Achieving high performance, energy efficiency, and cost-effectiveness while maintaining architectural flexibility is a critical challenge in the development and deployment of edge AI devices. Monolithic SoC designs struggle with this complex balance mainly due to low manufacturing yields (below 16%) at advanced 360 mm^2 process nodes. This paper presents a novel chiplet-based RISC-V SoC architecture that addresses these limitations through modular AI acceleration and intelligent system level optimization. Our proposed design integrates 4 different key innovations in a 30mm x 30mm silicon interposer: adaptive cross-chiplet Dynamic Voltage and Frequency Scaling (DVFS); AI-aware Universal Chiplet Interconnect Express (UCIe) protocol extensions featuring streaming flow control units and compression-aware transfers; distributed cryptographic security across heterogeneous chiplets; and intelligent sensor-driven load migration. The proposed architecture integrates a 7nm RISC-V CPU chiplet with dual 5nm AI accelerators (15 TOPS INT8 each), 16GB HBM3 memory stacks, and dedicated power management controllers. Experimental results across industry standard benchmarks like MobileNetV2, ResNet-50 and real-time video processing demonstrate significant performance improvements. The AI-optimized configuration achieves ~14.7% latency reduction, 17.3% throughput improvement, and 16.2% power reduction compared to previous basic chiplet implementations. These improvements collectively translate to a 40.1% efficiency gain corresponding to ~3.5 mJ per MobileNetV2 inference (860 mW/244 images/s), while maintaining sub-5ms real-time capability across all experimented workloads. These performance upgrades demonstrate that modular chiplet designs can achieve near-monolithic computational density while enabling cost efficiency, scalability and upgradeability, crucial for next-generation edge AI device applications.

Large Language Models (LLMs) are becoming increasingly data-intensive due to growing model sizes, and they are becoming memory-bound as the context length and, consequently, the key-value (KV) cache size increase. Inference, particularly the decoding phase, is dominated by memory-bound GEMV or flat GEMM operations with low operational intensity (OI), making it well-suited for processing-in-memory (PIM) approaches. However, existing in/near-memory solutions face critical limitations such as reduced memory capacity due to the high area cost of integrating processing elements (PEs) within DRAM chips, and limited PE capability due to the constraints of DRAM fabrication technology. This work presents a chiplet-based memory module that addresses these limitations by decoupling logic and memory into chiplets fabricated in heterogeneous technology nodes and connected via an interposer. The logic chiplets sustain high bandwidth access to the DRAM chiplets, which house the memory banks, and enable the integration of advanced processing components such as systolic arrays and SRAM-based buffers to accelerate memory-bound GEMM kernels, capabilities that were not feasible in prior PIM architectures. We propose Sangam, a CXL-attached PIM-chiplet based memory module that can either act as a drop-in replacement for GPUs or co-executes along side the GPUs. Sangam achieves speedup of 3.93, 4.22, 2.82x speedup in end-to-end query latency, 10.3, 9.5, 6.36x greater decoding throughput, and order of magnitude energy savings compared to an H100 GPU for varying input size, output length, and batch size on LLaMA 2-7B, Mistral-7B, and LLaMA 3-70B, respectively.

Heterogeneous chiplet-based systems improve scaling by disag-gregating CPUs/GPUs and emerging technologies (HBM/DRAM).However this on-package disaggregation introduces a latency inNetwork-on-Interposer(NoI). We observe that in modern large-modelinference, parameters and activations routinely move backand forth from HBM/DRAM, injecting large, bursty flows into theinterposer. These memory-driven transfers inflate tail latency andviolate Service Level Agreements (SLAs) across k-ary n-cube base-line NoI topologies. To address this gap we introduce an InterferenceScore (IS) that quantifies worst-case slowdown under contention.We then formulate NoI synthesis as a multi-objective optimization(MOO) problem. We develop PARL (Partition-Aware ReinforcementLearner), a topology generator that balances throughput, latency,and power. PARL-generated topologies reduce contention at the memory cut, meet SLAs, and cut worst-case slowdown to 1.2 times while maintaining competitive mean throughput relative to link-rich meshes. Overall, this reframes NoI design for heterogeneouschiplet accelerators with workload-aware objectives.

