Reasoning with Large Language Models (LLMs) has become a pivotal research topic because their logical abilities significantly surpass those of standard LLMs. LLM reasoning typically forms multiple chains of thought, action-by-action, and selects the best one as the final answer. However, the inference overhead of LLM reasoning is more than an order of magnitude higher than that of LLM. Despite the emerging shift towards memory-optimized algorithms and near-memory hardware, we still face the following challenges: (1) Existing memory-centric algorithms (e.g., KV cache technique) have low computational utilization (< 4% on NVIDIA A100 GPU) due to intensive memory access for inter-action data. (2) Emerging hardware architectures (e.g., near-memory processing) fail to fully utilize the inherent parallelism due to dependencies among models, leading to low utilization of memory bandwidth.To tackle these challenges, we propose Endor, a hardware-algorithm co-design to accelerate the inference of LLM reasoning efficiently. We identify that the auto-regressive decoding of LLM reasoning changes from the token level to the action level in terms of the computing paradigm. At the algorithm level, we propose a "nearly-decode-only" method which encompasses an efficient inter-action cache reuse method and a prediction-based pipeline optimization to reduce computation overhead. At the hardware level, we propose Endor-NMP, a near-memory accelerator featuring a score-aware cache management architecture and a heterogeneous mapping dataflow. Endor fully exploits both interaction and intra-action parallelism to improve memory bandwidth utilization. Experimental results demonstrate that neither existing algorithms nor hardware can achieve the expected acceleration. Endor achieves an end-to-end average speedup of 2.97× and 2.52× compared to the NVIDIA A100 GPU and advanced LLM accelerators on multiple models and datasets.

To alleviate the vulnerability of graph neural networks (GNNs) on unseen graphs, many works propose to integrate large language models (LLMs) into GNNs, called graph foundation models (GFMs). The LLM-enhanced GNN, a typical integration method of GFMs, has achieved state-of-the-art performance in most graph-related tasks. However, intensive general matrix multiplications (GEMMs) overhead of LLMs poses a significant challenge to end-to-end inference latency. The introduction of LLMs brings 100× more workload than original GNNs, with GEMMs accounting for more than 99%, becoming the bottleneck of end-to-end inference.

In recent years, Neural Networks developed rapidly to deal with tasks in the field of Computer Vision and Natural Language Process, etc. With the development of AI Generated Content, U-Net based Diffusion Models (DM) take image synthesis to new heights. U-Net performs the noise prediction of DM, the latency of which accounts for the majority of the end-to-end latency. Although FPGA has been proven to be a high performance platform to deploy NN, a series of facts still pose challenges for efficient U-Net based DMs deployment based on FPGA. The input vector length and type of the special function vary between different layers. The absence of model periodicity increases the granularity and complexity of operator scheduling. Skip-connection and residual connection inside model cause meta-data retaining in the memory, which is not conductive to avoiding memory overflow and decreasing total off-chip memory access.

The row-wise product algorithm shows significant potential for sparse matrix-matrix multiplication (SpGEMM) on hardware accelerators. Recent studies have made notable progress in accelerating SpGEMM using this algorithm. However, several challenges remain in accelerating imbalanced SpGEMM, where the distribution of non-zero elements across different rows is imbalanced. These challenges include: (1) the fixed dataflow of the merger tree, which leads to lower PE utilization, and (2) highly imbalanced data distributions, such as single rows with numerous non-zero elements, which result in intensive computations. This imbalance significantly challenges SpGEMM acceleration, leading to time-consuming processes that dominate overall computation time. In this paper, we propose AiSpGEMM to accelerate imbalanced SpGEMM on FPGAs. First, we improved the C2SR format to adapt it for imbalanced SpGEMM acceleration based on the row-wise product algorithm. This reduces off-chip memory bank conflicts and increases data reuse of matrix B. Secondly, we design a reconfigurable merger (R-merger) with flexible interconnects to improve PE utilization. Additionally, we propose an intra-row parallel merging algorithm and its corresponding hardware architecture, the parallel merger (P-merger), to accelerate intensive operations. Experimental results demonstrate that AiSpGEMM achieves a geometric mean (geomean) speedup of 5.8× compared to the state-of-the-art FPGA-based SpGEMM accelerator. In Geomean, AiSpGEMM achieves a 3.0× speedup and a 9.8× improvement in energy efficiency compared to the NVIDIA cuSPARSE library running on an NVIDIA A6000 GPU. Moreover, AiSpGEMM-21 demonstrated a 4× increase in average throughput compared to the same GPU.

