디지털솔루션

DAEBO Communication & Systems

Deep Learning and AI Infinite possibilities

Strengths of Deep Learning

Deep learning, which allows a computer to understand an infinite amount of data (including an image, voice, text, etc.), is showing the fastest growth in the field of AI. Various industry-leading companies are adopting deep learning in order to process exponentially increasing data, using machine learning algorithms and computer H/W. In this way, they are leveraging such enormous amount of data in developing product, services, or procedures, and gaining a competitive edge successfully.

nvidia

Application of
deep learning

GPU is turning a new business idea into a reality in various industries, including a smart city, the public sector, financing, manufacturing, retail business, medical, service, and so on. And NVIDIA's GPU for enterprise use serves as the brain of an AI-based business.

Applied fields

  • Automobiles
    Self-driving
    Pedestrian detection
    Lane tracking
    Vehicle detection
  • Financial
    services
    Fraud and error detection
    Risk analysis
    Transaction algorithm
  • Government/
    National defense
    Cyber security
    Facial recognition
    Video monitoring
    Graph analysis
  • Healthcare
    Cancer cell detection
    Disease identification
    New drug development
  • Educational/
    R&D
    Institutions
    Image analysis
    AI R&D
    Voice processing
  • AI start-up
    Video search
    Voice recognition
    Emotion analysis
    Recommendation/
    suggestion

DEVELOPMENT TRAINING

DEPLOYMENT INFERENCE

  • DESKTOP
  • CLOUD
  • DATA CENTER
  • SELF-DRIVING CARS
  • INTELLIGENT MACHINES

Creating value
for our customers

NVIDIA DGX and TESLA provide a data center with the most powerful deep learning platform and performance for high computing workload, allowing data scientists to explore AI across a desk workstation, data center, and cloud.

DGX deep learning stack

DGX deep learning stack

Deep learning framework
dgx
Deep learning user S/W
NVIDIA DIGITS™
Acceleration solution
dgx
Deep learning library
NVIDIA cuDNN and NCCL
Containerization tool
NVDocker
GPU driver
NVIDIA GPU Compute Driver Software
System
GPU-Optimized Linux Server OS

Deep learning solution line-up

DGX-2

Super computer system for the most powerful deep learning scaling NVIDIA DGX-2, the first two-petaflops system using sixteen GPUs, interconnected perfectly so as to increase the performance of deep learning ten-fold, is designed to overcome any limitations in AI speed and extension. Use of DGX-2, which is based on the architecture for AI-scale employing the NVIDIA® DGX™ S/W and NVIDIA NVSwitch technology, allows you to perform even the most complex AI projects in the world.

DGX-2 with deep learning performance 10 times faster

트레이닝 시간(Days) 성능 비교

Comparison of training time (days)

Workload : FairSeq, 55 epochs to solution PyTorch training performance.
DGX-2 성능
GPU 16x Tesla V100 SXM3 System Memory 1.5TB
GPU Memory 512GB Total (NVSwitch Technology) Network 8x 100Gb / sec
Infiniband / 100GigE
2x 10 / 25Gb / sec Ethernet
TFLOPS (FP16) 2,000 TFLOPS
NVIDIA CUDA Core 81,920 Storage OS : 2x 960GB NVME SSDs
Data : 30TB (8x 3.84TB) NVME SSDs
NVIDIA Tensor Core 10,240 Software Ubuntu Linux OS, CUDA Toolkit
NNVSwithces 12 System Weight 340 lbs / 154.2 kg
Max. Power Consumption 10,000 W System Dimensions H 440 x W 482 x L 834 mm
CPU 2x Xeon Platinum 8168
2.7 GHz, 24 Core
Operating Temp 5 ~ 35℃
16x Tesla V100 32GB SXM3 3 PetaFLOPS FP16 512GB GPU Memory NVLink (up to 16way) 12 NVSwitches with up to 2.4TB/sec of bisectional bandwidth

DGX-1V

Solution for enterprise-level AI R&DDGX-1V, a must for AI R&D, helps to accelerate a data center and simplify the deep learning workflow, allowing researchers to perform an experiment faster and train a bigger model. NVIDIA DGX-1, mounted with NVIDIA Volta™, offers industry-leading performance for AI and deep learning.

