ATLAS: The First Full-Stack Performance Evaluation Framework for 3D-DRAM Large Language Model Accelerators

ATLAS is the first full-stack simulation framework for 3D-DRAM large language model accelerators validated with real silicon, filling the gap in the field where public performance evaluation tools were missing. Built on commercial 3D-DRAM technology, it provides an open, universal, and high-precision performance analysis platform that supports any inference scenario, helping researchers conduct design space exploration and promoting the development and ecosystem formation of 3D-DRAM accelerator technology.

Large language model inference (especially the decoding phase) is memory-intensive, making bandwidth a key bottleneck; 3D-DRAM has become an ideal choice due to its high bandwidth density and energy efficiency ratio. However, current 3D-DRAM accelerators rely on closed-source evaluation tools, leading to fragmented modeling and results that are difficult to compare, which hinders technological progress.

ATLAS is built based on the characteristics of commercialized 3D-DRAM silicon chips and introduces a unified abstraction mechanism: at the system architecture level, it defines standardized component interfaces and interconnection models; at the programming primitive level, it provides general computing and storage operation abstractions, shielding hardware differences and supporting scenarios such as LLMs of different scales, single-user low-latency, and high-throughput batch processing.

ATLAS has been validated with silicon, with a simulation error ≤8.57% and a correlation coefficient with measured performance ranging from 97.26% to 99.96%. Design space exploration reveals that there is an optimal range for the ratio of memory bandwidth to computing units, and different batch sizes require adjusting the 3D-DRAM hierarchical scheduling strategy to leverage the high-bandwidth advantage.

The research team will open-source the ATLAS framework to break closed-source barriers and allow more researchers to participate; iteratively improve functions through community efforts; establish unified evaluation benchmarks to promote fair competition and cooperation, and drive the maturity of the field.

ATLAS marks a new stage in the research of 3D-DRAM LLM accelerators—from relying on closed-source tools to an open platform, from fragmented modeling to unified abstraction, from speculative design to data-driven optimization. It will promote the technology to find a better balance among performance, energy efficiency, and cost, paving the way for the inclusive application of LLMs.