SEOUL | Oct.7, 2016

In Kee Paek, 10/07/2016

Cloud Data Solution Architect

Microsoft

INTRO TO NVIDIA GPUS IN AZURE

2

GPU VIRTUALIZATION VISION

Deliver accelerated graphics & compute capabilities in Azure infrastructure

High end performance

Not “Swiss-army knife” offering

Helps achieve true “HPC in the Cloud”

Close partnership with NVIDIA

3

FINANCE FX Options

Risk Management

Fixed Income Analytics

Market data analysis

Hedge Fund management

4

MANUFACTURING OIL & GAS

Dynamic manipulation of Engineering

models and parts

Realistic material creation

Reservoir modelling and seismic

processing

Automotive engineering

5

MEDIA Stream high fidelity video games

Encoding and transcoding

Image processing

Social media sentiment analysis

6

RENDERING Visual Effects (VFX)

Ray-Tracing rendering

Advertising & Marketing

CAD Applications in Architecture

Simulations

7

TECHNOLOGY

DDA – Discrete Device Assignment

Introduced in Windows Server 2016 as part of Hyper-V

Pass-through PCIe devices directly to a Guest VM

Allows for close to bare-metal performance

8

DDA (DISCREET DEVICE ASSIGNMENT)

Entire device is mapped into the VM just as it would be running on bare metal

Allows for full access to capabilities of that device as well as allowing the device’s native driver to be used

Each device may be mapped to a single VM (1 device to 1 VM), but multiple devices can be mapped into the same VM (1 VM to many devices)

9

10

ARCHITECTURE

11

COMPUTE VIRTUAL MACHINES

12

TESLA K80 – “IT’S FAST…”

0x

5x

10x

15x

K80 CPU

Quantum Chemistry

Molecular Dynamics

PhysicsBenchmarks

13

RATE OF IMPROVEMENT

72%74%

84%

88%

93%

96%

65%

70%

75%

80%

85%

90%

95%

100%

2010 2011 2012 2013 2014 2015

GPU

65%

70%

75%

80%

85%

90%

95%

100%

8-2013 12-2014 5-2016

Accuracy

39%

45%

55%

62%66%

72%75%

79%83%

86%87.5%

30%

40%

50%

60%

70%

80%

90%

100%

Top Score

15

VISUALIZATION VIRTUAL MACHINES

16

REMOTE DESKTOP SERVICES

Graphics performance improvements through DDA

Enhanced support for OpenGL and OpenCL

Connection Broker capabilities to handle massive concurrent connections

Edge support

Pen support

Windows 10 and Mac preview available

17

NV VM DEMO

18

COLLABORATION WITH CNTK

First class citizen

Scalability – multi-GPU-multi-VM

Performance

Internal use-cases across various Microsoft properties and products

(DSVM) Data Science VM by Azure Machine Learning

N-Series and CNTK works really well together

19

CNTK PERFORMANCE ON DDA

2670

10560

18755

27575

35750

0

5000

10000

15000

20000

25000

30000

35000

40000

CPU 1 GPU 2 GPUs 3 GPUs 4 GPUs

Sam

ple

s p

er S

eco

nd

Resource

Avg. Samples/Sec Linear (Avg. Samples/Sec)

20

SUMMARY

High end platform

Support OpenGL and DirectX

Support OpenCL and CUDA

Support existing libraries and tools

Can utilize Azure higher-level services on top of GPU

Fulfill both Compute and Visualization use-cases

21

RESOURCES

http://blogs.technet.com/b/virtualization/archive/2015/11/23/discrete-device-assignment-gpus.aspx

https://azure.microsoft.com/en-us/blog/azure-n-series-preview-availability/

https://channel9.msdn.com/Shows/Tuesdays-With-Corey/Tuesdays-with-Corey-Azure-N-Series-GPU-Acceleration-in-the-public-cloud

SEOUL | Oct.7, 2016

THANK YOU