ELAU C400/10/1/1/1/00
For example, in the marketing industry, the majority of costs come from the production of marketing materials. Generative AI can shorten the production time of a marketing video from 10 hours to 5 minutes, which is equivalent to an efficiency increase of hundreds of times. However, the cost of using digital people in video materials ranges from millions to millions of yuan per customer, while the cost of hyper realistic digital person videos is around tens of thousands of yuan per second. Based on comprehensive calculations, the cost of video marketing materials produced using generative AI may not necessarily be lower than manual production.
Behind this, on the one hand, the training and reasoning of generative AI platforms are very costly, and these costs are ultimately borne by the application side. The current most popular generative AI platform, ChatGPT, has a single training cost of up to $1.4 million. In the words of renowned computer expert Wu Jun, “ChatGPT is equivalent to scrapping 3000 Tesla cars every time it is trained.”
Among the huge costs, computing power accounts for the largest share. Due to the popularity of large models, computing power has become a scarce commodity this year. The main reason is that the production capacity of GPUs that provide computing power cannot keep up with demand, and prices continue to skyrocket. This is equivalent to an additional increase in the training cost of generative AI.
Amidst the cost and benefit gap of generative AI, Amazon Cloud Technology is attempting to reshape the cost structure of cloud computing through innovation in self-developed chips, storage, and serverless technologies, making computing power more cost-effective. In terms of chips, Amazon Cloud Technology adopts a two legged approach of self-developed and partnering with Nvidia to reduce the uncertainty caused by chip shortages and optimize costs.
Among them, the Traineum2 chip is specifically designed for basic models and large language models with trillions of parameters or variables. Compared to the first generation Traineum chip, the performance has increased by four times, memory has increased by three times, and energy efficiency (performance per watt) has increased by two times. In terms of cooperation with Nvidia, Amazon Cloud Technology became the first cloud manufacturer to equip Nvidia’s GH200 Grace Hopper superchip in the cloud. The platform of cooperation between the two parties allows joint customers to expand to thousands of GH200 superchips.
Amazon Cloud Technology’s processor Amazon Graviton4, cloud storage service Amazon S3 Express One Zone, and three serverless service innovations launched on re: Invent 2023 can all help enterprise users achieve high cost-effectiveness goals. Compared with the previous generation product, the performance of Amazon Graviton4 has improved by up to 30%, independent cores have increased by more than 50%, memory bandwidth has increased by 75%, Amazon S3 Express One Zone data access speed has increased by up to 10 times, and data request costs have been reduced by 50%.
Enterprise users pursue the robustness, economy, security, and usability of their applications, which is completely different from the approach of generative AI such as big language models, which are trained at high computational costs to achieve higher capabilities. Faced with the more pragmatic needs of enterprise users, Amazon Cloud Technology attempts to reshape the cost structure of cloud computing, enhance the practicality of generative AI from the perspective of balancing cost and benefits, and meet the needs of enterprise users.