At the GTC China 2020 Games on December 17,AliPan Yue, head of cloud heterogeneous computing products, said that demand in the artificial intelligence industry has clearly changed. The inference demand for AI services on the cloud accounts for more than 50% for the first time. From GN4 four years ago to the newly released GN7 this year, the heterogeneous computing power of a single instance has increased 100 times.
Heterogeneous computing is the calculation method that makes the most efficient use of AI. Pan Yue explained that Alibaba Cloud provides a hyperscale heterogeneous computing cluster on the cloud, which can support 10 billion heterogeneous computations per second, and can read more than 530 million images within 1 second; recognizing 330 million seconds/92,000 hours of speech is equivalent to dictation of 1,723 “A Dream of the Red Mansion” or 260,000 short stories; it can translate 40 million sentences within 1 second, which is equivalent to translating 42 “Hamlet.”
Four years ago, 80% of AI demand on the cloud came from training businesses. However, in 2020, AI inference business on the cloud already accounted for more than half. Pan Yue said that this also marks Alibaba Cloud's heterogeneous computing entering a new stage, that is, heterogeneous infrastructure on the cloud integrating software and hardware.
Specifically, users need not only deep learning framework software, but also hardware compatible with it to utilize the GPU's computing power. For example, high-speed deployment tools such as Alibaba Cloud's Dragon AI Accelerator and FastGPU bring flexible computing scheduling through pooling computing power, and cloud servers are evolving from specific configurations to serverless. These can help customers improve performance by 2x to 10x in training scenarios, 2x to 4x performance in inference scenarios, and save at least 50% of costs.
Take Alibaba Cloud IoT's image classification business as an example. Alibaba Cloud Shenlong AI accelerated the AIACC team and worked closely with the IoT intelligent business R&D team to increase the distributed training performance of large-scale image classification by 5 times, significantly improving algorithm development efficiency and drastically reducing costs.
“From the earliest inference accounting for less than 20%, the AI business now accounts for more than half; originally, only artificial intelligence startups explored and landed in various industries; heterogeneous businesses on the cloud ranged from a single deep learning training scenario to today's diverse scenarios such as training reasoning, cloud desktop, and graphic image design. Alibaba Cloud's heterogeneous computing is a witness and practitioner of the digital and intelligent transformation of enterprises.” Pan Yue said.