DGX-1 Volta with deep learning performance 140 times faster

트레이닝 시간 성능 비교

Comparison of training time (days)

Workload : ResNet-50, 90 epochs to solution | CPU Server: Dual Xeon E5-2699v4, 2.6GHz
DGX-1V 성능
GPU 8x Tesla V100 SXM2 System Memory 512 GB
GPU Memory 256GB Total Network 2x 10Gbe
4x Infiniband EDR
TFLOPS (FP16) 1,000 TFLOPS
NVIDIA CUDA Core 40,960 Storage 4x 1.92 TB SSD RAID 0
NVIDIA Tensor Core 5,120 Software Ubuntu Linux OS, CUDA Toolkit
NNVSwithces 12 System Weight 134 lbs / 60.8 kg
Max. Power Consumption 3,500 W System Dimensions H 131 x W 444 x L 866 mm
CPU 2x Xeon E5-2698 v4
2.2 GHz, 20 Core
Operating Temp. 10 ~ 35℃
8 x Tesla V100 32GB SXM2 1 PetaFLOPS FP16 256GB GPU Memory NVLink (up to 8 way)

DGX Station

No-noise desk-side super workstation for advanced AI developmentNVIDIA DGX Station™ is an office-use super computer for advanced AI development, and is the only one of its kind.
This system, designed for an office environment and based on the software stack, the same as the one mounted on all DGX systems,
allows you to perform R&D projects effectively and easily.

NVIDIA AI system selection guideline (On-Premises)

AI WORKSTATION
DGX Station
Tesla V100 32GB
The Personal
AI Supercomputer
AI DATA CENTER
DGX-1V
Tesla V100 32GB
The Essential
Instrument for AI Research
DGX-2
Tesla V100 32GB
The World Most Powerful
AI System for the Most Complex
AI Challenges

DGX Station with deep learning performance 72 times faster

딥러닝 트레이닝 가속

Deep leaning training accelerated

DGX Station performance projected based on DGX(with Tesla V100) Workload : ResNet50, 90 epochs to solution | CPU Server: Dual Xeon E5-2699 v4, 2.6GHz, Projections subject to change.
DGX-1V 성능
GPU 4x Tesla V100 PCI-e
GPU Memory 128GB Total Network 2x 10Gbe
TFLOPS (FP16) 500 TFLOPS Storage OS : 1.92TB SSD
Data : 3x 1.92TB SSD
NVIDIA CUDA Core 20,480 Software Ubuntu Linux OS, CUDA Toolkit
NVIDIA Tensor Core 2,560 소음 < 35 dB (liquid cooler, super quiet)
Max. Power Consumption 1,500 W (for general office-use) System Weight 88 lbs / 40 kg
CPU Xeon E5-2698 v4
2.2 GHz, 20 Core
System Dimensions H 639 x W 256 x L 518 mm
System Memory 256 GB Operating Temp. 10 ~ 30℃
4 x Tesla V100 32GB PCIe 500 TeraFLOPS FP16 128GB GPU Memory NVlink (up to 4 way)

TESLA

The most advanced data center GPUNVIDIA®Tesla® acceleration computing platform provides an advanced data center with performance
for accelerating both AI and high computing workload.

  • Deep learning training time for one-day work

    트레이닝 시간 성능 비교

    Time to Solution in Hours
    Lower is Better

    Server Config : Dual Xeon E5-2699 v4 2.6GHz | 8X Tesla P100 or
    V100 | ResNet-50 Training on MXNet for 90 Epochs with 1.28M
    ImageNet dataset.
  • Deep learning performance 47 times faster

    트레이닝 시간 성능 비교

    Performance Normalized
    to CPU

    Workload : ResNet-50 | CPU: 1X Xeon E5-2690v4 @ 2.6GHz |
    GPU : add 1X NVIDIA Tesla P100 or V100
tesla 성능
TESLA Tesla V100 SXM2 Tesla V100 Tesla P100 Tesla P40 Tesla P4
Tesla V100 SXM2 Tesla V100 Tesla P100 Tesla P40 Tesla P4
Intended Use DL Trainning / HPC DL Trainning / HPC DL Trainning / HPC DL Inference DL Inference
CUDA Core 5120 5120 3584 3840 2560
Double Precision 7.8 TFLOPS 7 TFLOPS 4.7 TFLOPS
Single Precision 15.7 TFLOPS 14 TFLOPS 9.3 TFLOPS 12 TFLOPS 5.5 TFLOPS
Half Precision 125 TFLOPS 112 TFLOPS 18.7 TFLOPS 47 TFLOPS 22TFLOPS
GPU Memory 16GB/32GB 16GB/32GB 16GB/12GB 24GB 8GB
Memory Bandwidth 900 GB/s 900 GB/s 732 GB/s
549 GB/s
346 GB/s 192 GB/s