and data center energy consumption and storage changes

Deepseek's efficient AI training has aroused widespread discussion in the AI community and led to fluctuations in AI-related stocks. However, we should not be surprised by the progress Deepseek has made in development. The various computing, network, memory and storage technologies that promote AI training today have a long history of innovation, thus improving efficiency and reducing power consumption.

These trends will continue in both hardware and software, enabling data centers to get twice the result with half the effort. They will also enable more organizations to easily carry out AI training, use existing data centers to achieve more functions, and drive the growth of digital storage and memory to support more AI training.

Behind the data center growth forecast is that it is estimated that data centers that will perform heavy AI tasks in the future may require several gigawatts (GW) of power consumption. This is equivalent to the estimated consumption of 5.8 in San Francisco, California giva the power is equivalent. In other words, it is expected that the power demand of a single data center will be comparable to that of a big city. This prompted the data center to consider autonomous power generation, using renewable and non-renewable energy sources, including modular nuclear reactors.

What if we can improve the efficiency of AI training and reasoning in future data centers, thus slowing down the expected power consumption growth? More efficient AI training methods such as Deepseek can make AI training easier to deploy and complete more training with less energy consumption.

Deepseek implements efficient training by adopting the "Mixture of Experts" architecture. Compared with other AI models, Deepseek significantly reduces resource usage. In this architecture, a special sub-model processes different tasks, effectively allocates computing loads, and activates relevant parts of the model only for each input, this reduces the need for large amounts of computing power and data. This method is combined with technologies such as intelligent memory compression and training of only the most critical parameters, enabling them to achieve high performance with less hardware, shorter training time and lower power consumption.

More efficient AI training will reduce the investment in creating new models, thus enabling more organizations to carry out more AI training. Even if the training data is compressed, more models mean that more storage and memory are needed to store the training data. With the support of more efficient AI training, the demand for AI digital storage will continue to grow. There may be more room for efficiency improvement in AI training. In addition to the methods used by Deepseek, the further development of AI training methods and algorithms can help us limit the future AI energy demand.

A report recently released by the U.S. Department of Energy (compiled by Lawrence Berkeley National Laboratory) studied the historical trend and forecast of power consumption in U.S. data centers from 2014 to 2028, as follows: until around 2018, the percentage of total power generation consumed by the data center has been fairly stable, less than 2%. The growth trend of cloud computing, especially all kinds of AI, has pushed the power consumption to 4.4% in 2023. It is estimated that this proportion will increase to 6.7 to 12.0% by 2028. This growth may bring heavy pressure to our power grid.

During the period before 2018, although computing and other data center activities increased, architecture and software changes (such as virtual machines and containers) as well as the high efficiency achieved by dedicated processing and new expansion and network technologies, it can limit the total energy consumption of the data center.

AI and other growing computing applications require more and more digital storage and memory to store the data being processed. Storage and memory consume electricity. The following figure in the Department of Energy report shows the estimated value and forecast trend of digital storage energy consumption in data centers from 2014 to 2028.

The chart is based on IDC data and shows that the growth has been faster since 2018. It is estimated that the power consumption will increase by about 2 times by 2028, most of which comes from SSD based on NAND flash memory. This may be due to the growth of SSD in data center applications, especially primary storage, because of their higher performance, however, this increase may be due to more intensive writes and reads to SSDs to support AI and similar workflows, write and read operations on an SSD consume more energy than when you do not access the SSD.

HDD is increasingly used for secondary storage for data retention (data will not be processed immediately at this time). Even if the total storage capacity of these devices increases, their energy efficiency becomes higher and higher. As a result, SSD may account for nearly half of data center storage energy consumption by 2028.

However, the increase in storage and memory power consumption in these predictions is much lower than the increase in power consumption required for AI model GPU processing. New storage and memory technologies, such as memory and storage pooling and software management storage Allocation, which may create more efficient storage and memory usage methods for AI applications, thus helping to implement more efficient AI modeling.

Deepseek and similar more efficient AI training methods can reduce the power demand of data centers, make AI modeling easier to deploy, and increase data storage and memory requirements. Even if it is possible to achieve higher efficiency, this will help improve the sustainability of the data center.

references: Coughlin, Thomas. "Deepseek AI Will Increase Data Storage And Make AI More Accessible." Forbes, 6 Feb. 2025, https:// www. Forwarded. Accessed 8 Feb. 2025.

The above content is from Andy730 public account