Kirin: Improving ANN efficiency with SNN Hybridization

Artificial neural networks (ANNs), particularly large language models (LLMs), demonstrate powerful inference capabilities but consume substantial energy. Conversely, spiking neural networks (SNNs) exhibit exceptional energy efficiency due to their binary and event-driven characteristics, thus motivating the study of ANN-to-SNN conversion. In this process, quantization plays a pivotal role, mapping LLMs’ floating-point parameters to discrete SNN parameters via the temporal dimension of the time window. However, several challenges remain in the conversion process: (i) converting high bit-width quantization values into binary spikes requires longer time windows, increasing system latency; and (ii) the inherent trade-off between the information loss of single-spike schemes and the energy costs of multi-spike ones in SNN. To address these challenges, we propose Kirin, a integer and spike hybrid based SNN to achieve accuracy lossless ANN-to-SNN conversion with time and energy efficiency. Specifically, we first propose a Spike Matrix Hybridization strategy that encoding low bit-width parameters that leading to small time window size into binary spikes while preserving the rest in integer format, thereby reducing the overall latency of SNN execution. Second, we introduce a silence threshold mechanism to regulate the timing of single-spike firing, ensuring the output is mathematically equivalent to the LLM’s output and preserves accuracy. Experimental results demonstrate that Kirin, under a W4A4&8 quantization setting, achieves near-FP16 accuracy while reducing energy consumption by up to 84.66% and shortening time steps by 93.75%.

💡 Research Summary

The paper addresses the growing energy consumption problem of large language models (LLMs) by proposing Kirin, a novel framework that combines the energy‑efficiency of spiking neural networks (SNNs) with the high‑accuracy of artificial neural networks (ANNs). Traditional ANN‑to‑SNN conversion pipelines first quantize floating‑point parameters and then encode them as binary spike trains over a time window whose length grows exponentially with the quantization bit‑width (T = 2^b). This creates two major bottlenecks: (i) outlier weights or activations that require high‑bit quantization force the entire network to use a long time window, dramatically increasing latency; (ii) temporal coding schemes such as Time‑to‑First‑Spike (TTFS) allow only a single spike per neuron, causing severe information loss across layers, while rate‑based coding avoids loss but incurs a high spike rate and thus high energy consumption.

Kirin solves both issues with two complementary mechanisms. First, the Spike Matrix Hybridization strategy detects outliers using a robust MAD‑based algorithm and applies mixed‑precision quantization: normal values are quantized to low bits (e.g., 4‑bit) and encoded as TTFS spikes, while high‑bit outliers (e.g., 8‑bit) are kept as integer values and processed with conventional multiply‑accumulate (MAC) operations. By selecting the matrix with fewer outliers as the “spike matrix,” the method minimizes the number of integer MACs, preserving the overall energy advantage of SNNs. Empirically, outliers constitute only a few percent of parameters (≈2.68 % in Llama‑2‑7B), so the added MAC cost is negligible.

Second, Kirin introduces a Silence Threshold mechanism within the integrate‑and‑fire (IF) neuron. In standard TTFS, once the membrane potential crosses the firing threshold the neuron emits a spike and then remains silent for the rest of the window, discarding any further accumulated potential. Kirin modifies the IF dynamics so that when the potential reaches a threshold derived from the quantization scale, the neuron may stay silent (i.e., “silence”) while still accumulating potential. This ensures that the final output mathematically matches the de‑quantized floating‑point value, effectively preserving the information that would otherwise be lost in a single‑spike scheme, while retaining the low‑energy characteristics of TTFS.

The overall workflow consists of: (1) outlier detection and mixed‑precision quantization; (2) selection of spike vs. integer matrices; (3) TTFS encoding of low‑bit values; (4) IF processing with the silence threshold; and (5) reconstruction of the final output. Experiments under a W4A4&8 quantization setting demonstrate that Kirin achieves near‑FP16 accuracy, reduces energy consumption by up to 84.66 %, and shortens the required time steps by 93.75 % compared with existing SNN conversion methods. These results show that Kirin effectively breaks the traditional trade‑off triangle among accuracy, latency, and energy in ANN‑to‑SNN conversion, making it a promising candidate for energy‑constrained inference on edge devices and next‑generation AI accelerators. The authors also discuss future directions, including dynamic adjustment of the hybrid ratio for models with higher outlier prevalence, hardware‑level optimizations, and extending the approach to non‑language tasks.