ArchAgent: The AI Architect Redefining the Future of Supercomputing

The Complexity Paradox in Modern Computing

The era of generative Artificial Intelligence (AI) and High-Performance Computing (HPC) has brought an unprecedented challenge for system architects. As Large Language Models (LLMs) grow in parameters, the infrastructure required to train and run them becomes exponentially complex. Today, designing a supercomputer is not just about choosing a processor; it is about navigating an ocean of variables including compute architectures (GPUs, RISC-V accelerators), memory technologies (HBM, SRAM), and interconnect protocols (NVLink, InfiniBand, Ultra Ethernet).

As researchers Murali Emani and Kashyap R Manjusha observe, no single architecture is optimal for all tasks simultaneously. What works for training an LLM may be inefficient for inference or coupled scientific simulations. The resulting design space is so vast that manual exploration has become impractical, demanding a new approach that utilizes AI itself to design the next generation of hardware.

ArchAgent: A Hierarchy of Specialists

To resolve this impasse, Murali Emani and Kashyap R Manjusha developed ArchAgent, a hierarchical multi-agent framework based on LLMs. The brilliance of the system lies in its division of labor. Instead of trying to solve the design problem as a single block, ArchAgent uses a “Planner Agent” that decomposes the problem into three critical domains: chip, scale-up (within the node), and scale-out (across nodes).

Each of these domains is assigned to a specialist agent. The Chip Agent focuses on microarchitecture, deciding on core counts and local memory types. The Scale-Up Agent manages how these components communicate within a single server, optimizing bandwidths like those of NVLink. Finally, the Scale-Out Agent designs the large-scale network, handling latencies and global interconnect topologies. This structure allows the system to explore the design space much more intelligently than exhaustive search algorithms or traditional optimization methods.

The Analytical Feedback Loop

One of the greatest risks in using LLMs for engineering is the tendency toward hallucination—generating data that looks correct but is physically impossible. Murali Emani and Kashyap R Manjusha mitigated this risk elegantly: ArchAgent does not “guess” performance. The framework closes the feedback loop with a deterministic analytical evaluator.

Whenever the agents propose a new system configuration, that proposal is submitted to rigorous mathematical models that calculate actual performance and power consumption. The LLM receives these numerical results and uses them to refine its next suggestion. This process ensures that all performance estimates are derived from validated analytical models rather than probabilistic model predictions. It is the union of generative creativity with the rigor of computational physics.

Results and Energy Efficiency

The results obtained by the research duo are impressive. In a mixed workload scenario (AI + HPC) under a strict 1 MW power budget, ArchAgent discovered heterogeneous designs that achieved a 1.93x weighted geometric mean speedup over a homogeneous baseline based on traditional GPUs.

Even more notable is the efficiency of the discovery process. While conventional optimization methods, such as “hill climbing,” required 100 evaluations to reach a 1.65x speedup, ArchAgent needed only 18 evaluations to find a superior solution. This demonstrates that agentic AI not only finds better solutions but does so drastically faster, saving precious computational time in the design process.

Impact on the Deep Tech Industry

The research by Murali Emani and Kashyap R Manjusha signals a paradigm shift in system architecture. We are moving from an era of general-purpose hardware to an era of highly specialized and heterogeneous systems, where different types of processors and memories coexist to maximize efficiency.

ArchAgent proves that agentic AI can act as a hardware co-designer, allowing engineers to explore frontiers that were previously invisible. For the Deep Tech sector, this means accelerating the innovation cycle for chips and supercomputers, reducing costs, and, crucially, addressing the challenge of energy sustainability in data centers. The future of computing will not just be built by humans, but orchestrated by intelligent agents that understand the complexity of silicon better than any previous tool.

Scientific Source: Agentic AI for Heterogeneous System Design Across Chip, Scale-Up, and Scale-Out Domains — Authors: Murali Emani, Kashyap R Manjusha. Available on arXiv.