Video Generation Model (VGM), as a representative of multi-modal large models, has revolutionized the productivity of video content creation. VGMs are compute-bound due to adopting the Diffusion Transformer (i.e., DiT) structure. Sparsification is a common method for accelerating compute-intensive models. Still, sparse VGMs cannot fully exploit the effective throughput (i.e., TOPS) of GPUs. FPGAs are good candidates for accelerating sparse deep learning models. However, existing FPGA accelerators still face low throughput ( < 2TOPS) on VGMs due to the significant gap in peak computing performance (PCP) with GPUs ( > 21× ). To achieve a higher throughput than GPUs, FPGA-based acceleration of sparse VGMs still faces the following challenges: large redundancy in activations, low performance of DSPs under hybrid precision, and under-utilization using static compilation for online compression.

A few overlay processors for transformer networks emerge to achieve reconfigurable architectures and dynamic instructions. However, these processors consistently neglect exploring network sparsity, while existing sparse accelerators inefficiently utilize resources with separate computation parts. Furthermore, mainstream compilers for instruction generation are intricate and demand significant engineering efforts. In this work, we propose CSTrans-OPU, an FPGA-based overlay processor with full compilation for transformer networks via sparsity exploration. Specifically, we customize a multi-precision processing element (PE) array with DSP-packing for unified computation format with full resource utilization. Additionally, the introduced sorting and computation mode selection modules make it possible to explore the token sparsity. Moreover, equipped with a user-friendly compiler, CSTrans-OPU enables model parsing, operation fusion, model quantization, instruction generation and reordering directly from model files. Experimental results show that CSTrans-OPU achieves 6.92-20.06× speedup and 182.48× higher energy efficiency compared with CPU, and 1.47-3.85× latency reduction with 4.63-52.53× better energy efficiency compared with GPU. Furthermore, we observe up to 4.28× better latency and 4.94× higher energy efficiency compared with previously customized accelerators, and can be up to 1.93× faster and 4.39× more energy efficient than FPGA processors. To the best of our knowledge, our CSTrans-OPU is the first overlay processor for transformer networks considering sparsity.

Stable Diffusion has become one of the mainstream image synthesis algorithms. The mainstream computing platform for Stable Diffusion is GPU. However, the deployment of Stable Diffusion on GPU still faces the problems of power consumption. With dedicated hardware design and optimization, FPGA based Stable Diffusion accelerator can achieve better performance of energy efficiency. In this paper, we propose SDAcc for realizing efficient inference of Stable Diffusion on FPGA. SDAcc is 4.40× faster than CPU. Compared to GPU and CPU, SDAcc achieves 1.27× and 19.66× energy efficiency improvement, respectively.

Logic synthesis is a crucial step in electronic design automation tools. The rapid developments of reinforcement learning (RL) have enabled the automated exploration of logic synthesis. Existing RL based methods may lead to data inefficiency, and the exploration approaches for FPGA and ASIC technology mapping in recent works lack the flexibility of the learning process. This work proposes ESE, a reinforcement learning based framework to efficiently learn the logic synthesis process. The framework supports the modeling of logic optimization and technology mapping for FPGA and ASIC. The optimization for the execution time of the synthesis script is also considered. For the modeling of FPGA mapping, the logic optimization and technology mapping are combined to be learned in a flexible way. For the modeling of ASIC mapping, the standard cell based optimization and LUT optimization operations are incorporated into the ASIC synthesis flow. To improve the utilization of samples, the Proximal Policy Optimization model is adopted. Furthermore, the framework is enhanced by supporting MIG based synthesis exploration. Experiments show that for FPGA technology mapping on the VTR benchmark, the average LUT-Level-Product and script runtime are improved by more than 18.3% and 12.4% respectively than previous works. For ASIC mapping on the EPFL benchmark, the average Area-Delay-Product is improved by 14.5%.

