This presentation reviews the top challenges facing your data science team and the benefits an NVIDIA DGX Station can offer. Learn how your teams can get the performance of GPU-powered deep learning, without the dependency on a data center server. Discover how to help your researchers and developers save time and avoid wrestling with IT, and gain instant productivity with an optimized deep learning software stack, that’s up and running in just a couple hours instead of weeks. Explore the benefits of a personal supercomputer that fits neatly desk-side, with the versatility to take your team from experimentation to training at scale, to deep learning inference and insights.
Transcript: New from BookNet Canada for 2024: BNC CataList - Tech Forum 2024
5 Reasons Why Your Data Science Team Needs The DGX Station
1. 5 REASONS WHY YOUR DATA
SCIENCE TEAM NEEDS THIS
POWERFUL DEEP LEARNING
WORKSTATION
THE PERSONAL AI SUPERCOMPUTER
NVIDIA® DGX Station™
2. Challenges:
- You need the power of a server, but need
the convenience and ease of a workstation.
- Traditional platforms lack the speed to
experiment fast, and iterate faster.
- Researchers don’t have time to waste on IT,
troubleshooting code, or optimizing a deep
learning stack.
- You need a software stack that lets your
models follow you across desk, data center
and cloud.
- Today’s platforms lack the versatility to
support you from experimentation to
training to inference.
3. “The School of Computer Science and Communication has a powerful
compute server. However, for our current projects we need a compute
server that we have exclusive access to.”
— Professor, Computer Science Faculty of an European University
“[We’re looking for a] … a workstation configured to create and train convolutional neural networks to be set up in
the department that will have the ability to download and store multiple images and reports. [..] Training of an
algorithm utilization high resolution medical images requires a large memory bandwidth such as that
available in NVIDIA GPUs.”
— Scientist, a Well Known North American University’s Medical Center
“I use deep learning to secure cyber assets. Access to a deep learning
workstation will increase the speed of innovation and improve security.”
— Scientist at large research lab
DATA SCIENTISTS HAVE TOLD US…
4. TOP 5 REASONS YOU CAN ACCELERATE DEEP LEARNING
VALUE AND INSIGHTS WITH NVIDIA DGX STATION
5. #1: DESIGNED FOR WHERE YOU WORK
The power of 400 CPUs – at your desk
Consumes only 1500W, drawing 1/20th
the power of a traditional workstation
Emitting only 1/10th the noise of other
workstations
1
Download Infographic
6. #2: 3X FASTER THAN THE FASTEST WORKSTATIONS
2
Learn More Watch Video
Water-cooled performance – the
only workstation built on 4 Tesla
V100’s
3X the performance of today’s
fastest GPU workstations
30% faster training over non-DGX
stack solutions
5X increase in I/O performance
with 4-way NVLink vs. PCIe-
connected GPU’s
480 TFLOPS
30%
5X
3X
7. #3: EFFORTLESS PRODUCTIVITY
3
Access popular deep
learning frameworks,
NVIDIA-optimized
for maximum performance
DGX containers
enable easier
experimentation and
keep base OS clean
Develop on DGX
Station, scale on
DGX-1 or the
NVIDIA GPU Cloud
Watch Video
8. - Single, unified stack for deep learning frameworks
- Predictable execution across platforms
- Pervasive reach
#4 - COMMON SOFTWARE STACK ACROSS DGX FAMILY
DEEP LEARNING FRAMEWORKS
DGX Station DGX-1 NVIDIA Cloud Service
NVIDIA
GPU Cloud
DEEP LEARNING USER SOFTWARE
NVIDIA DIGITS™
NVIDIA DEEP LEARNING SDK
CONTAINERIZATION TOOL
NVIDIA Docker
Docker
GPU DRIVER
NVIDIA Driver
SYSTEM
Host OS
4
9. #5 – DEEP LEARNING FROM DEVELOPMENT TO INFERENCE
Accelerated Deep Learning Value with DGX Solutions
Experiment
Tune/
Optimize
Deploy Train Insights
Procure
DGX
Station
Install /
Compile
Training
Productive
Experimentation
Fast Bring-up
DGX/CSP
DGX Station
From Desk
installed optimized
Inference
to Insights
refine, re-train
5
11. RESEARCHER
“I felt I won the software stack lottery as NVIDIA-
docker was already installed. I immediately pulled a
container and started work on a CNTK NCCL project,
the next day pulled another container to work on a
TF biomedical project. I haven’t looked back at how
to reimage because felt too productive.”
12. DEVELOPER
“For the numbers, it’s taking about 1-2 hrs to train a
152 layer ResNet on a ~20GB dataset, which is pretty
good and keeping me active with experiments rolling,
just on the workstation. It feels right for this work to
allow fast iteration. The last time I did some serious
model architecture/tuning work it took halfdays to days
on Kepler GPUs.”
13. VENTURE BEAT
NVIDIA also said it was launching its Optix 5.0 SDK on the Nvidia DGX AI
workstation. That will give designers, artists, and other content-creation
professionals the rendering capability of 150 standard central processing unit
(CPU) servers. By running Nvidia Optix 5.0 on a DGX Station, content creators
can significantly accelerate training, inference and rendering (meaning both AI
and graphics tasks). – Dean Takahashi, Venture Beat
14. TOP 5
1. You can now get the right kind of power, conveniently at your desk
2. Now you can save time and money by starting your
experimentation in hours, not weeks, powered by DGX Stack
3. Be more productive with your training and inference work that
goes from desk to data center to cloud
4. Get breakthrough performance and precision – powered by Volta
5. Flexibility to do AI work at the desk, data center, or edge
The Fastest Personal Supercomputer for Researchers and Data Scientists
Access popular deep learning frameworks, NVIDIA-optimized for maximum performance
DGX containers enable easier experimentation and keep base OS clean
Develop on DGX Station, scale on DGX-1 or the NVIDIA Cloud
DGX Station is perfect for individual researchers or data scientists, or small groups thereof, that need flexibility and want full control over their compute resources.
As both system share the same architecture, the DGX Station provides the freedom to seamlessly scale workloads to the infinite compute power of a DGX-1 cluster.
A typical deep learning workflow could like the following:
DGX Station: build a promising model. Iterate frequently in hours. Use 2 to 4 GPUs.
DGX-1: Make the model work with real data and optimise data and optimise data and optimise data and optimise data and optimise it. Iterate in days to weeks. Use 4 to 8 GPUS.
DGX-1: Train the model in production, and prepare it for your customers. Iterate from hours to weeks. Use 64 to 128 GPUs in a DGX-1 cluster.
And eventually:
Cloud/Datacenter/Embedded devices: Serve the trained model and provide functionality to your customers. In milliseconds. Use P40s or P4 or TX1.