At its annual GPU Technology Conference (GTC), NVIDIA announced several new products to solidify its position in the artificial intelligence (AI) space. The American company unveiled the new Hopper Architecture, a few products including the H100 GPU, Grace CPU Superchip, and the Omniverse Cloud for real-time design collaboration and simulation capabilities.
The announcement of the new lineup of next-generation products comes on the heels of a couple of setbacks NVIDIA faced in recent days. Soon after its $40 billion bid to acquire Arm fell through, 19 GB of its data was leaked online by the Lapsus$ cyber extortion gang. Let’s look at what NVIDIA has in store at the spring event.
NVIDIA Hopper Architecture
Named after rear admiral and computer science pioneer Grace Hopper, the Hopper architecture will replace NVIDIA’s Ampere Architecture. Ampere was introduced two years ago at NVIDIA’s 2020 GTC just as the COVID-19 pandemic gained momentum. NVIDIA has maintained its dominance in GPUs in these two years as demand remained consistently high, thanks to widespread computing needs.
The company now believes it’s time for an updated accelerated computing platform that uplifts its AI/ML workloads investment. Specifically, the Hopper Architecture features new Transformer Engines designed for natural language processing (NLP) and computer vision tasks.
NVIDIA Hopper Architecture (H100 Die)
Transformers are deep learning models that are used in the analysis of complex data associated with NLP processes. It adheres to the self-attention principle and powers GPT-3 and previous versions by OpenAI and others. NVIDIA leveraged this to instill Transformer Engines in the Hopper Architecture to help accelerate machine learning model training.
Rapid transformer training is essential because of the exponential rate at which connections in emerging transformers are scaling up. GPUs based on Hopper Architecture, combined with the latest communication link, have the potential to train a transformer up to 9x faster.
See More: CES 2022: Chip Makers Intel, AMD, and NVIDIA Unveil Their Latest Wares
NVIDIA H100 GPU
This latest GPU is based on the Hopper architecture and is NVIDIA’s flagship acceleration GPU targeted for server implementations. The chipmaker said H100 is 6x faster than the previous A100.
NVIDIA H100 consists of 80 billion transistors, has a memory bandwidth of 3 TB/s and is the first GPU to support PCle Gen5 and utilize HBM3. It is based on the advanced TSMC 4N process. “Twenty H100 GPUs can sustain the equivalent of the entire world’s internet traffic, making it possible for customers to deliver advanced recommender systems and large language models running inference on data in real time,†NVIDIA noted.
H100 features confidential computing capabilities, thus allowing federated learning/training for privacy-sensitive applications and DPX instructions for accelerated dynamic programming.
You can find technical details of the H100 GPU on NVIDIA’s website. Below is AnandTech’s comparison of H100 with two of its predecessors:
|
NVIDIA Accelerated Computing AI GPUs |
|||
| Specifications | H100 | A100 |
V100 |
|
Manufacturing Process |
TSMC 4N | TSMC 7N | TSMC 12nm FFN |
| Architecture | Hopper | Ampere |
Volta |
|
Memory Bandwidth |
3 TB/sec | 2 TB/sec | 900 GB/sec |
| Transistor Count | 80 Billion | 54.2 Billion |
21.1 Billion |
|
Memory Clock |
4.8 Gbps HBM3 | 3.2 Gbps HBM2e | 1.75 Gbps HBM2 |
| VRAM | 80 GB | 80 GB |
16 GB/32 GB |
|
Interconnect |
NVLink 4 | NVLink 3 | NVLink 2 |
| Thermal Design Power | 700 W | 400 W |
300 W/350 W |
|
Boost Clock |
Estimated 1.78 GHz | 1.41 GHz | 1.53 GHz |
| Tensor Cores | 528 | 432 |
640 |
NVIDIA H100 is expected to be available later in 2022.
Grace CPU Superchip
Also intended for server implementations, Grace CPU Superchip is NVIDIA’s first data center CPU based on the Arm Neoverse. Like the H100 GPU, these CPUs will also be powered by the Hopper architecture for high-performance computing (HPC).