Field-programmable gate arrays (FPGAs) are an ideal candidate for accelerating graph neural networks (GNNs). However, the FPGA redeployment process is time-consuming when updating or switching between diverse GNN models across different applications. Existing GNN processors eliminate the need for FPGA redeployment when switching between different GNN models. However, adapting matrix multiplication types by switching processing units decreases hardware utilization. In addition, the bandwidth of DDR limits further improvements in hardware performance. This article proposes a highly flexible FPGA-based overlay processor for GNN accelerations. Graph-OPU provides excellent flexibility and programmability for users, as the executable code of GNN models is automatically compiled and reloaded without requiring FPGA redeployment. First, we customize the compiler and instruction sets for the inference process of different GNN models. Second, we customize the datapath and optimize the data format in the microarchitecture to fully leverage the advantages of high bandwidth memory (HBM). Third, we design a unified matrix multiplication to handle both sparse-dense matrix multiplication (SpMM) and general matrix multiplication (GEMM), enhancing Graph-OPU performance. During Graph-OPU execution, the computational units are shared between SpMM and GEMM instead of being switched, which improves the hardware utilization. Finally, we implement a hardware prototype on the Xilinx Alveo U50 and test the mainstream GNN models using various datasets. Experimental results show that Graph-OPU achieves up to 1,654 \(\times\) and 63 \(\times\) speedup, as well as up to 5,305 \(\times\) and 422 \(\times\) energy efficiency boosts, compared to implementations on CPU and GPU, respectively. Graph-OPU outperforms state-of-the-art (SOTA) end-to-end overlay accelerators for GNN, reducing latency by an average of 1.36 \(\times\) and improving energy efficiency by 1.41 \(\times\) on average. Moreover, Graph-OPU exhibits an average 1.45 \(\times\) speed improvement in end-to-end latency over the SOTA GNN processor. Graph-OPU represents an in-depth study of an FPGA-based overlay processor for GNNs, offering high flexibility, speedup, and energy efficiency.

Existing implementations of transformer networks by field-programmable gate array (FPGA) focus only on attention computation, or suffer from fixed model structure without flexibility. In this article, we propose an FPGA-based overlay processor, named Transformer-OPU for general accelerations of transformer networks. Experimental result shows that our Transformer-OPU achieves 5.19-15.06× and 1.14-2.89× speedup compared with CPU and GPU, respectively. We also observe 1.10-2.47× better latency compared with previously customized FPGA accelerators, and is 1.45× faster than NPE.

Existing accelerators for transformer networks with field-programmable gate array (FPGA) either focus only on attention computation or suffer from fixed data streams without flexibility. Moreover, compression and approximation methods of transformer networks have the potential for further optimization. In this article, we propose a low-latency FPGA-based overlay processor, named LTrans-OPU for general accelerations of transformer networks. Specifically, we design a domain-specific overlay architecture, including a computation unit for matrix multiplication of arbitrary dimensions. An instruction set customized for our overlay architecture is also introduced, dynamically controlling data flows by generated instructions. In addition, we introduce a hybrid pruning method common to various transformer networks, along with an efficient non-linear function approximation method. Experimental results show that our design is rather competitive and has low latency. LTrans-OPU achieves 11.10-32.20× speedup compared with CPU and 2.44-6.18 × latency reduction compared with GPU. We also observe 2.36-12.43 × lower latency compared with customized FPGA/ASIC accelerators, and can be 3.10× faster than NPE.

