DATE: AUG 16 2023/ SOURCE: YICAI
(Yicai) Aug. 16 -- Chinese voice recognition firm iFlytek and Huawei, a telecoms giant, have released a hardware device called SparkDesk to help companies build their own large language models.
The device, which combines hardware and software, can be used to train an LLM. "Every enterprise, every industry, every school, and every hospital can have the opportunity to build their own exclusive LLMs," per Liu Qingfeng, chairman of Hefei-based iFlytek.
SparkDesk has three options of parameters: 13 billion, 65 billion, and 175 billion, which define how input data is treated. iFlytek is building a developer ecosystem around the model, and more than 4,000 developer teams have joined the platform.
The machine involves an artificial intelligence framework, a self-developed training algorithm, a closed-loop architecture, and the ability to answer questions, generate dialogue, and make smart recommendations. The pair will jointly identify and polish the training libraries, Liu said.
iFlytek is working with Huawei to build a cluster of domestic computing power to train LLMs on an ultra-large scale, and the system is benchmarked against Nvidia’s A100 chip, the chairman said. The American chip giant's A100 is currently the most widely used semiconductor in LLM training.
The Chinese AI company intends to open the system for university research and development, Liu said, responding to a question of whether the firm will turn SparkDesk into an open-source platform.