As expected, Nvidia took advantage of the GTC on Tuesday to lift the veil on a new architecture: the Nvidia Hopper, in honor of Grace Hopper. The Nvidia H100 chip is the first to use this architecture.
Two years after the introduction of the Ampere architecture that brilliantly revitalized the GeForce RTX 3000, A-series chips for data centers, and RTX A-series chips for professional PCs, Nvidia is introducing a new architecture: Hopper.
Architecture for professionals
We must first make an important clarification. While Ampere was an architecture designed for both the gaming market and the professional market, Nvidia decided to work on two architectures for its new generation of products. On Tuesday during GTC, the developer conference, Nvidia unveiled the Hopper for professionals, and we expect the Ada Lovelace architecture for the GeForce RTX 4000 by the start of the school year.
However, it’s still very interesting to discover what Hopper has to offer in order to understand what Nvidia has to offer to the general public in the coming years.
80 billion transistors, 4nm manufacturing
This is the announcement of all superlatives for Nvidia. The firm introduced both the Hopper architecture and the H100 chip, which was the first to use the architecture. The chip, which will still integrate 80 billion transistors, will make it the largest and most powerful accelerator in the world, according to Nvidia. We’re under the 114 billion transistors of the Apple M1 Ultra chip, but the latter is an SoC that integrates many components.
To make this chip, Nvidia uses TSMC and their new 4nm manufacturing process. Recall that Nvidia has chosen Samsung as a partner for the production of Ampere chips using the 8nm process technology. This chip can be paired with the latest generation HBM3 memory providing 3TB/s bandwidth. Nvidia says that with this bandwidth, only 20 Hopper H100 chips are required to support global Internet traffic. It is enough to present AI applications on a global scale and in real time.
Another characteristic element of Hopper is the transition to Nvidia’s 4th generation NVLink, which allows scaling up to 256 H100 chips in the data center, improving throughput by up to 9 times over the previous generation. The H100 chip is the first chip to natively support PCI Express 5.0 to maximize throughput on servers using this protocol.
Faced with Ampere, Nvidia has no hesitation in announcing a 3-4x increase in processing power to reach dizzying levels: 4000 teraflops in FP8, 2000 teraflops in FP16 and 60 teraflops in FP32/FP64.
Nvidia transformer engine
At the heart of the Nvidia Hopper architecture is the Transformer Engine, which should greatly enhance the chip’s AI capabilities and performance. The name directly refers to the transformer, an increasingly popular and promising technology in the field of artificial intelligence. Google BERT or GPT-3 OpenAI is based on the transformer model. This model is increasingly being used in fields related to translation, image analysis, or health research.
According to Nvidia, the integration of the Transformer Engine should make it possible to speed up artificial neural networks up to 6 times compared to the old generation.
Nvidia offered several performance comparisons between the next generation H100 and the Ampere A100 chip. With serial GPU scaling, Nvidia promises to increase performance by up to nine times, reducing training time from 7 days to 20 hours.
Launch in Q3
Nvidia Hopper will be available to professionals from the third quarter. The firm has already signed contracts with leading partners in the sector such as Dell, Fujistsu, Supermicro, Cisco, Atos and Hewlett Packard Enterprise to offer products based on the H100.
To follow us, we invite you to download our app for Android and iOS. You can read our articles, files and watch our latest YouTube videos.