TreeGRNG: Binary Tree Gaussian Random Number Generator for Efficient Probabilistic AI Hardware

A new hardware-oriented Gaussian random number generator, TreeGRNG, replaces arithmetic units with constant comparators to cut energy per sample by a factor of 3.7 and boost throughput per unit area by a factor of 5.8, according to a paper submitted in June 2026 [1]. The design targets Bayesian Neural Networks (BNNs), which can monitor uncertainty in decision-making but require a Gaussian Random Number Generator (GRNG) inside each neuron. State-of-the-art GRNG algorithms rely on multiple arithmetic operations and large look-up tables, making them difficult to implement in ultra-low-power hardware [2]. TreeGRNG sidesteps that bottleneck by using a binary tree structure built from constant comparators, which the authors describe as “ultra-low-cost” [2]. Alongside the hardware savings, the paper reports that the optimized TreeGRNG surpasses existing generators in distribution accuracy [2]. The authors also note a flexibility advantage: designers can adjust the shape of the sampled probability distribution, a capability that extends beyond traditional GRNGs and points toward future probabilistic AI designs [2]. The TreeGRNG design has been released as open-source [2]. The paper was submitted to arXiv on 15 June 2026 under the Computer Science > Hardware Architecture category [1]. arXiv, which hosts the majority of machine-learning preprints, allows authors to link associated code, models, and demos through integrations with platforms such as Hugging Face [4][5]. Hugging Face’s paper pages can surface related artifacts—models, datasets, and interactive Spaces—alongside a preprint, and the platform has collaborated with arXiv to embed demos directly on abstract pages [5]. The TreeGRNG work arrives as the broader AI hardware landscape faces pressure to reduce energy consumption. Recent large-language-model releases have drawn attention to training and inference costs. DeepSeek, for instance, reported training its V3 model for roughly US$6 million, a fraction of the estimated US$100 million cost for OpenAI’s GPT-4, using approximately one-tenth the computing power of Meta’s comparable Llama 3.1 model [7]. While TreeGRNG addresses a different layer of the stack—probabilistic inference at the extreme edge—the emphasis on energy efficiency and open-source availability echoes wider industry trends [2][7].