American multinational technology company Nvidia has unveiled a new generation of artificial intelligence chips and software at its developer’s conference in San Jose, aiming to strengthen its position as the leading supplier for AI companies.
The announcement comes amid Nvidia’s remarkable growth, with its share price soaring five-fold and total sales more than tripling since the AI boom initiated by OpenAI’s ChatGPT in late 2022. Nvidia’s high-end server GPUs play a crucial role in training and deploying large AI models, attracting significant investments from tech giants like Microsoft and Meta.
The new AI graphics processors, named Blackwell, are spearheaded by the GB200 chip, set to be released later this year. Nvidia’s push for more powerful chips aims to drive new orders, especially as demand for the current “Hopper” H100s remains high. In addition to hardware advancements, Nvidia introduced revenue-generating software called NIM, designed to streamline AI deployment and provide customers with added value.
Nvidia executives envision the company evolving from a chip provider to a platform provider, akin to Microsoft or Apple, enabling other companies to build software on Nvidia’s infrastructure.According to Nvidia enterprise VP Manuvir Das, the introduction of new software marks a significant shift towards a commercial software business model, making it easier for developers to utilize Nvidia GPUs across various applications.
Das highlighted the accessibility of Nvidia’s new software, ensuring compatibility with a wide range of GPUs to reach a broader audience of developers and users.
Nvidia Blackwell
Nvidia’s Blackwell-based processors, used by the multi-node, liquid-cooled, rack-scale system – GB200 NVL72, represents a significant leap in performance for AI companies, boasting an impressive 20 petaflops in AI performance compared to the 4 petaflops offered by the H100. This substantial increase in processing power will empower AI companies to tackle larger and more complex models, according to Nvidia. Key to the Blackwell chip is its specialized “transformer engine,” purpose-built to efficiently run transformers-based AI, a foundational technology crucial for ChatGPT and similar models.
The Blackwell GPU is notably sizable, integrating two separately manufactured dies into a single chip, all manufactured by TSMC. Moreover, Nvidia will offer the Blackwell GPU as part of an entire server package called the GB200 NVLink 2, which combines 72 Blackwell GPUs with other Nvidia components tailored specifically for training AI models.
Major cloud service providers, including Amazon, Google, Microsoft, and Oracle, are set to offer access to Nvidia’s GB200 through their cloud services. The GB200 configuration includes two B200 Blackwell GPUs paired with one Arm-based Grace CPU, facilitating powerful computing capabilities for AI applications.
Nvidia announced that Amazon Web Services (AWS) plans to construct a server cluster equipped with 20,000 GB200 chips, demonstrating a significant commitment to leveraging Nvidia’s latest technology for cloud computing.
The GB200 system is capable of deploying a 27-trillion-parameter model, surpassing the scale of even the largest existing models, such as GPT-4, which is reported to have 1.7 trillion parameters. This unprecedented scale suggests that larger models with more parameters and data could potentially unlock new capabilities, a notion supported by many artificial intelligence researchers.
Nvidia GB200 Blackwell Price
Nvidia has not disclosed the pricing details for the new GB200 chip or the systems incorporating it. However, analysts estimate that Nvidia’s previous Hopper-based H100 chip ranges in cost between $25,000 and $40,000 per chip. Entire systems utilizing these chips can cost as much as $200,000, according to these estimates. Until Nvidia officially announces the pricing for the GB200 and associated systems, these estimates provide a reference point for the potential cost of the new hardware.
NIM – Nvidia Inference Microservice
Along with the GB200 chip, Nvidia also introduced a new product called NIM, or Nvidia Inference Microservice, as part of its Nvidia enterprise software subscription. NIM streamlines the utilization of older Nvidia GPUs for inference tasks, which involve running AI software.
This allows companies to leverage their existing investment in Nvidia GPUs, potentially saving costs. Inference typically requires less computational power compared to the initial training of AI models. NIM empowers companies to run their own AI models rather than relying on AI services from external providers like OpenAI.
The aim is to encourage customers who purchase Nvidia-based servers to subscribe to Nvidia enterprise, priced at $4,500 per GPU per year for a license.Nvidia will collaborate with AI companies such as Microsoft or Hugging Face to optimize their AI models for compatibility with Nvidia chips.
Developers can then efficiently deploy these models on their own servers or cloud-based Nvidia servers using NIM, streamlining the process without extensive configuration requirements. According to the company, the software will also facilitate AI operations on GPU-equipped laptops, offering flexibility in deployment options beyond cloud-based servers.
