This article is also available in:
French
At GTC 2024, NVIDIA unveiled the new Blackwell architecture dedicated to AI. The company promises significant power gains and energy savings.
In addition to new chips, NVIDIA also unveils AI microservices that rely on existing hardware.
NVIDIA: To Infinity, and Beyond
The NVIDIA Blackwell architecture will provide better performance. Two chips have been unveiled, the B200 and the GB200. The B200 chip provides 20 petaflops of FP4 power, with 208 billion transistors. The GB200, which NVIDIA describes as a “superchip,” is actually the assembly of two B200s and a Grace CPU.
Compared to the H100 chip, also developed by NVIDIA and already by AI companies such as OpenAI, the GB200 should provide about 7 times more computing power. According to NVIDIA, training an AI will be 4 times faster for LLMs (large language models) like ChatGPT.
What about energy consumption?
As we know, AI training requires a lot of electricity. Whichs begs the question: how much energy do these new chips require?
According to NVIDIA the GB200 will require “up to 25 times less” energy than the H100.
More power, less energy, but at what cost?
The question remains how much companies will have to spend to acquire one of these new chips. NVIDIA does not provide a figure at the moment. The H100 from the previous generation is sold between $20,000 and $40,000 per chip. We should also mention that that companies will also be able to use the new chips through various cloud services like those from AWS, Google, Microsoft, Oracle. AWS, in particular, will set up a server cluster with about 20,000 GB200 chips.
One more thing
In addition to these hardware announcements, NVIDIA unveiled NIM (Nvidia Inference Microservice), a service for developers with pre-trained models. NIM is designed to rely on existing NVIDIA CUDA GPUs. In other words, there’s no need to purchase new, expensive hardware.
It is possible to experiment with these microservices for free at ai.nvidia.com. You will then have to subscribe to NVIDIA AI Enterprise 5.0 to use these new tools.
NVIDIA CEO Jensen Huang also gave some insight as to his overall stretagy. According to him, NVIDIA will become an “AI foundry,” that is, the equivalent for the industry and businesses using AI of what TSMC does for NVIDIA by building its chips.
Takeaway
Clearly, NVIDIA is focused on remaining a key player in the rise of AI, both in terms of hardware and software. The goal is also, of course, to grow keep growing the revenue of the company and to increase the wealth of its shareholders. As a reminded, NVIDIA shares are worth about 20 times what they used to 5 years ago.
For more information about these announcements, you can checkout the video below, which features the highlights of the two-hour long keynote.
