Massive Spikes in LLMs are Bias Vectors: Mechanistic Uncovering and Spike-Free Quantization

Massive activation spikes in Large Language Models (LLMs) severely degrade quantization by stretching dynamic ranges. While prior hypotheses characterize these as high-level scalar biases, we argue that they are merely the scalar intermediates of rigid, structural vector biases in the spike-carrying tokens. We show that these tokens converge to constant vectors after normalization that drive the attention sink and value-state drain mechanisms. We geometrically substantiate this by analyzing the coordination of projection weights: $W_K$ contrastively amplifies the vector, $W_Q$ aligns semantic tokens toward it, and $W_V$ projects it into the spectral null-space. Furthermore, we reveal that the model actively preserves these structural biases against Rotary Positional Embedding (RoPE) perturbations by localizing them in "zones of rotational stability" utilizing low-frequency bands and coherent channel pairs. Leveraging this, we propose INSERTQUANT, a post-training quantization (PTQ) framework that clamps spikes and restores their function via pre-computed template vectors. This renders activations strictly spike-free, enabling robust low-bit quantization with high fidelity. INSERTQUANT achieves parity with state-of-the-art per-tensor quantization methods on LLMs and uniquely generalizes beyond text to other modalities such as ViTs.

PACiM: A Sparsity-Centric Hybrid Compute-in-Memory Architecture via Probabilistic Approximation

Approximate computing emerges as a promising approach to enhance the efficiency of compute-in-memory (CiM) systems in deep neural network processing. However, traditional approximate techniques often significantly trade off accuracy for power efficiency, and fail to reduce data transfer between main memory and CiM banks, which dominates power consumption. This paper introduces a novel probabilistic approximate computation (PAC) method that leverages statistical techniques to approximate multiply-and-accumulation (MAC) operations, reducing approximation error by 4X compared to existing approaches. PAC enables efficient sparsity-based computation in CiM systems by simplifying complex MAC vector computations into scalar calculations. Moreover, PAC enables sparsity encoding and eliminates the LSB activations transmission, significantly reducing data reads and writes. This sets PAC apart from traditional approximate computing techniques, minimizing not only computation power but also memory accesses by 50%, thereby boosting system-level efficiency. We developed PACiM, a sparsity-centric architecture that fully exploits sparsity to reduce bit-serial cycles by 81% and achieves a peak 8b/8b efficiency of 14.63 TOPS/W in 65 nm CMOS while maintaining high accuracy of 93.85/72.36/66.02% on CIFAR-10/CIFAR-100/ImageNet benchmarks using a ResNet-18 model, demonstrating the effectiveness of our PAC methodology.