IDRARs. Set di dati: ImageNet2012-1K: http://www.image-net.org/challenges/LSVRC/2012 /. ChexNet*: https://stanfordmlgroup.github.io/projects/chexnet/. Prestazioni misurate con: OMP_NUM_THREADS=24 HOROVOD_FUSION_THRESHOLD=134217728 export I_MPI_FABRICS=tmi, export I_MPI_TMI_PROVIDER=psm2 \ mpirun -np 512 -ppn 2 python resnet_main.py –train_batch_size 8192 –train_steps 14075 –num_intra_threads 24 –num_inter_threads 2 — mkl=True –data_dir=/scratch/04611/valeriuc/tf-1,6/tpu_rec/train –model_dir model_batch_8k_90ep –use_tpu=False –kmp_blocktime 1. https://ai.intel.com/diagnosing-lung-disease-using-deep-learning/.

9

CERN: Misurazioni effettuate il 17/05/18 su Stampede2/TACC: https://portal.tacc.utexas.edu/user-guides/stampede2. Nodi di elaborazione: processore Platinum Intel® Xeon® 8160 a 2 socket con 24 core ciascuno a 2,10 GHz per un totale di 48 core per nodo, 2 thread per core, L1d 32K; cache L1i 32K; cache L2 1024K; cache L3 33792K, 96 GB di DDR4, Intel® Omni-Path Host Fabric Interface (Intel® OP HFI), a doppio binario. Software: Intel® MPI Library 2017 Update 4 Intel® MPI Library 2019 Anteprima tecnica OFI 1.5.0PSM2 con Multi-EP, 10 Gbit Ethernet, SSD locale da 200 GB, Red Hat* Enterprise Linux 6.7. TensorFlow* 1.6: realizzato e integrato dalla sorgente: https://www.tensorflow.org/install/install_sources Modello: CERN* 3D GANS https://github.com/sara-nl/3Dgan/tree/tf Set di dati: CERN* 3D GANS https://github.com/sara-nl/3Dgan/tree/tf Prestazioni misurate su 256 con 256 nodi con: OMP_NUM_THREADS=24 HOROVOD_FUSION_THRESHOLD=134217728 export I_MPI_FABRICS=tmi, export I_MPI_TMI_PROVIDER=psm2 \ mpirun -np 512 -ppn 2 python resnet_main.py –train_batch_size 8 \ –num_intra_threads 24 –num_inter_threads 2 –mkl=True \ –data_dir=/path/to/gans_script.py –kmp_blocktime 1. https://www.rdmag.com/article/2018/11/imagining-unthinkable-simulations-without-classical-monte-carlo.

10

Miglioramento delle prestazioni di throughput con l’inferenza di 275 volte con Intel® Optimization for Caffe* rispetto a BVLC-Caffe*: Misurazioni Intel in data 11/12/2018. CPU processore Platinum Intel® Xeon® 8180 a 2 socket a 2,50 GHz (28 core), HT attivata, Turbo disattivata, 192 GB di memoria totale (12 slot * 16 GB, Micron 2666 MHz), SSD Intel® SSDSC2KF5, Ubuntu 16.04 Kernel 4.15.0-42.generic; BIOS: SE5C620.86B.00.01.0009.101920170742 (microcodice: 0x0200004d); Topologia: Resnet-50 riferimento: FP32, BVLC Caffe* (https://github.com/BVLC/caffe.git) commit 99bd99795dcdf0b1d3086a8d67ab1782a8a08383 Prestazioni attuali: INT8, Intel® Optimizations for Caffe* (https://github.com/Intel/caffe.git) commit: Caffe* commit: e94b3ff41012668ac77afea7eda89f07fa360adf, MKLDNN commit: 4e333787e0d66a1dca1218e99a891d493dbc8ef1.

Il software e i carichi di lavoro usati nei test delle prestazioni potrebbero essere stati ottimizzati a livello prestazionale solo sui microprocessori Intel. I test delle prestazioni, come SYSmark* e MobileMark*, sono calcolati utilizzando specifici sistemi informatici, componenti, software, operazioni e funzioni. Qualunque variazione in uno di questi fattori può comportare risultati diversi. Consultare altre fonti di informazione e altri test delle prestazioni per una valutazione completa dei prodotti che si desidera acquistare, comprese le prestazioni di tali prodotti se abbinati ad altri prodotti. Per ulteriori informazioni, visitare il sito Web all'indirizzo www.intel.it/benchmarks.

Avviso sull'ottimizzazione: i compilatori Intel possono o meno garantire lo stesso livello di ottimizzazione per i microprocessori non Intel per quanto riguarda ottimizzazioni non esclusive dei soli microprocessori Intel®. Queste ottimizzazioni includono i set di istruzioni SSE2, SSSE3 e SSE3 e altre ottimizzazioni. Intel non garantisce la disponibilità, la funzionalità o l'efficacia di qualsiasi ottimizzazione su microprocessori non prodotti da Intel. Le ottimizzazioni dipendenti dai microprocessori in questo prodotto sono pensate per l'uso con microprocessori Intel. Alcune ottimizzazioni non specifiche per la microarchitettura Intel sono riservate ai microprocessori Intel. Fare riferimento alle Guide dell'utente e alle Guide di riferimento dei prodotti applicabili per ulteriori informazioni sugli specifici set di istruzioni a cui si riferisce questo avviso.

I risultati prestazionali potrebbero non riflettere tutti gli aggiornamenti di sicurezza pubblicamente disponibili. Per i dettagli, consultare le informazioni sulla configurazione. Nessun prodotto è totalmente sicuro.

Intel, il logo Intel, i processori scalabili Xeon e Deep Learning Boost sono marchi di Intel Corporation o di società controllate da Intel negli Stati Uniti e/o in altri Paesi.
*Altri marchi e altre denominazioni potrebbero essere rivendicati da terzi.
© Intel Corporation.