Transformer models have gained significant attention due to their power in machine learning tasks. Their extensive deployment has raised concerns about the potential leakage of sensitive information during inference. However, when being applied to Transformers, existing approaches based on secure two-party computation (2PC) bring about efficiency limitations in two folds: (1) resource-intensive matrix multiplications in linear layers, and (2) complex non-linear activation functions like $\mathsf{GELU}$ and $\mathsf{Softmax}$. This work presents a new two-party inference framework $\mathsf{Nimbus}$ for Transformer models. Specifically, we propose a new 2PC paradigm to securely compute matrix multiplications based on an outer-product insight, which achieves $2.9\times \sim 12.5\times$ performance improvements compared to the state-of-the-art (SOTA) protocol. Furthermore, through a new observation of utilizing the input distribution, we propose an approach of low-degree polynomial approximation for $\mathsf{GELU}$ and $\mathsf{Softmax}$, which improves the performance of the SOTA polynomial approximation by $2.9\times \sim 4.0\times$, where the average accuracy loss of our approach is 0.08\% compared to the non-2PC inference without privacy. Compared with the SOTA two-party inference, $\mathsf{Nimbus}$ improves the end-to-end performance of $BERT_{base}$ inference by $2.7\times \sim 4.7\times$ across different network settings.

Weights and activation engaged inmatrix multiplication are also quantized. B.2 Problematicquantizationnodes In this subsection, we give some simple and direct studies to elaborate on the most problematic tensors (outputs ofLayerNorm structures andGELU). On the one hand, we compare the cosine similarity between the FP value and the quantized one foreachoutput. For example, "Query" means disabling the quantization of output at Query modulesacross12layers. Infact,wefindthatthisnotonly emergesinfine-tuned models but is also obvious in the pre-trained ones.

Specifically,we first propose a customized homomorphic encryption-based protocol for matrix multiplication that crucially relies on a novel compact packing technique.

Further,inthemiddlestage,i.e. fromthe7thto9thlayers,head tokens draw more attention from other tokens.

Two other important techniques are mixed precision training [36] and in-place activated BatchNorm [53]. Mixed precision training involves training using both 32-bit and 16-bit IEEE floating point numbers depending onthenumerical sensitivityofdifferent layers [36].

Previous Quantization-Aware Training (QAT) methods for vision transformers rely on expensive retraining to recover accuracy loss in non-linear layer quantization, limiting their use in resource-constrained environments. In contrast, existing Post-Training Quantization (PTQ) methods either partially quantize non-linear functions or adjust activation distributions to maintain accuracy but fail to achieve fully integer-only inference. In this paper, we introduce IPTQ-ViT, a novel PTQ framework for fully integer-only vision transformers without retraining. We present approximation functions: a polynomial-based GELU optimized for vision data and a bit-shifting-based Softmax designed to improve approximation accuracy in PTQ. In addition, we propose a unified metric integrating quantization sensitivity, perturbation, and computational cost to select the optimal approximation function per activation layer. IPTQ-ViT outperforms previous PTQ methods, achieving up to 6.44\%p (avg. 1.78\%p) top-1 accuracy improvement for image classification, 1.0 mAP for object detection. IPTQ-ViT outperforms partial floating-point PTQ methods under W8A8 and W4A8, and achieves accuracy and latency comparable to integer-only QAT methods. We plan to release our code https://github.com/gihwan-kim/IPTQ-ViT.git.

Transformer-based models have revolutionized computer vision (CV) and natural language processing (NLP) by achieving state-of-the-art performance across a range of benchmarks. However, nonlinear operations in models significantly contribute to inference latency, presenting unique challenges for efficient hardware acceleration. To this end, we propose QUARK, a quantization-enabled FPGA acceleration framework that leverages common patterns in nonlinear operations to enable efficient circuit sharing, thereby reducing hardware resource requirements. QUARK targets all nonlinear operations within Transformer-based models, achieving high-performance approximation through a novel circuit-sharing design tailored to accelerate these operations. Our evaluation demonstrates that QUARK significantly reduces the computational overhead of nonlinear operators in mainstream Transformer architectures, achieving up to a 1.96 times end-to-end speedup over GPU implementations. Moreover, QUARK lowers the hardware overhead of nonlinear modules by more than 50% compared to prior approaches, all while maintaining high model accuracy -- and even substantially boosting accuracy under ultra-low-bit quantization.

Current transformer accelerators primarily focus on optimizing self-attention due to its quadratic complexity. However, this focus is less relevant for vision transformers with short token lengths, where the Feed-Forward Network (FFN) tends to be the dominant computational bottleneck. This paper presents a low power Vision Transformer accelerator, optimized through algorithm-hardware co-design. The model complexity is reduced using hardware-friendly dynamic token pruning without introducing complex mechanisms. Sparsity is further improved by replacing GELU with ReLU activations and employing dynamic FFN2 pruning, achieving a 61.5\% reduction in operations and a 59.3\% reduction in FFN2 weights, with an accuracy loss of less than 2\%. The hardware adopts a row-wise dataflow with output-oriented data access to eliminate data transposition, and supports dynamic operations with minimal area overhead. Implemented in TSMC's 28nm CMOS technology, our design occupies 496.4K gates and includes a 232KB SRAM buffer, achieving a peak throughput of 1024 GOPS at 1GHz, with an energy efficiency of 2.31 TOPS/W and an area efficiency of 858.61 GOPS/mm2.