Thread by @amogh42

Jensen Huang, @nvidia CEO just kicked off #GTC2020 with a keynote that covered several groundbreaking announcements and partnerships. The future of #AI and #Nvidia is incredibly exciting. Here's a thread of the (many) key takeaways. Let's get started!🧵👇

The main focus this time was on #AI & high performance computing in the data center & on the edge. This #GTC2020 Nvidia is releasing 80 new and updated SDKs. CUDA, Nvidia's toolkit for GPU powered applications, has been downloaded 20M times, 6M in 2020 alone.

#Omniverse, Nvidia's platform for simulation & collaboration is now in open beta. Using Omniverse teams can simultaneously work in Blender, Maya, Unreal etc and view the results rendered in real time in a common interface.

It is cloud native and also supports ray tracing. It also supports AI agent plugins like animating a character's face automatically based on any dialogue audio. Omniverse can also be used for training robots and other AI in a virtual world with real world physics simulation.

Drug discovery is a $1.25 Trillion industry that is being disrupted by AI. About $2 billion is spent on R&D per drug with a 12.5 year development time and a 90% failure rate. So why is drug discovery so hard?

1. It's hard to find the exact protein implicated in a disease
2. It's hard to find the right small molecule that can bind with the protein to activate/deactivate it
3. It's hard to get the small molecule inside the cell
4. It's hard to predict how the body will react to a drug

Nvidia Clara Discovery is a suite of tools for every stage of the drug discovery process. It includes components like RAPIDS for data analysis, Bio-Megatron language models, and Clara for imaging. Nvidia AI can use simulations to understand the biological machinery of proteins.

Nvidia is building the Cambridge-1, a 400 petaflops supercomputer for healthcare research which will be the fastest in UK. First partners are @AstraZeneca, @GSK, @KingsCollegeLon, @NHSEngland and @oxfordnanopore.

@GSK and @nvidia are partnering for the world's first AI drug discovery lab. Biology is incredibly complex and GSK has gathered more data in the first quarter of 2020 than in its entire 300 year history combined. Processing this data for drug discovery is the goal.

To enable large scale AI compute infrastructure, Nvidia is releasing the DGX Superpod supercomputer architecture blueprint. This will enable partners to build supercomputers using Nvidia tech. @cdacindia is already building India's fastest supercomputer using DGX Superpod.

Computation required to train state of the art AI models has increased 30,000x in less than 5 years.
Nvidia's Megatron BERT Transformer model for #NLP scored 91% on the RACE reading comprehension dataset (30,000 passages and 100k MCQ questions). The average human scores 70%

Nvidia's goal is to democratize the 3 pillars of AI -
1. Data pre-processing, feature engineering & training
2. Inference optimized for accuracy, throughput, response time & memory size
3. Pretrained models & engines for problems like conversational AI & self driving vehicles.

NGC container registry provides optimized & containerized NVIDIA computing stacks for every cloud. So far it's got 1 million downloads, 250k users over 2 years and 4X growth. NGC will be available within Azure, AWS and GCP marketplaces for easy use on any cloud.

2020 is the 10th anniversary of the 1st Nvidia GPU in @awscloud. Nvidia AI inference computation is growing 10x every 2 years and this year surpassed total CPU compute in the cloud. In 3 years Nvidia GPUs are expected to represent 90% of the total cloud AI inference compute.

@Microsoft is adopting Nvidia AI stack to create smart experiences in Office such as text prediction, grammar correction, and Q&A. Office will be connected to Nvidia GPUs in Azure with inference running for millions of simultaneous users.

Jarvis, Nvidia's conversational AI stack is now in open beta. It includes all basic elements like speech to text, Megatron - BERT based natural language understanding and human sounding text to speech. All delivered in under 300ms latency using the Triton inference server.

For live video calls Nvidia introduced Maxine, a cloud native video streaming AI platform. By mapping facial features, sending only data about changes in expression and reanimating the files at the receiver's end, video call bandwidth can be reduced & call quality improved by 10x

It can also reorient your face so that you're making eye contact with each person individually. AI driven facial animations can automatically move a 3D character's lips. Combined with real time language translation & close captioning, video calls are going to be revolutionized.

After acquiring Mellanox, Nvidia is set to introduce the game changing Data Processing Unit. A DPU is a programmable datacenter infrastructure processing chip. An estimated 30% of CPU cores in a datacenter are consumed running the infrastructure software.

Nvidia's Bluefield 2 DPU separates the application processing domain from the infrastructure domain on a single chip. DOCA is a programmable datacenter infrastructure processor architecture. DOCA SDK lets developers write custom apps for datacenter networking, storage, security.

Nvidia also announced a major partnership with @VMware. They will be porting VMware on the Bluefield 2 DPU. The Bluefield 2X chip will add the Ampere GPU compute to Bluefield 2 to boost the infrastructure computing with CUDA and Nvidia AI.

Nvidia also announced Merlin, a petabyte scale end to end accelerated recommendation AI. The core engine is Nvidia Rapids whose API is modelled after #Pandas, #Scikitlearn and #XGBoost. It's currently being used in @TencentGlobal's ecommerce operations.

Nvidia EGX Edge #AI platform is used for controlling fleets of robots in autonomous factories. The EGX integrates Bluefield 2 DPU with Ampere GPU to turn any standard OEM server into an AI data center. Jensen Huang calls it the iPhone moment for industrial edge computing.

Nvidia Fleet Command is a new cloud based platform for managing AI at the edge. You can manage edge AI deployments from browser across large geographies.

The Jetson Nano will now retail at $59 only.

Nvidia's Isaac Sim is a virtual world to design and train robots. Robots can be trained in simulated environments before the same AI models are deployed in actual robots in live environments.

Nvidia Drive is an autonomous vehicle training platform. Training simulations can be run in Drive Sim built on Omniverse. Nvidia has partnered with Mercedes so that starting from 2024, Mercedes' self driving cars will be powered by Nvidia Drive.

Training the car in Drive Sim involves using an RTX GPU to generate a virtual world view for each sensor. This data is sent to the Drive AV computer which runs inference and moves the car in Drive Sim which in turn generates new views of the world.

Drive AV is running the actual self driving car models and does not know it is driving in a virtual world. The trained AI models can the be tested and deployed in actual cars in a real world environment.

@threadreaderapp please unroll