AI Nvidia

AI in Computex Taipei 2018,

Deepu Talla, Nvidia

Computing petaflop

New tensor core GPU  computing

 

1075 images per sec by single chip

15500 images per sec by single node

6250 images per sec by inference

1.1   Milisec fastest inference

14 minutes fastest scale

 

Rene Haas, arm, IP products group

Time to train alexnet, from 2012 6days to 18 min in march.

Cambrian explosion

1Conventional networks

2Recurrent network

3Generative adversarial network

4Reinforcement learning

5New species

 

Key characteristics for data center of AI:

Programmability

Latency

Accuracy

Size

Thruput

Energy efficiency

Rate of learning

 

Nvidia AI inference

Tensor RT 4

Tensorflow integration

Kaidi optimization

 

30M hyperscale services

Many selection of nvidia server platforms

 

AI application is revolutionalizing industries:

Automotives

Smart cities

Healthcare

Consumer IoT (home, thermosat, wearable, appliance)

 

Nvidia driving 

Audi, Toyota, benz. Volvo etc 370+ partners.

More than $10T transportation industry

 

End-to-end platform:

Collect and process data (lidar,radar)

Train models

Simulate (special exception situation finding and solving)

Drive

 

$2T smart city industry

Air quality, heat, face recognition, traffic management

 

$7T healthcare industry for diagnosis, recognition and simulation

1 triilion consumer intelligence devices

Nvidia provides modular –configurable design