Transformer-based models suffer from large num-ber of parameters and high inference latency, whose deployment are not green due to the potential environmental damage caused by high inference energy consumption. In addition, it is difficult to deploy such models on devices, especially on resource constrained devices such as FPGA. Various model pruning methods are proposed to shrink the model size and resource consumption, so as to fit the models on hardware. However, such methods often introduce floating point of operations (FLOPs) as an agent of hardware performance, which is not accurate. Furthermore, structural pruning methods are always in a single head-wise or layer-wise pattern, which fails to compress the models to the extreme. To resolve the above issues, we propose a green BERT deployment method on FPGA via hardware-aware and hybrid pruning, named g-BERT. Specifically, two hardware-aware metrics are introduced by High Level Synthesis (HLS) to evaluate the latency and power consumption of inference on FPGA, which can be optimized directly while pruning. Moreover, we simultaneously consider pruning of heads and full encoder layers. To efficiently find the optimal structure, g-BERT applies differentiable neural architecture search (NAS) with a special 0–1 loss function. Compared with the BERT-base, g-BERT achieves$2.1\times$speedup,$1.9\times$power consumption reduction and$1.8\times$model size reduction with comparable accuracy, on par with the state-of-the-art methods.

There are already some works on accelerating transformer networks with field-programmable gate array (FPGA). However, many accelerators focus only on attention computation or suffer from fixed data streams without flexibility. Moreover, their hardware performance is limited without schedule optimization and full use of hardware resources. In this article, we propose a flexible and efficient FPGA-based overlay processor, named FET-OPU. Specifically, we design an overlay architecture for general accelerations of transformer networks. We propose a unique matrix multiplication unit (MMU), which consists of a processing element (PE) array based on modified DSP-packing technology and a FIFO array for data caching and rearrangement. An efficient non-linear function unit (NFU) is also introduced, which can calculate arbitrary single input non-linear functions. We also customize an instruction set for our overlay architecture, dynamically controlling data flows by instructions generated on the software side. In addition, we introduce a two-level compiler and optimize the parallelism and memory allocation schedule. Experimental results show that our FET-OPU achieves 7.33-21.27× speedup and 231× less energy consumption compared with CPU, and 1.56-4.08× latency reduction with 5.85-66.36× less energy consumption compared with GPU. Furthermore, we observe 1.56-8.21× better latency and 5.28-6.24× less energy consumption compared with previously customized FPGA/ASIC accelerators and can be 2.05× faster than NPE with 5.55× less energy consumption.

Field-programmable gate arrays (FGPAs) are widely used because of the superiority in flexibility and lower non-recurring engineering cost. How to optimize the routing architecture is a key problem for FPGA architects because it has a large impact on FPGA area, delay, and routability. In academia, the routing architecture is mainly based on the connection blocks (CBs) and switch blocks (SBs), whereas most research has focused on SB architectures, such as Wilton, Universal, and Disjoint SB patterns. In this article, we propose a novel unidirectional routing architecture—general interconnection block (GIB)—to improve FPGA performance. With the GIB architecture, logic block (LB) pins can directly connect with the adjacent GIBs without programmable switches. Inside a GIB, LB pins can connect to the routing channel tracks on the four sides of a GIB. In particular, the logic pins from different neighboring LBs that connect to the same GIB can connect with each other with only one programmable switch. In addition, we enhance VTR to support the GIB with bent wires and develop a searching framework based on the simulated annealing algorithm to search for a near-optimal distribution of wire types. We evaluate the GIB architecture on VTR 8 with the provided benchmark circuits. The experimental results show that the GIB architecture with length-4 wires can achieve 9.5% improvement on the critical path delay and 11.1% improvement on the area-delay product compared to the VTR CB-SB architecture with length-4 wires. After exploring mixed wire types, the optimized GIB architecture can further improve the delay by 16.4% and area-delay product by 17.1% compared to the CB-SB architecture with length-4 wires.

