Domain Flow vs Stored Program

Interactive animations comparing GPU-style stored program architectures with Domain Flow dataflow architectures for AI inference

The Fundamental Problem

Modern GPUs are powerful, but they pay a heavy tax for their flexibility. To handle memory latency, GPUs maintain thousands of threads, each with dedicated registers, managed by complex schedulers. Only ~12% of a GPU's energy goes to actual computation - the rest is instruction overhead.

Domain Flow architectures take a different approach. Instead of hiding latency with thread-level parallelism, they eliminate it through system-level scheduling. Data flows through compute in predictable patterns, controlled by credits rather than instructions. The result: 6-7× better energy efficiency for AI inference workloads.

GPU (Stored Program)

  • Request-reply memory access
  • 64K registers to hold waiting threads
  • Complex warp scheduler
  • Instruction fetch, decode, issue every cycle
  • Bank conflicts cause stalls
  • ~12% compute efficiency

KPU (Domain Flow)

  • Credit-based data flow
  • Small stream buffers only
  • No scheduler needed
  • No instructions - data just flows
  • Deterministic, contention-free
  • ~83% compute efficiency
6.7×
Less Energy per Op
72%
Silicon for Compute
0
Instructions to Decode
2.7×
Smaller Die Size

Explore the Three Concepts

Progressive Deep Dive

Why This Matters for AI Inference

AI inference is dominated by matrix operations: matrix multiply, convolutions, attention. These are domain-specific workloads with predictable data access patterns. GPUs treat them as general-purpose programs, paying the full instruction tax on every operation.

Domain Flow recognizes that AI inference is a data transformation problem, not a control flow problem. By scheduling data movement at the system level and flowing data through compute in predictable patterns, we eliminate the need for instruction infrastructure entirely.

The result: Same compute, fraction of the power. This enables AI at the edge, in mobile devices, and in data centers where power is the limiting factor.