It leverages NVLink®-C2C, a new high-speed, low-latency, chip-to-chip interconnect used to connect two CPU chips. “A new type of data center has emerged — AI factories that process and refine mountains of data to produce intelligence,†said Jensen Huang, founder and CEO of NVIDIA.
NVIDIA Grace GPU Superchip
NVIDIA said Grace CPU Superchip that uses CPU-GPU integration could deliver twice the memory bandwidth and energy efficiency of chips by competitors. It features 144 Arm cores in a single socket, 1 TB/s bandwidth, and consumes 500 W, making it fairly energy efficient. Grace CPU Superchip will be available in the first half of 2023.
See More: What Is Artificial Intelligence (AI)? Definition, Types, Goals, Challenges, and Trends in 2022
Enterprise AI and Supercomputing
NVIDIA’s infrastructure/hardware-level enterprise AI offerings go hand in hand with software services. The Santa Clara-based company has announced updates for Maxine and Rava, both leveraged by a host of companies for respective AI-driven applications.
NVIDIA Maxine, for instance, is used for noise cancellation, virtual backgrounds, etc., by Avaya. Uses cases of Maxine, a GPU-accelerated software development kit (SDK), include bringing AI-powered capabilities to video conferencing, content creation and streaming applications. NVIDIA Riva is used in speech AI applications such as virtual assistants, real-time transcription, and chatbots.
At the infrastructure level, NVIDIA introduced DGX H100 Systems based on the namesake GPUs. WIth both infrastructure and software advancements, NVIDIA hopes to realize its ambitions to develop the world’s fastest AI supercomputer.
Dubbed NVIDIA Eos, this supercomputer is expected to commence operations later this year. It features 576 DGX H100 systems with 4,608 DGX H100 GPUs. At 18.4 exaflops of AI computing performance and 275 petaflops of regular scientific computing performance (HPL), Eos will provide 4x faster AI processing than the current fastest Fugaku supercomputer in Japan.
The A100-based NVIDIA Selene delivers 2.8 exaflops of AI peak performance and 63.460 petaflops on HPL. NVIDIA is building Eos to advance research in climate science, digital biology and the future of AI.
See More: The Metaverse is Here…But is the Hardware Ready?
NVIDIA Omniverse Cloud
NVIDIA’s metaverse plans announced last year through the Omniverse suite of software received a significant bump in global availability, applicability, and scalability. The company announced Omniverse Cloud to enable customers to develop simulations over the cloud.
NVIDIA Omniverse Cloud is perhaps the most important announcement considering the chipmaker is traditionally not a metaverse or virtual reality company. However, NVIDIA does have vested interests in the sphere, given it manufactures the underlying hardware. So why not offer tools that streamline development on the NVIDIA infrastructure?
Omniverse Cloud includes a suite of cloud services that enables content creators, artists, and developers to access the Nvidia Omniverse platform readily.
NVIDIA Omniverse Cloud
NVIDIA Omniverse is “a real-time physically accurate world simulation and 3D design collaboration platform.†According to the vice president of Omniverse Platform Development at NVIDIA, Richard Kerris, since its launch in April 2021, the Omniverse ecosystem has expanded ten times concerning companies and creator involvement.
Over 150,000 individuals have downloaded NVIDIA Omniverse to date. With Omniverse Cloud, the tools for developing 3D content can run even on run-of-the-mill computers that do not have GeForce or RTX hardware by NVIDIA or any other high-performance capabilities.Â
Kerris said, “By having all of Omniverse in the cloud, it becomes available to anybody, no matter what kind of platform you’re on, whether you’re on a Chromebook, a Mac or tablet.â€
The full Omniverse Cloud collection of services is currently under development.
Let us know if you enjoyed reading this news on LinkedInOpens a new window , TwitterOpens a new window , or FacebookOpens a new window . We would love to hear from you!