Alibaba plans model of 5 to 10 trillion parameters, launches AI chip
Chief executive Eddie Wu set out the plan at Alibaba Cloud's Apsara Conference in Hangzhou, alongside a new accelerator and a 20 gigawatt data centre target.
Alibaba plans to train a model with between 5 trillion and 10 trillion parameters, chief executive Eddie Wu said Tuesday at the company's Apsara Conference in Hangzhou. Reuters reported the remarks. Wu gave no timetable for the training run and did not say what hardware would carry it.
Alibaba's chip unit T-Head announced an accelerator called the Zhenwu V900 at the same event. Bloomberg reported that Alibaba puts its performance at three times that of the previous part, and says a single cluster can scale to 500,000 of them. Both figures are the company's own.
Wu also said Alibaba Cloud aims to hold more than 20 gigawatts of data centre capacity worldwide by 2032, according to a market wire carrying his remarks. Bloomberg reported the same target. Alibaba describes the Zhenwu V900 as the most powerful AI chip in China.
No independent testing of the chip supports that description. Alibaba has not published a specification sheet for the Zhenwu V900, and the three-times figure is measured against a part the company also makes. Fast market wires carried the chip's designation inconsistently on Tuesday, with one reporting it as the V300.
The parameter range is a statement of intent rather than a shipped model. Wu did not say whether the count refers to total parameters or to the smaller number active on each token, a distinction that changes what the figure means. Alibaba has not said when training would begin.