OpenAI's "Next Big Thing": NVIDIA's Most Powerful Chip B100, OpenAI's Most Powerful Model GPT-5

After the world's most powerful AI chip, the NVIDIA H200, made its debut, the industry has begun to anticipate the even more powerful B100 chip from NVIDIA. At the same time, OpenAI, the hottest AI startup this year, has started developing a more powerful and complex GPT-5 model.

In its latest research report, Guotai Junan pointed out that the unlimited performance of the B100 and GPT-5 is expected to be released in 2024, bringing unprecedented productivity upgrades. The institution expressed optimism about AI entering a period of rapid development, with visibility continuing until 2024.

Compared to previous generations of products, how powerful are the B100 and GPT-5? NVIDIA and OpenAI have given some previews: the speed of the B100 may be more than four times that of the H100, and GPT-5 may achieve superintelligence.

In a recent special speech at the Supercomputing 2023 conference in the United States, NVIDIA compared the performance of the B100 and A100, H100, and H200 running GPT-3 175B, and came to some astonishing conclusions. Among them:

The inference speed of the H200 is 18 times that of the A100 released in 2020, and the B100 based on the next-generation Blackwell architecture, although NVIDIA's provided chart does not show specific numbers, the technology website Tweaktown compared the bar charts of the B100 and H100 and found that the B100 is at least three times faster than the H100.

As for the configuration, according to the Tweaktown article, there are reports that the B100 will embed 178 billion transistors internally, which is more than twice the current number of GPU transistors, challenging the limits of chip integration.

The B100 will use Micron's latest HBM3e memory, the same as the recently launched H200. In addition, the H200 supports up to 141GB of HBM3e memory with a memory bandwidth of up to 4.8TB/s, so it is expected that the B100 will have a similar memory configuration.

The news also states that the B100 will adopt a Chiplet design internally, which is NVIDIA's first attempt and will compete with AMD's upcoming Instinct MI300.

Is GPT close to superintelligence?

Shifting the focus to OpenAI, the hot startup in the AI industry has already begun working on a more powerful and complex GPT-5 model.

OpenAI CEO Sam Altman confirmed this in a recent media interview and claimed that it could achieve superintelligence.

However, Altman has not yet set a timetable for the release of GPT-5, nor has he determined what functionalities it may have, as it is impossible to know until it is completed.

The capabilities of GPT-4 are already impressive, but it is only approaching human-level knowledge and understanding. The next generation of AI models will not only surpass humans in terms of knowledge, but also be on par with humans in their ability to reason and process complex ideas.

This is also known as Artificial General Intelligence (AGI), which is not just about mimicking given information, but also about the ability to express new and original things.

So, the question arises: what can GPT-5 achieve? Can it help bridge the gap between humans and AGI?

Altman also mentioned that building GPT-5 would require further investment from Microsoft.

Microsoft has already invested over $10 billion in OpenAI, and Altman hopes for more funding to follow, as the path to AGI is still long and requires the construction of a significant amount of computing power.