Contact
AI/ML Hardware Accelerators

Dataflow Architectures for Neural Networks

How data moves through an accelerator's compute array – its 'dataflow' – is one of the most fundamental architectural choices in AI accelerator design, and it significantly affects both performance and energy efficiency.

VIDYUTT• July 22, 2026• 2 min read
Figure 1: Different dataflow strategies – weight stationary, output stationary, and row stationary – optimize for different reuse patterns.
Figure 1: Different dataflow strategies – weight stationary, output stationary, and row stationary – optimize for different reuse patterns.

What Dataflow Architectures Involve

In a weight-stationary dataflow, weights stay fixed in each processing element while activations stream through, minimizing weight-related memory traffic. In an output-stationary dataflow, partial sums stay fixed in place while weights and activations stream through, reducing the energy cost of accumulating results. Row-stationary and other hybrid approaches try to balance reuse across weights, activations, and partial sums simultaneously, rather than optimizing for just one.

Why It Matters

The best dataflow choice depends heavily on the specific characteristics of the target neural network layers – their filter sizes, channel counts, and reuse opportunities. An accelerator optimized for one dataflow pattern may perform poorly on network architectures that don't match its assumptions, making this an important early architectural decision tied closely to the intended workload.

More in AI/ML Hardware Accelerators

View all
AI/ML Hardware Accelerators Aug 1, 2026 • 2 min read Quantization and Precision Trade-offs in Hardware Neural networks don't necessarily need full 32-bit floating point precision to produce useful results. Quantization – using lower-precision number formats – is one of the most effective techniques for reducing an accelerator's memory footprint and energy consumption. Read article AI/ML Hardware Accelerators Jul 27, 2026 • 2 min read Scaling AI Accelerators: From Edge to Data Center AI acceleration isn't one-size-fits-all – the design considerations for a battery-powered edge sensor are almost entirely different from those for a data center training cluster, even though both may run similar neural network algorithms. Read article AI/ML Hardware Accelerators Jul 23, 2026 • 2 min read Memory Bottlenecks in AI Accelerators: The Real Constraint Raw compute throughput (TFLOPS) gets most of the attention in AI accelerator marketing, but in practice, memory bandwidth is often the real constraint that determines an accelerator's actual, achieved performance. Read article
Build your idea into silicon.
From RTL to real, packaged chips through the VACS program.
Start Your Silicon Journey