Modern FPGAs utilize complex routing architectures to optimize the area, critical path delay, and power consumption. General Routing Block (GRB) models the routing resources of modern FPGAs, enabling the design of better routing architectures than previous academic FPGAs based on the CB-SB model. However, the design space of the GRB model is too large to be explored manually. In this paper, we propose GRAEBO, a design space exploration (DSE) algorithm for FPGA routing architectures based on Bayesian optimization, which can optimize and accelerate the DSE by balancing exploration and exploitation. Moreover, we design pruning rules to further improve the DSE efficiency, which can serve as a multi-fidelity acceleration method. GRAEBO obtains better area, delay, and area-delay product than a 142-channel baseline CB-SB architecture, with improvements of 8%, 19%, and 26%, respectively. Compared to the GRB architecture found by the simulated annealing algorithm, GRAEBO achieves 9% smaller area, 5% shorter delay, and 13% better area-delay product on the VTR benchmarks.

Logic synthesis is a crucial step in electronic design automation tools for integrated circuit design. In recent years, the development of reinforcement learning (RL) has enabled the designers to automatically explore the logic synthesis process. Existing RL based methods typically use conventional on-policy models, which leads to data inefficiency. Moreover, the exploration approach for FPGA technology mapping in recent works lacks the flexibility of the learning process. In this work, we propose ESE, a reinforcement learning based framework to efficiently learn the logic synthesis process. The framework supports the modeling for both the logic optimization and the FPGA technology mapping. The reward functions and terminal conditions in the RL environment are designed to efficiently guide the optimization of the metrics and execution time. For the modeling of FPGA mapping, the logic optimization and technology mapping are combined to be learned in a flexible way. Moreover, the Proximal Policy Optimization model is adopted to improve the utilization of samples. The proposed framework is evaluated on several common benchmarks. For the logic optimization on the EPFL benchmark, compared with previous works, the proposed method obtains an 11.3% improvement in the average quality (node-level-product) and reduces the execution time by 13.7%. For the FPGA technology mapping on the VTR benchmark, our method improves the average quality (LUT-level-product) by 14.8%, and reduces the execution time by 14.4% compared with the recent work.

In FPGAs, the programmable interconnect is implemented by multiplexers (MUXes), which have a large impact on the area and delay. In academia, large MUXes are extensively used in intra and inter clusters, resulting in significant FPGA area overhead and load for routing wires. In this paper, we model the interconnect from routing wires and CLB feedbacks to LUT inputs as an input block (IB), and implement the IB and the switch block (SB) with the 2-level MUX topology. Applying the 2-level MUX topology in FPGA routing architecture enables us to explore a larger design space for the area and delay, because the 2-level MUX topology can tradeoff between MUX sizes, connectivity degree, and the input bandwidth. We carefully design a baseline 2-level MUX routing architecture and evaluate it by running place and route experiments with VTR benchmarks. To optimize the baseline 2-level MUX routing architecture, we explore one design parameter at a time by keeping others fixed and perform subsequent explorations based on previous optimal design parameters. The results show that the optimized 2-level MUX routing architecture can achieve 1% shorter critical path delay (CPD) at the cost of 3% area overhead compared to the CB-SB FPGA architecture with 1-level MUX topology.

In error-tolerant applications such as low-precision DNNs and digital filters, approximate arithmetic circuits can significantly reduce hardware resource utilization. In this work we propose an embedded block for field-programmable gate arrays, called APIR-DSP, which incorporates an approximate 9×9 hard multiplier based on the PIR-DSP architecture to improve speed and reduce area. In addition, a DSP unit evaluation platform based on Yosys and VPR which packs multiply accumulate operations into DSP blocks is developed. Using this tool we synthesis designs from Verilog implementations of matrix multiplication in DeepBench and the DoReFaNet low-precision neural network and show that APIR-DSP significantly reduces DSP resources and improves hardware utilization and performance compared with the Xilinx DSP48E2 embedded block. Compared with exact multiplication, it is shown that accuracy loss is optimized with the SNR of an FIR filter being reduced by 1.03 dB. For DNNs, accuracy loss for AlexNet is 0.31% on CIFAR10 dataset and no accuracy loss for LeNet on MNIST dataset is observed. Synthesis results show that the APIR-DSP enjoys an area reduction of 21.60%, critical path reduction of 4.85% and power consumption is reduced by 2.80%, compared with PIR-DSP.

