Since the 1960s, three distinct architectures have been used to accelerate computational tasks for DSP systems: Microprocessors, Field Programmable Gate Arrays (FPGA), and Coarse Grained Reconfigurable Arrays (CGRA), all with variations optimizing the problem domain with specialization , parallelism , and configurability .

Early DSP history was myopically focused on specialization in Microprocessor architectures primarily due to limited area on die . The first single-chip DSP, the TMC 0280, was developed in 1978 with a dedicated multiply accumulate (MAC) unit , and dedicated complex operators are a mainstay of DSP architectures to this day. The TMS 32010 adopted the Harvard Architecture in 1982 to satisfy intensive IO bandwidth requirements , and numerous variations appeared shortly thereafter . The DSP 32 added floating point arithmetic to deal with data scaling issues in 1984 , and the DSP 56001 found a better solution in 1987 with saturating fixed-point arithmetic on a wide datapath . The DSP 32 also added register indirect addressing modes to compress memory addresses in the instruction words, and the DSP 56001 implemented circular buffers in memory to optimize delay computations.

With shrinking process technology nodes yielding more transistors on die, DSP architectures shifted focus toward parallelism . The TMS320C20 had a pipelined datapath to target data parallelism in 1993 . In 1996, the TMS320C8x added multiple cores to optimize task parallelism . Then in 1997, the DSP16xxx introduced a two lane pipelined Very Long Instruction Word (VLIW) architecture .

In the 2000s, the DSP market saw a fundamental shift. First, Intel introduced DSP extensions for their general purpose processors targeting non-embedded applications in 1999 . Second, Xilinx introduced FPGAs to the DSP market with the development of the Xilinx Virtex-II targeting embedded high-performance applications in 2001 . While difficult to program, FPGAs are much more flexible, have orders of magnitude better performance and energy consumption, and may be reconfigured in the field. As a result, specialized microprocessor DSP architectures were relegated to embedded low-performance problem domains. Since then, FPGA innovations have focused on application specific operator integration and network optimization , ease of use , embedded and non-embedded system integration , and run-time and partial reconfigurability .

While the dominance of FPGAs has demonstrated that array architectures are the right solution for the problem domain, CGRAs show the potential for significant improvements across the board . Historically, bit-parallel CGRAs have extremely limited capacity due to routing resource requirements. Digit-serial CGRAs solve the capacity issues by reducing the width of the datapath. However, they also sacrifice configurability in the face of complex timing and control requirements. This has led to a variety of systolic array architectures that accelerate extremely specific computational tasks. However, solving these configurability issues could open the door to a diverse set of new capabilities on mobile platforms.

The concepts introduced by Von Neumann in 1945 , remain the centerpiece of computer architectures to this day. His programmable model for general purpose compute combined with a relentless march toward increasingly efficient devices cultivated significant long-term advancement in the performance and power-efficiency of general-purpose computers. For a long time, chip area was the limiting factor and raw instruction throughput was the goal, leaving energy largely ignored. However, technology scaling has demonstrated diminishing returns, and the technology landscape has shifted quite a bit over the last 15 years.

Around 2007, three things happened. First, Apple released the iPhone opening a new industry for mobile devices with limited access to power. Second, chips produced with technology nodes following Intel's 90nm process ceased scaling frequency () as the power density collided with the limitations of air-cooling (). For the first time in the industry, a chip could not possibly run all transistors at full throughput without exceeding the thermal limits imposed by standard cooling technology. By 2011, up to 80% of transistors had to remain off at any given time .

Third, the growth in wire delay relative to frequency introduced new difficulties in clock distribution. Specifically, around the introduction of the 90nm process, global wire delay was just long enough relative to the clock period to prevent reliable distribution across the whole chip ().

As a result of these factors, the throughput of sequential programs stopped scaling after 2005 (). The industry adapted, turning its focus toward parallelism. In 2006, Intel's Spec Benchmark scores jump by a 135% with the transtion from NetBurst to the Core microarchitecture, dropping the base clock speed to optimize energy and doubling the width of the issue queue from two to four, targeting Instruction Level Parallelism (ILP) instead of raw execution speed of sequential operations . Afterward, performance grows steadily as architectures continue to optimize for ILP. While Spec2000 focused on sequential tasks, Spec2006 introduced more parallel tasks .

By 2012, Intel had pushed most other competitors out of the Desktop CPU market, and chips following Intel's 32nm process ceased scaling total transistor counts. While smaller feature sizes supported higher transistor density, it also brought higher defect density () causing yield losses that make larger chips significantly more expensive ().

Today, energy has superceded area as the limiting factor and architects must balance throughput against energy per operation. Furthermore, improvements in parallel programs have slowed due to a combination of factors (). First, all available parallelism has already been exploited for many applications. Second, limitations in power density and device counts have put an upper bound on the amount of computations that can be performed at any given time. And third, memory bandwidth has lagged behind compute throughput, introducing a bottleneck that limits the amount of data that can be communicated at any given time () .