Memory d-Matrix 3D DRAM (3DIMC) for AI

Startup d-Matrix has announced that it is developing a new type of memory designed for AI tasks.

3D digital computing technology

The 3D digital computing technology in the new 3DIMC memory is designed to accelerate data output by AI systems significantly. At this stage, HBM is used for all AI-related tasks.

HBM functions by placing memory modules on top of each other, thus achieving a fairly high level of performance and effective connection of memory crystals. This type of memory performs its tasks, but it is not the best option for computing.

d-Matrix is currently testing the Pavehawk 3DIMC silicon chip. At the moment, the hardware for in-memory computing consists of LPDDR5 memory chips and DIMC chiplets installed on top, which are connected via an interposer.

This configuration makes it possible to perform calculations directly in memory. AI models are evolving rapidly, and traditional memory systems limit bandwidth and are becoming increasingly expensive. Thus, the new approach, where memory is better integrated with computing systems, increases bandwidth and overall efficiency.

Pavehawk 3DIMC is being tested in laboratory conditions, and a new generation of Raptor memory is also expected to be released, which will also be based on the chiplet model. Raptor is expected to accelerate AI data output by almost 10 times compared to traditional memory systems.