Neuromorphic computing promises brain-like efficiency, yet today's multi-chip systems scale over PCBs and incur orders-of-magnitude penalties in bandwidth, latency, and energy, undermining biological algorithms and system efficiency. We present DarwinWafer, a hyperscale system-on-wafer that replaces off-chip interconnects with wafer-scale, high-density integration of 64 Darwin3 chiplets on a 300 mm silicon interposer. A GALS NoC within each chiplet and an AER-based asynchronous wafer fabric with hierarchical time-step synchronization provide low-latency, coherent operation across the wafer. Each chiplet implements 2.35 M neurons and 0.1 B synapses, yielding 0.15 B neurons and 6.4 B synapses per wafer.At 333 MHz and 0.8 V, DarwinWafer consumes ~100 W and achieves 4.9 pJ/SOP, with 64 TSOPS peak throughput (0.64 TSOPS/W). Realization is enabled by a holistic chiplet-interposer co-design flow (including an in-house interposer-bump planner with early SI/PI and electro-thermal closure) and a warpage-tolerant assembly that fans out I/O via PCBlets and compliant pogo-pin connections, enabling robust, demountable wafer-to-board integration. Measurements confirm 10 mV supply droop and a uniform thermal profile (34-36 °C) under ~100 W. Application studies demonstrate whole-brain simulations: two zebrafish brains per chiplet with high connectivity fidelity (Spearman r = 0.896) and a mouse brain mapped across 32 chiplets (r = 0.645). To our knowledge, DarwinWafer represents a pioneering demonstration of wafer-scale neuromorphic computing, establishing a viable and scalable path toward large-scale, brain-like computation on silicon by replacing PCB-level interconnects with high-density, on-wafer integration.

EdgeCortix, a Tokyo-based artificial intelligence (AI) chip startup, is riding a wave of interest to foster Japanese semiconductors with defense applications. EdgeCortix, which has won a contract tied to the U.S. Department of Defense, on Wednesday secured government subsidies of 3 billion ( 21 million) to develop energy-efficient AI chiplets for commercialization in 2027. The contract may help revenue more than double this year, founder Sakyasingha Dasgupta said. The products, designed to help robots make real-time decisions and fill the country's labor shortages, target mass production at Taiwan Semiconductor Manufacturing Co.'s plant in Japan. The subsidies are on top of 4 billion in support the semiconductor designer won in November to make chips for next-generation communication systems.

The insatiable appetite of Artificial Intelligence (AI) workloads for computing power is pushing the industry to develop faster and more efficient accelerators. The rigidity of custom hardware, however, conflicts with the need for scalable and versatile architectures capable of catering to the needs of the evolving and heterogeneous pool of Machine Learning (ML) models in the literature. In this context, multi-chiplet architectures assembling multiple (perhaps heterogeneous) accelerators are an appealing option that is unfortunately hindered by the still rigid and inefficient chip-to-chip interconnects. In this paper, we explore the potential of wireless technology as a complement to existing wired interconnects in this multi-chiplet approach. Using an evaluation framework from the state-of-the-art, we show that wireless interconnects can lead to speedups of 10% on average and 20% maximum. We also highlight the importance of load balancing between the wired and wireless interconnects, which will be further explored in future work.

We study the application of emerging chiplet-based Neural Processing Units to accelerate vehicular AI perception workloads in constrained automotive settings. The motivation stems from how chiplets technology is becoming integral to emerging vehicular architectures, providing a cost-effective trade-off between performance, modularity, and customization; and from perception models being the most computationally demanding workloads in a autonomous driving system. Using the Tesla Autopilot perception pipeline as a case study, we first breakdown its constituent models and profile their performance on different chiplet accelerators. From the insights, we propose a novel scheduling strategy to efficiently deploy perception workloads on multi-chip AI accelerators. Our experiments using a standard DNN performance simulator, MAESTRO, show our approach realizes 82% and 2.8x increase in throughput and processing engines utilization compared to monolithic accelerator designs.

