“Softwares will be written by software running on AI computers” 

Jensen Huang, CEO, NVIDIA

Dubbed as the first kitchen keynote, last year NVIDIA CEO announced DLSS, which revolutionised RTX, RTX servers, NVIDIA Jarvis, A100 and DGX A100. This year too, Jensen Huang’s keynote speech acted as the symbolic ribbon cutting to one of the most anticipated events in the tech world.

The keynote speech highlighted the following developments:

NVIDIA opens gates to its omniverse

Image credits: NVIDIA

Built on top of RTX, Omniverse is a platform to connect multiple 3D worlds into a shared virtual world. It takes its inspiration from the term “metaverse”. “Pieces of the early-metaverse vision are already here in massive online social games like Fortnite or user-created virtual worlds like Minecraft,” said Huang.

“Omniverse is a platform built from the ground up to be physically based. It is fully path traced. Physics is simulated with NVIDIA PhysX, materials are simulated with NVIDIA MDL, and Omniverse is fully integrated with NVIDIA AI,” he added. 

The platform is cloud-native, multi-GPU scalable and runs on any RTX platform and can be streamed remotely on any device.   

Foray into data centre markets

The biggest takeaway from the keynote is the launch of NVIDIA’s first data centre CPU, Grace. Each CPU delivers 300 SPECint with a total of over 2,400 SPECint rate CPU performance from an eight-GPU DGX. 

The chipmaker has also announced the first-ever DPU made for AI and accelerated computing—Bluefield 3, which allows organisations to create applications with industry-leading performance and data centre security.

According to NVIDIA, Bluefield-3 can achieve 400 GB/s and has 10x processing capability of Bluefield-2. It is optimised for multi-latent, cloud-native environments software-defined, hardware-accelerated networking, storage, security and management services. It delivers equivalent data centre services of up to 300 CPU cores. 

Huang also announced DGX station, which has four 80GB A100 GPUs having more memory and bandwidth than the original DGX station. Meant for AI research, it also has refrigerated liquid cooling for its Epyc CPU and four A100 GPUs.

The station runs at a quiet 37 dB while only utilising up to 1,500 W of power. It delivers up to 2.5 petaFLOPS of floating-point performance. One of the most impressive features is that it transfers 8TB per second.

Talking about the performance, Huang said that the Megatron could linearly scale training up to 1 trillion parameters on the DGX SuperPOD with advanced optimisations and parallelisation algorithms. 

MLOps and Enterprise AI

NVIDIA EGX enterprise platform: The platform enables both existing and modern, data-intensive  applications to be accelerated and secure on the same infrastructure. It delivers the power of accelerated computing from the data centre to edge with a range of optimised hardware, an easy to deploy application and management software and a vast ecosystem of partners who offer EGX in their products.

Huang also gave a glimpse of frameworks that encapsulate the entire workflow to customise AI models. It applies transfer learning to your data to fine-tune models. “No one has all the data – sometimes it’s rare, sometimes they are trade secrets. No model can be trained for every skill. And the specialised ones are the most valuable,” said Huang as he introduced Nvidia TAO. 

TAO has a federated learning system allowing multiple users to train a shared model while having data privacy. Using TensorRT, it optimises the model for the target GPU system.

A Quantum cue

NVIDIA’s cuQuantum— an acceleration library designed for simulating quantum circuits, is meant for both tensor network solver and state vector solvers. It is also optimised to scale to large GPU memories, multiple GPUs and multiple DGX nodes. Developers can use it to speed up quantum circuit simulations based on state vector, density matrix, and tensor network methods by order of magnitude.

Stay tuned to AIM for more updates from GTC 2021.