White Papers

ManualsBrandsDell ManualsConverged InfrastructureHigh Performance Computing Solution Resources

PowerEdge Servers and Deep Learning Domains:

The Impact of Scaling Accelerators

Evaluating Performance Using MLPerf Benchmark

To accurately harvest Artificial Intelligence (AI) performance data it is

critical to select a benchmark that is qualified to accurately test multiple

domain types. MLPerf is a new and broad Machine Learning (ML) and

Deep Learning (DL) benchmark suite that is gaining popularity and

adoption for its multi-domain capabilities and representative models. The

current version (v0.5) covers five domains associated with AI subsets, as

seen in Figure 1: image classification, object detection, language

translation, reinforcement learning and recommendation.

For each given domain, MLPerf will measure performance by assessing

and comparing total training times; the amount of time that it takes to train

a neural net model for a given domain to reach target accuracy. Dell EMC

team benchmarked various PowerEdge servers that have GPU

compatibility to help customers pick the appropriate GPU infrastructure

that will achieve their requirements. We used multi-GPU training to

highlight the shortest amount of training time needed to reach target

accuracy the fastest for the MLPerf.

Tech Note by

Matt Ogle

Ramesh Radhakrishnan

Summary

With deep learning

principles becoming a

widely accepted practice,

customers are keen to

understand how to select

the most optimized server,

based on GPU count, to

accommodate varying

machine and deep learning

workloads.

This tech note delivers test

results that portray how

scaling NVIDIA GPU’s on

PowerEdge server

configurations will impact

performance for various

deep learning domains, and

how these results outline

general guidelines for

constructing an optimized

deep learning platform.

Direct from Development

Server and Infrastructure

Engineering

Figure 1: Domains covered within the MLPerf v0.5 benchmark

Summary of content (3 pages)

PAGE 1
Direct from Development Server and Infrastructure Engineering PowerEdge Servers and Deep Learning Domains: The Impact of Scaling Accelerators Tech Note by Matt Ogle Ramesh Radhakrishnan Summary With deep learning principles becoming a widely accepted practice, customers are keen to understand how to select the most optimized server, based on GPU count, to accommodate varying machine and deep learning workloads.
PAGE 2
Server DSS 8440 PE T640 PE R740 Precision 5820 # of CPU's 2 2 2 1 # of GPU's 8 4 3 2 GPU Type V100 (16GB) V100 (32GB) V100 (16GB) GV100 (32GB) GPU Interconnect PCIe PCIe PCIe PCIe Figure 2: PowerEdge CPU & GPU details for each tested configuration Every benchmark ran on single node PowerEdge servers, as seen in Figure 2. Each server was loaded with either 2, 3, 4 or 8 Tesla V100 PCIe GPU’s, and these configurations ran until the unique domain being tested reached the target accuracy.
PAGE 3
Recommendations 1. The training times and scaling behavior vary between different domains and models o Using superior accelerators would be advantageous for the domains that require the most time o In order to pick the appropriate server and number of GPU’s, it is useful to understand the models and domains being used. 2.