What is it about?

Turquoise is a simulator for specialized hardware accelerators for sparse tensor algebra, a type of computation widely used in scientific computing, graphics, artificial intelligence, and machine learning. Designing specialized hardware accelerators is a complex process that requires evaluating many possible architectures using simulators. The tool considered the state-of-the-art for this purpose is TeAAL. However, TeAAL can take several days to evaluate a single candidate design, making this exploration process very time-consuming. While existing software simulators incur significant overhead from emulating hardware behavior, Turquoise maps a simplified version of the architecture directly onto an FPGA, a type of reconfigurable hardware commonly used to implement specialized accelerators. This approach allows Turquoise to perform simulations that would take TeAAL several days in just a few minutes.

Featured Image

Why is it important?

Linear algebra is at the heart of many modern computing applications, from scientific simulations and engineering to computer graphics, artificial intelligence, and machine learning. Matrix and tensor operations are fundamental to processing data, solving complex problems, and training and running AI models. As these applications continue to grow in complexity and scale, performing these computations efficiently becomes increasingly important. General-purpose processors (CPUs) often cannot deliver the performance these increasingly demanding applications require while maintaining reasonable energy efficiency. Graphics processing units (GPUs) offer greater parallelism, but they can consume substantial amounts of power and are not always well suited to sparse computations, where much of the data is zero and processing it can be wasteful. Domain-specific accelerators (DSAs) address these limitations by tailoring hardware to specific computational workloads, optimizing how operations are performed and how data is stored and moved. Finding the most effective accelerator design, however, is not straightforward. Different architectural choices, such as the number of processing units, memory organization, dataflow, and caching strategies, can significantly affect performance. Design space exploration (DSE) allows researchers to systematically evaluate different configurations and identify the most promising designs. This exploration is only useful if it can be performed at a reasonable cost. Evaluating a single design can already take hours or days with traditional software simulators, making it impractical to investigate a large number of alternatives. Faster simulation tools like Turquoise enable researchers to explore more architectures in less time, accelerate the development of specialized hardware, and make better-informed design decisions, ultimately helping bring more efficient computing systems to real-world applications.

Perspectives

The core idea of Turquoise is to take advantage of FPGAs to simulate sparse tensor algebra accelerators directly in reconfigurable hardware. Rather than relying on software to emulate the behavior of the target architecture, Turquoise implements a simplified version of that architecture on an FPGA. This reduces the abstraction layers between the architecture being studied and the hardware executing the simulation, enabling faster and more direct evaluation of different accelerator designs.

Francesco Pesce
Politecnico di Milano

Read the Original

This page is a summary of: Turquoise: A Hardware-Accelerated Architectural Simulator for Sparse Algebra DSAs, ACM Transactions on Architecture and Code Optimization, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3860332.
You can read the full text:

Read

Contributors

The following have contributed to this page