Hybrid Analog–Analog Reservoir Computing: Bridging GPU Firmware Physics and Memristive Neuron Dynamics

 We present a hybrid analog–analog architecture for neuromorphic reservoir computing in which both a memristive
  neuron substrate (128 NS-RAM-modelled LIF neurons on FPGA) and the host GPU (AMD Radeon 8060S, RDNA4) contribute
  their native physics to computation. GPU firmware noise layers (VRM 1/f noise, SMN thermal fluctuations, kernel
  jitter, clock-domain crossing artifacts) are injected as neuromodulatory current into the FPGA neuron bank via a
  bidirectional UDP Ethernet bridge at 2 kHz.

  Across 83 experiment groups (619 tests), we demonstrate: 81% waveform classification with 128 neurons;
  self-organised criticality (σ=1.027, driven 27× closer to critical by GPU 1/f noise than white noise); causal
  emergence (2.87× effective information ratio); directed cross-substrate information flow (0.122 bits transfer
  entropy); a 7-level substrate comparison ladder showing cross-substrate fusion achieves the best temporal
  regression (NARMA-10 NRMSE 28% better than FPGA alone); and a GPU-only neuromorphic reservoir exploiting seven
  microarchitectural mechanisms (branch divergence, LDS bank conflicts, TLB persistence, PLL jitter, wavefront
  scheduling, atomic serialisation, memory coalescing) that achieves 97.7% waveform classification in a
  four-population architecture.

  The platform is designed for hardware substitution: replacing the FPGA neuron model with real NS-RAM devices
  requires adapting the physical interface, not the analysis pipeline. All RTL, bridge code, and a reservoir
  computing demo are released as open source.