Routing architecture has a significant impact on the area, critical path delay and power consumption of modern FPGAs. The most common routing architecture of island-style FPGAs in academia is the CB-SB model, which is not effective to model complex routing architectures in modern FPGAs. To improve the routability and performance of the existing routing model, we propose a new routing model called General Routing Block (GRB) to model complex commercial FPGAs. In the proposed model, all routing resources can be divided into three modules: general switch block (GSB), input connection block (ICB) and output connection block (OCB). The GSB and ICB are extended from the SB and CB with more flexible and richer connections. The OCB is a new module that provides novel connections for the LB output pins. We support bent wire architecture to reduce the delay, and two-level MUXes with output sharing to achieve a better trade-off between the area and flexibility. Moreover, to explore the trade-offs of different design spaces and find better architectures, an architecture exploration platform based on the simulated annealing algorithm is proposed to efficiently explore the enormous design space specified by a set of parameters. The results of global design space exploration show that the architecture with the proposed GRB model reduces the critical path delay by 15.5% and area-delay product by 14.8% compared to the length-4 CB-SB architecture based on the VTR benchmarks. After further local subspace explorations, the best architecture can achieve an 18.7% improvement on the critical path delay and a 23.8% improvement on the area-delay product, which represents a significant improvement over other routing architectures.

Field Programmable Gate Arrays (FPGAs) are widely used for their flexibility and short time to market. FPGA routing architecture design is the key problem due to the fact that it plays a dominant role in the area, delay and power. Most of modern FPGAs are island-style which provide abundant vertical and horizontal tracks to guarantee the circuit designs can be routed successfully. Most connections in placed netlists are diagonal which may lead to passing through extra turning switches, resulting in increased delay cost and high routing density. In this paper, we propose a hexagon-based honeycomb FPGA routing architecture to improve the routability and performance. In honeycomb architecture, there are three kinds of routing channels which can provide more freedom to decrease the turning switches on the routing paths. In addition, the router lookahead algorithm is enhanced to support the honeycomb architecture which is then evaluated by the enhanced VTR with provided benchmarks. The experimental results show that the honeycomb architecture can improve the minimum routing channel width by 7.7% compared with traditional rectangular architecture with length-1 wires. In addition, the honeycomb architecture can achieve 9.9% improvement on the routed wirelength, 11.5% on the critical path delay and 12.4% on the area-delay product.

We propose two tiers of modifications to FPGA logic cell architecture to deliver a variety of performance and utilization benefits with only minor area overheads. In the first tier, we augment existing commercial logic cell datapaths with a 6-input XOR gate in order to improve the expressiveness of each element, while maintaining backward compatibility. This new architecture is vendor-agnostic, and we refer to it as LUXOR. We also consider a secondary tier of vendor-specific modifications to both Xilinx and Intel FPGAs, which we refer to as X-LUXOR+ and I-LUXOR+ respectively. We demonstrate that compressor tree synthesis using generalized parallel counters (GPCs) is further improved with the proposed modifications. Using both the Intel adaptive logic module and the Xilinx slice at the 65nm technology node for a comparative study, it is shown that the silicon area overhead is less than 0.5% for LUXOR and 5-6% for LUXOR+, while the delay increments are 1-6% and 3-9% respectively. We demonstrate that LUXOR can deliver an average reduction of 13-19% in logic utilization on micro-benchmarks from a variety of domains. BNN benchmarks benefit the most with an average reduction of 37-47% in logic utilization, which is due to the highly-efficient mapping of the XnorPopcount operation on our proposed LUXOR+ logic cells.

