When a DSP Outperforms a Hardware Accelerator

Embedded CPUs became ubiquitous because they combine flexibility, solid performance, low power consumption, and cost‑effectiveness. Switching from a separate microprocessor or microcontroller to an embedded CPU was a clear win for most applications. Yet, no CPU can run every algorithm efficiently—software can grow arbitrarily complex, and execution time or power budget may become unacceptable.

Source: CEVA

To address this, manufacturers introduced hardware accelerators—dedicated circuits that perform common operations, like floating‑point math, far faster than software on a general‑purpose CPU. The concept quickly expanded to include cryptography, regular‑expression handling, graphics, and more.

While accelerators deliver speed, they sacrifice the flexibility that software offers. They are hard‑wired, so modifying an algorithm, fixing a bug, or adapting to market changes requires redesigning hardware—a costly and time‑consuming process.

Ideally, we want the best of both worlds: accelerate algorithms while still writing them in software. DSPs (Digital Signal Processors) can achieve this for many common, stream‑based tasks.

DSPs excel in streaming data operations such as audio filtering, PDM‑to‑PCM conversion, acoustic echo cancellation, and stream‑based ciphers like SNOW and ZUC used in LTE. In signal‑processing contexts, they handle complex matrix calculations for channel estimation between base stations and mobile devices, optimizing transmissions for reliability. More broadly, DSPs suit any application that benefits from wide parallelism, such as AES cryptography.

DSPs are particularly powerful for streaming computation, complex math (matrices, floating point), and high parallelism—areas where they can rival or surpass hardware accelerators. In many cases, a DSP implementation is smaller and cheaper, and its power consumption is still far below a CPU‑based solution. A single DSP can host multiple acceleration functions that need not run simultaneously, eliminating the need for several dedicated accelerators. For even greater throughput, multi‑core DSPs provide the same scalability as multi‑core CPUs.

Source: CEVA

Perhaps the greatest advantage is programmability. Like a CPU core, a DSP can be programmed in C, allowing developers to adapt or upgrade software without hardware changes. With a capable compiler and simulator, optimizing for parallelism becomes manageable, enabling rapid bug fixes, feature updates, and improved customer satisfaction—all while maintaining low power and cost.

Another benefit is multi‑function support. For example, Global Navigation Satellite System (GNSS) processing benefits from DSPs. In mobile devices, adding GNSS via software becomes possible; in fixed devices, a DSP‑based GNSS can run during idle periods, saving area and power compared to a dedicated hardware unit.

DSPs are not a universal replacement for all accelerators. Some functions, such as very large filters, may still favor hard‑wired solutions. Nonetheless, for a wide range of algorithms, DSPs approach or match accelerator performance and power, often at lower cost and with far greater flexibility. They are a compelling option worth serious consideration.

This article is the second in a series that began with “Why DSPs are Suddenly Everywhere” and will be followed by “Decisions, Decisions: Hardware Accelerator or DSP?”.