With the growing need for higher memory bandwidth and computation density, 2.5D design, which involves integrating multiple chiplets onto an interposer, emerges as a promising solution. However, this integration introduces significant challenges due to increasing data rates and a large number of I/Os, necessitating advanced optimization of the power distribution networks (PDNs) both on-chip and on-interposer to mitigate the small signal noise and simultaneous switching noise (SSN). Traditional PDN optimization strategies in 2.5D systems primarily focus on reducing impedance by integrating decoupling capacitors (decaps) to lessen small signal noises. Unfortunately, relying solely on frequency-domain analysis has been proven inadequate for addressing coupled SSN, as indicated by our experimental results. In this work, we introduce a novel two-phase optimization flow using deep reinforcement learning to tackle both the on-chip small signal noise and SSN. Initially, we optimize the impedance in the frequency domain to maintain the small signal noise within acceptable limits while avoiding over-design. Subsequently, in the time domain, we refine the PDN to minimize the voltage violation integral (VVI), a more accurate measure of SSN severity. To the best of our knowledge, this is the first dual-domain optimization strategy that simultaneously addresses both the small signal noise and SSN propagation through strategic decap placement in on-chip and on-interposer PDNs, offering a significant step forward in the design of robust PDNs for 2.5D integrated systems.

Emerging multi-model workloads with heavy models like recent large language models significantly increased the compute and memory demands on hardware. To address such increasing demands, designing a scalable hardware architecture became a key problem. Among recent solutions, the 2.5D silicon interposer multi-chip module (MCM)-based AI accelerator has been actively explored as a promising scalable solution due to their significant benefits in the low engineering cost and composability. However, previous MCM accelerators are based on homogeneous architectures with fixed dataflow, which encounter major challenges from highly heterogeneous multi-model workloads due to their limited workload adaptivity. Therefore, in this work, we explore the opportunity in the heterogeneous dataflow MCM AI accelerators. We identify the scheduling of multi-model workload on heterogeneous dataflow MCM AI accelerator is an important and challenging problem due to its significance and scale, which reaches O(10^18) scale even for a single model case on 6x6 chiplets. We develop a set of heuristics to navigate the huge scheduling space and codify them into a scheduler with advanced techniques such as inter-chiplet pipelining. Our evaluation on ten multi-model workload scenarios for datacenter multitenancy and AR/VR use-cases has shown the efficacy of our approach, achieving on average 35.3% and 31.4% less energy-delay product (EDP) for the respective applications settings compared to homogeneous baselines.

Chiplet-based systems, integrating various silicon dies manufactured at different integrated circuit technology nodes on a carrier interposer, have garnered significant attention in recent years due to their cost-effectiveness and competitive performance. The widespread adoption of reinforcement learning as a sequential placement method has introduced a new challenge in determining the optimal placement order for each chiplet. The order in which chiplets are placed on the interposer influences the spatial resources available for earlier and later placed chiplets, making the placement results highly sensitive to the sequence of chiplet placement. To address these challenges, we propose a learning to rank approach with graph representation, building upon the reinforcement learning framework RLPlanner. This method aims to select the optimal chiplet placement order for each chiplet-based system. Experimental results demonstrate that compared to placement order obtained solely based on the descending order of the chiplet area and the number of interconnect wires between the chiplets, utilizing the placement order obtained from the learning to rank network leads to further improvements in system temperature and inter-chiplet wirelength. Specifically, applying the top-ranked placement order obtained from the learning to rank network results in a 10.05% reduction in total inter-chiplet wirelength and a 1.01% improvement in peak system temperature during the chiplet placement process.