Quantisation is a key optimisation strategy to improve the performance of floating-point deep neural network (DNN) accelerators. Digital signal processing (DSP) blocks on field-programmable gate arrays are not efficiently utilised when the accelerator precision is much lower than the DSP precision. Through three modifications to Xilinx DSP48E2 DSP blocks, we address this issue for important computations in embedded DNN accelerators, namely the standard, depth-wise, and pointwise convolutional layers. First, we propose a flexible precision, run-time decomposable multiplier architecture for CNN implementations. Second, we propose a significant upgrade to DSPDSP interconnect, providing a semi-2D low precision chaining capability which supports our low-precision multiplier. Finally, we improve data reuse via a register file which can also be configured as FIFO. Compared with the 27 × 18-bit mode in the Xilinx DSP48E2, our Precision, Interconnect, and Reuseoptimised DSP (PIR-DSP) offers a 6× improvement in multiplyaccumulate operations per DSP in the 9 × 9-bit case, 12× for 4 × 4 bits, and 24× for 2 × 2 bits. We estimate that PIR-DSP decreases the run time energy to 31/19/13% of the original value in a 9/4/2-bit MobileNet-v2 DNN implementation.

Routing architecture design is crucial in the early exploration of FPGA fabrics. In the modern unidirectional routing architectures, the signal passes through a programmable switch while turning into the channel perpendicular to the current one at the intersections, which may lead to excessive turning switches in the routing paths and cause the high latency. Most of the researches focus on the routing topology where the wire segments span in either vertical or horizontal direction. In this paper, we propose the bent routing pattern, where the bent wire segments can span in both vertical and horizontal channels without passing through any turning switch. The bent routing topology is designed to keep the regularity and symmetry, and then evaluated by the enhanced VTR. To optimize the architecture with the bent routing pattern, we develop a stochastic searching method based on the simulated annealing algorithm. The results show that the architecture with the mixture of bent wires and straight wires can achieve 9% shorter critical path delay and 11% area-delay product savings on average compared to the architecture with only straight wires.

Interlayer liquid cooling network has been considered as one of the effective cooling mechanisms for heat dissipation in three dimensional integrated circuits (3D ICs). In this paper, an optimization approach is proposed for 3D IC interlayer cooling network design with consideration of cooling energy minimization and thermal constraints. First, a channel patterning technique is proposed which adopts straight-channel patterns and corner-barrier patterns to efficiently reduce the temperature of hotspots in the chip. Second, an iterative channel pruning process is proposed for cooling network optimization that reduces the temperature gradient and cooling energy. Furthermore, algebraic multigrid-preconditioned generalized conjugate residual solver is applied for thermal simulation in this optimization process to achieve better computational efficiency. Experimental results have shown that the proposed optimization approach can produce optimal interlayer cooling networks with 11.3%-61.3% cooling energy savings and 49.3%-59.4% pumping pressure savings in comparison with the first place winner of 2015 computer-aided design contest at International Conference on Computer-Aided Design.

In this paper we propose an optimized bitstream compression algorithm based on LZ and a novel architecture of decompressor, the proposed algorithm improves the Compression Ratio by fully utilizing the regularity of configuration bits of CLB (Configurable Logic Box) in FPGA and using the variable length Golomb coding method. The experimental results show that the Optimized method can improve the Compression Ratio of LZSS by 32.3% for bitstream with high regularity and 10.3% for bitstream with low regularity, and our approach shows a higher flexibility than the BMC+RLE arithmetic when compressing the bitstream with high regularity for various FPGA. Moreover, we design a two-buffer-window decompressor to download the compressed bitstreams. In order to increase the throughput of the proposed decompressor, we design a multi-stage data selector in it. The post-simulation of the decompressor shows that its throughput is up to 9280 Mbps under 65nm CMOS process. And that is 4352Mbps when verified on a Virtex-5 FPGA.