ScienceGPT it is a model of artificial intelligence developed and trained with the computing power and data of one of the best supercomputer internationally.
Its official name is AuroraGPT and refers directly to supercomputer Aurora, still under development. To date there is not much confirmation on whether this works Large Language Model (LLM).
At the same time i Aurora technical data and the mass of data used to train the AI they are so impressive that you can imagine extraordinary performances.
What is ScienceGPT, the AI model of the Aurora supercomputer
American researchers fromArgonne National Lab I’m working on an AI model based on Aurora scientific data: one of the fastest supercomputers in the world, with a computing power of 2 ExaFLOPS.
This means that the supercomputer is capable of performing more or less a quintillion calculations per second, thanks to its over nine thousand knots. For the rest Aurora is equipped with 10 petabytes of memory and boasts one 230 petabyte storage space. The petabyte is a computer measurement unit that corresponds to one million GB.
The system is also composed of two Intel Xeon Max processorsa unified memory architecture and well sei GPU Intel Max. Impressive numbers, which help to better understand a monster consumption of 60 MW.
The AI model nicknamed ScienceGPT it’s actually called AuroraGPT. A system of generative artificial intelligence which uses data from the Aurora supercomputer and boasts approx a trillion parameters: another record figure, which makes this LLM among the most impressive at an international level.
How ScienceGPT works
According to the statements of the vice president and general manager HPC Wire Ogi Brkicuser-side ScienceGPT will work quite similarly to ChatGPT.
But it still remains to understand if the model will be multimodal: that is, whether it will be able to support the creation of multimedia content to the generation of original text.
Added to this are at least a couple more elements of uncertainty. The first is related to LLM training times, which will still take months. Also because, at present, operations are limited to only 256 of Aurora’s approximately 10,000 nodes.
The second is related to development of the same supercomputer, considering that work on Aurora is also underway. And that, originally, the completion of the system had been planned for 2018.
Looking at the design data, Aurora is destined to become the second most powerful supercomputer ever after Frontier. The Datasheet available they speak of power from 60,000 Intel GPUs, 10,000 processing nodes, 166 racks and more than 80,000 network nodes.
At the same time, several insiders imagine that, once completed, the system it will even be superior to Frontier. And in this case it is reasonable to think that too ScienceGPT will far exceed expectations.
To know more: Artificial Intelligence: what it is and what it can do for us