At the 2021 NVIDIA GTC conference Jensen Huang spoke from his kitchen to say that we need a metaverse (the Omniverse, with attribution to SF author Neil Stephenson), a digital twin of the real world and that NVIDIA is building the tools to make this possible.
He spoke about democratizing high performance computing and of course, his talk included cool videos of special effects and showing how the Omniverse can be used to teach robots how to be robots and to design better factories. An important element in the Omniverse is creating ever more realistic models with “digital twins” that can accurately describe the real world and can be used for various optimization and failure analysis applications.
Jensen talked about 4 stacks of technology and their applications. These are RTX visual computing platform that is being used to create the Onmiverse; computing and accelerator platforms (DGX, Grace, BlueField and DOCA); the EGX accelerated computing platform applied to Jarvis, Merlin, Maxine, Morpheus and NVIDIA AI; and autonomous vehicle products Hyperion, Atlan and Orlin.
NVDIA said that AI model sizes are getting bigger in order to make more accurate models. The size of the largest models doubles every 2.6 months and will be over 100 trillion parameters by 2023. The growth in the amount of data that needs to be processed to support these huge models is pushing the limits of existing computer architectures. The figure below, from Jensen’s talk, show growing use of GPU based modeling for more and more complex models.
To meet these processing needs NVIDIA introduced its GRACE CPU for giant-scale AI and high-performance computing (HPC) applications.
An important element in this performance is the speed that data goes from and to memory. The GRACE chip supports over 500GB/s using LPDDR5 with error correction coding. NVIDIA says that trillion parameter models can be trained in 3 days versus about a month with older hardware and can provide real-time inference on a 500 billion parameter model. NVIDIA says that it has two supercomputer customers for this product including the world’s fastest supercomputer for AI, the ALPS in Switzerland.
In addition, NVIDIA said it will be offering expanded ARM processing in the cloud using AWS Graviton2 with NVMDIA GPUs to deliver high scale performance streaming, to support applications such as Android gaming. The company is also offering cuQuantum that provides a library of quantum computing simulation capabilities to use today’s computers to develop the computers of tomorrow.
NVIDIA also announced its BlueField-3 digital processing unit (DPU), shown below, a special purpose processing device for software defined networking, storage and cybersecurity acceleration for multi-tenant, cloud-native data centers that is expected to sample in Q1 2022. In the meantime, BlueField-2 DPUs are available today.
The 400GbE/NDR DPU, BlueField-3, delivers unmatched networking performance. It features 10x the accelerated compute power of the previous generation, with 16x Arm A78 cores and 4x the acceleration for cryptography. BlueField-3 is also the first DPU to support fifth-generation PCIe and offer time-synchronized data center acceleration. One BlueField-3 DPU delivers the equivalent data center services of up to 300 CPU cores, freeing up valuable CPU cycles to run business-critical applications. BlueField-3 also provides real-time network visibility, detection and response for cyber threats and acts as the monitoring (telemetry) agent for the just announced NVIDIA Morpheus, an AI-enabled cloud-native cybersecurity platform.
BlueField-3 also enables NVIDIA DOCA, which gives developers an open software platform for building software-defined, hardware-accelerated networking, security and management applications running on BlueField DPUs. Dell, Inspur, Lenovo and Supermicro are integrating BlueField DPUs into their systems. Cloud services providers are using BlueField DPUs to accelerate various workloads. BlueField-3 support will be available from storage providers DDN, NetApp, VAST and WekaIO.
According to the company’s press release, “BlueField-3 lets every enterprise deliver applications at any scale with industry-leading performance and data center security. It is optimized for multi-tenant, cloud-native environments, offering software-defined, hardware-accelerated networking, storage, security and management services at data-center scale.”
NVIDIA was showing its DRIVE Atlan, which it called an AI data center on wheels for next generation autonomous vehicles that combines many NVIDIA products that provides more than 1,000 TOPS and data-center-grade security and is targeted for automakers 2025 models. The company says that it, “fuses AI and software with the latest in computing, networking and security for unprecedented levels of performance and security.”
DRIVE Atlan includes NVIDIA’s next generation AMPERE GPU, ARM CPU cores as well as seep learning and computer vision accelerators to provide automakers to create software-defined vehicles that can be upgraded through secure over-the-air updates.
The 2021 GTC keynote by Jensen Huang talked about how the company’s GPUs, CPUs and other accelerator technologies are turning Science Fiction concepts such as the Metaverse into reality and create data centers on wheels.
This is a syndicated post. Read the original post at Source link .