[GTC 2020] GPU Technology Conference Unusual Kitchen Keynote Speech Report by NVIDIA CEO Jensen Huang!

This year is also the season for the GPU Technology Conference (GTC), one of the world's largest events in the fields of AI, deep learning, IoT, robots, and autonomous driving. If it is an average year, the conference will be held on a large scale in California, USA. Experts and intellectuals in the field discuss AI, deep learning technology, and the future of peripheral technologies such as cloud and enterprise computing, and related companies and partners all exhibit their latest technologies. It's an event. This year, unfortunately, the spread of the COVID-19 is still not over in Santa Clara. The GTC keynote speech was also streamed on YouTube from the CEO's home kitchen at 10:00 pm on May 14, 2020 Japan time in the form of H2H (Home2Home), reflecting the social situation.

GTC, a conference for developers, offers a variety of cutting-edge technology sessions, including GPUs, and a number of AI sessions, which have been attracting attention recently. In addition to the keynote speech by NVIDIA CEO Jensen Huang, this year's H2H also featured a number of sessions.

Therefore, we would like to introduce three of the latest technologies Macnica is focusing on from the CEO's keynote speech.

➀ GPU "NVIDIA A100" with new Ampere architecture

The GPU "NVIDIA A100" that first adopted the new Ampere architecture, which is the successor to Volta.

In 2017, we announced a GPU architecture called Volta and started selling the GPU "Tesla V100" for data centers based on it. "Tesla V100" realized high computational performance during AI learning / inference by adding a computing unit specialized for AI called Tensor core in addition to computing unit for CUDA core parallel operation.

The Ampere architecture A100 announced this time adopts a new Tensor core, and adopts a mechanism called "TF32" that can perform FP32 calculations in a hybrid method of FP32 and FP16. The performance during calculation has been greatly improved.

A100 is a GPU that uses TSMC's 7nm process, implementing 54 billion transistors, which is more than double the previous generation Volta, and 312 TFLOPS (teraflops) during FP32 (single-precision floating point) calculations through hybrid calculation. , and has a 20-fold improvement in computational performance compared to conventional products.

In addition, the server "NVIDIA DGX A100" (hereinafter referred to as DGX A100) for AI learning equipped with A100 was also announced. This is the latest product in the DGX series, following the DGX-1 (equipped with 8 Tesla V100s) and the DGX-2 (equipped with 16 Tesla V100s). With the adoption of A100, an architecture (MIG: Multi-Instance GPU) that can be applied to all data center use cases is implemented, and it is expected to be used for a wide range of applications.

Features of the new “NVIDIA DGX A100” model

• Equipped with 8 new GPU "A100"
• Achieves about 20 times the performance of conventional Volta generation GPUs in AI learning and inference
• 6 NVSwitches connecting GPUs
• Equipped with 9 Mellanox network card products "ConnectX-6 VPI"
• The CPU uses AMD's "2nd generation EPYC" (development code name: "Rome")

② NVIDIA x Mellanox next-generation data center construction model

Also announced was a vision to build a next-generation "DGX SuperPOD" that combines 140 DGX A100 systems with 170 Mellanox Quantum InfiniBand switches. Despite the cluster's powerful AI computing performance, it is said that it can be built in three weeks.

In addition, NVIDIA has expanded 'NVIDIA DGX SATURNV' as a data center showcase. In addition to the 1800 DGX system, the company announced the addition of three DGX A100 SUPERPODs (560 DGX A100s) to add even more processing power.

③ EGX edge AI platform

For the coming 5G era, it is an urgent issue for companies to put edge AI to practical use to reduce network load and achieve low-latency and secure processing for analysis of the huge amount of data collected by IoT devices. It is a well-known fact that

The EGX A100 is the first product for edge AI based on the Ampere architecture. By equipping the server with the EGX A100, it is now possible to process and protect large volumes of data streamed from edge sensors in real time. increase.

The EGX A100 combines blazing performance with the superior networking and security features of the NVIDIA Mellanox® ConnectX-6 Dx SmartNIC. With the advent of NVIDIA Mellanox's time-triggered transmission technology (5T for 5G), large amounts of data can be processed in real time, making standard servers into AI-capable supercomputers.

For example, it is expected to be used in public facilities such as airports and train stations to manage hundreds of cameras with an edge server equipped with "EGX A100".

In addition to the above, the keynote speeches included next-generation NVIDIA RTX, a demonstration that makes full use of ray tracing technology that expresses realistic illumination, shadows, and vivid colors, and NVIDIA Jarvis, a framework for multimodal next-generation conversational AI robots. , the Misty demo that embodies AI, and the latest information on autonomous vehicle development were also introduced.

Click here to watch the keynote

At the end

At GTC, which has become an annual event, you will be able to come into contact with a wide variety of information, including the latest NVIDIA technology, the current state of GPUs, and future roadmaps.

After watching the GTC Kitchen keynote speech this time, the provision of a server that enables a more powerful AI learning environment and the creation of a compact edge AI platform environment have greatly expanded the introduction frontage for using AI in business. I felt that it might have spread to I'm really looking forward to where NVIDIA is headed in the future and what kind of world they will provide using AI!

Click here for inquiries about announced products