March 1, 2011 - Technology media outlet smartprix published a blog post yesterday (February 28) reporting that OpenAI In a video introducing the GPT-4.5 model, the accidental leak of the GPT-6 Training may be required GPU numbers, suggesting that it will be much larger than ever before.
Note: At the 2:26 mark of the GPT-4.5 model introduction video, the words "Num GPUs for GPT 6 Training" appear in the chat transcript of OpenAI's demonstration of GPT 4.5 capabilities.
Although the video does not explain this in any way, the"Num" may suggest an unprecedented number, with the outlet speculating as high as 100,000 GPUs.
Previously, it was reported that OpenAI used about 10,000 GPUs in training GPT-3, and as the model continues to be iterated, the demand for computational resources continues to increase.
GPT 4.5, internally codenamed "Orion", has made significant progress in terms of naturalness and reduction of "illusions", and its number of parameters may exceed 3 to 4 trillion. It is estimated that GPT 4.5 was trained using 30,000 to 50,000 NVIDIA H100 GPUs, at a cost of about $750 million to $1.5 billion (IT Home Note: Currently about 5.469 billion to 10.937 billion yuan).
The meaning of "Num" in the screenshot is unclear, but it could stand for "Numerous", suggesting that GPT-6 is being trained on a much larger scale than ever before. Of course, it's also possible that this is just a smokescreen for OpenAI, like the previous use of "Strawberry" as a codename for the o1 series.