A joint study by DeepSeek and Peking University has drawn attention with the introduction of Engram technology, which aims to fundamentally reduce memory dependency in artificial intelligence systems. The new method makes it possible to overcome the severe resource shortage in the industry by minimizing the need for high-bandwidth memory, one of the most expensive resources in training large language models.
The editorial reports that Engram technology separates memory storage from computation processes, protecting the model's GPU memory from unnecessary data load and implementing a mechanism to search and retrieve only the required information. Experts have long stated that the sharp increase in HBM demand has caused DRAM prices to increase fivefold in just 10 weeks.
The new method offers several technical advantages. Engram frees access to static data from real-time computation dependencies, allowing the model's computational power to be directed towards more complex logical processes rather than simple operations. Tests conducted on a model with 27 billion parameters have shown that this approach provides a significant performance increase in line with industry standards.
Researchers state that allocating approximately 20-25 percent of the total parameter budget to the Engram module yields better results compared to traditional models. This technology allows for increased overall memory capacity by working in parallel with more affordable SSD solutions like Phison and new CXL standards, enabling systems to operate quickly and stably without occupying expensive HBM resources.
According to experts, Engram technology can provide significant support for artificial intelligence development, especially in regions where HBM supply is difficult and expensive. This approach can also contribute to preventing sharp price fluctuations in the DRAM market and make the training of large-scale models more economically accessible.