International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 09 | Sep 2025
p-ISSN: 2395-0072
www.irjet.net
ML-Surrogate: An Intelligent Design Space Exploration Framework for Parameterizable GPUs Selva Lakshman Murali, Sanjana Ramesh -------------------------------------------------------------------------***------------------------------------------------------------------------Abstract – The end of Dennard scaling and the slowing of Moore's Law have ushered in an era of domainspecific hardware, where architectures are tailored for specific applications. Parameterizable GPU generators are at the forefront of this movement, offering the potential to create bespoke processors. However, this flexibility presents a combinatorial explosion of design choices, making design space exploration (DSE) a primary bottleneck. Traditional DSE, relying on manual tuning or exhaustive simulation, is prohibitively time-consuming and often fails to uncover optimal configurations. This paper proposes ML-Surrogate, a novel framework that leverages machine learning to intelligently and efficiently navigate the vast DSE landscape for GPUs. We train a gradient boosting ensemble to create high-fidelity surrogate models that predict key performance indicators—performance (GFLOPS), power (Watts), and area (mm2)—from a given set of architectural parameters. These predictive models are then integrated within a Bayesian Optimization loop to rapidly identify Pareto-optimal GPU designs. Using the open-source Vortex GPU platform as our testbed, we demonstrate that ML-Surrogate can identify superior designs using up to 80% fewer hardware evaluations (costly simulation and synthesis runs) compared to a traditional random search, effectively transforming a months-long exploration process into a matter of days.
I.
Graphics Processing Units (GPUs) represent one of the most successful parallel architectures, and their programmability has made them indispensable for scientific computing, machine learning, and data analytics. The logical next step in this evolution is the creation of application-specific GPUs. Hardware generators, such as the RISC-V Rocket Chip generator [2] for CPUs, have provided a blueprint for creating flexible hardware. This has inspired similar efforts in the GPU space, leading to powerful, open-source, and parameterizable platforms like Vortex [3] and Nyuzi [4]. These generators allow a designer to specify high-level parameters—such as the number of cores, cache sizes, or issue widths—and automatically produce synthesizable Register-Transfer Level (RTL) code. While powerful, this flexibility presents a formidable challenge: the design space is astronomically large. A moderately complex generator can easily have 10-15 tunable parameters, leading to trillions of possible hardware configurations. Manually exploring this space is impossible. The de facto standard, running extensive simulations on a grid of randomly selected points, is computationally brutal and offers no guarantee of finding globally optimal designs. This verification and exploration bottleneck is arguably the single greatest barrier to realizing the full potential of generated hardware. To break this impasse, we require a more intelligent approach. This paper introduces such an approach: the ML-Surrogate framework. The core idea is to replace the expensive, slow process of hardware evaluation (simulation and synthesis) with a fast, accurate machine learning model. This "surrogate" can be queried thousands of times per second, allowing us to use sophisticated optimization algorithms to search the design space efficiently.
INTRODUCTION
For decades, the semiconductor industry has been propelled by the predictable cadence of Moore's Law and Dennard scaling, delivering ever more powerful and efficient general-purpose processors. This paradigm is now facing fundamental physical limits [1]. As performance gains from simply shrinking transistors diminish, the industry is pivoting towards architectural specialization. Domain-Specific Architectures (DSAs), from Google's TPUs to custom accelerators for deep learning, have demonstrated that tailoring hardware to a specific workload can yield orders-of-magnitude improvements in performance and energy efficiency.
© 2025, IRJET
|
Impact Factor value: 8.315
Our contributions are threefold: 1.
|
We present a complete, closed-loop framework, ML-Surrogate, that integrates a state-of-the-art GPU generator with an ML-driven DSE engine.
ISO 9001:2008 Certified Journal
|
Page 184