Real-Time Visualisation of Large-Scale Neural Networks: GPU-Accelerated Methods for Desktop, Extended Reality, and Browser Platforms

Large-scale brain research programs, including the BRAIN Initiative and the Human Brain Project, have produced a sharp rise in the volume of neural recordings. Modern electrode arrays can now capture activity from millions of neurons in a single session. Visualisation tools have not kept up. ViSimpl, probably the closest comparable single-machine tool, handles roughly 800,000 particles, and NeuroVis only reaches interactive frame rates for networks of around 200 neurons. A GPU-accelerated visualisation framework was developed in Unity 3D to enable real-time rendering of million-neuron datasets. The work is organised around five studies. In the first, NeuroConstellation, the focus is raw scalability: 11 parallel GPU compute kernels combined with indirect instancing through DrawMeshInstancedIndirect allow the desktop implementation to render all 5.28 million human hippocampal CA1 neurons at 60 to 400 FPS on a laptop-class GPU (the NVIDIA RTX 2000 Ada, with 8 GB of VRAM). All six visualisation modes and five analysis metrics execute concurrently with rendering at a 10 Hz update rate. Each metric was cross-checked against a Python reference implementation, with deviations staying below 0.1%. Second, Chapter 6 describes the port to mixed reality through HoloNeV. Deploying on Microsoft HoloLens 2 via Holographic Remoting from an NVIDIA A5500 workstation, we visualised 288,027 mouse hippocampus CA1 neurons at over 90 FPS with a position buffer of just 4.39 MB, representing a 93.7% reduction compared to naive storage. Third, we extended the mixed reality framework to visualise dynamic neural activity through three complementary real-time visualisation modes on HoloLens 2: animated spike propagation revealing temporal dynamics, static activity gradient mapping cumulative spike distributions, and hybrid visualisation combining both perspectives simultaneously. The system implements GPU compute shader-based activation calculations processing 288,027 neurons in under 3 milliseconds per frame whilst maintaining 60–75 FPS during temporal playback of 1,527,485 spike events spanning 981.9 ms. Fourth, we deployed the complete framework on Varjo XR-4 hardware with Ultraleap hand tracking for immersive virtual reality visualisation of human hippocampus CA1 neural activity at brain scale (5.28 million neurons, 23 million spike events). Integration with Ultraleap hand tracking at 120 Hz provides naturalistic six-degree-of-freedom manipulation through confidence-based gesture detection (threshold 0.85, hold time 0.3 seconds), supporting single-hand translation and dual-hand rotation and scaling with smooth interpolation. Fifth, we ported the visualisation pipeline from Unity to the web browser through WebGPU, translating all six validated compute shader modes from HLSL to WGSL without algorithmic modification. The user needs nothing beyond a WebGPU-capable browser. Rendering performance across all five platforms came down to a few specific decisions in the pipeline. The most consequential was adopting a zero-copy GPU memory architecture. Neural position and spike data get uploaded to GPU buffers exactly once, at load time, and no CPU-GPU transfers happen after that point during rendering. Manipulating point clouds of hundreds of thousands of elements in mixed or virtual reality does not resemble desktop mouse-based navigation in any useful way. A proxy container interaction paradigm was developed to address this, where wireframe bounding volumes act as spatial affordances that users grasp and move rather than trying to select or drag individual neurons. Taken together, these contributions show that mixed reality, virtual reality, and browser-based deployment are not just alternative display options for neuroscience data but workable research platforms in their own right.