CS Paper Radar

Weekly top 20 · cs.AR

2026-W47 Nov 16 – Nov 22, 2026

Ranked from 1,465 papers submitted on 2 days — week in progress, ranks can still change.

Score = weighted mean ofCitations 20%Upvotes 10%Code stars 10%Venue 15%Code/data 10%Novelty 15%Momentum 15%Substance 5%how it works
1

One Mean Can't Rule Them All

cs.AR ·2026-11-16 ·Ahmed Nematallah
published
Abstract · 9 pages, 4 figures, 2 tables. Author's accepted manuscript, licensed CC BY 4.0. Published in IEEE Computer, vol. 59, no. 2, pp. 49-58, Febru…

There has been extensive debate on how to summarize performance results, especially within the computer architecture community. This article presents a systematic approach to choosing the correct mean, aiming to answer this question for all computer scientists once and for all.

0.65
2

ThermLeT: Transformer-based Temperature Prediction for 2.5D Chiplet Architecture

cs.AR cs.AIcs.LG ·2026-11-16 ·Varun Darshana Parekh, Anusha Devulapally, Sivani Devarapalli, Cassius Henderson, Shimeng Yu +1 more
publishedcode/data
Abstract · Published at VLSID 2026. Code and dataset: https://github.com/MicroSystemsDesignLab/ThermLeT

The drive toward heterogeneous 2.5D chiplet packages promises unmatched performance and customization, yet the resulting density amplifies thermal challenges. To facilitate rapid design space exploration, we present ThermLeT, a multimodal transformer-based, scale invariant surrogate that integrates power density maps, spatial coordinate embeddings, material properties and heat transfer rate to predict steady state temperature in 13 ms, about four orders of magnitude faster than conventional finite element solvers such as ANSYS. To train the transformer model, we also create a dataset of 6,221 chiplet layouts perturbed from three representative architectures. On a 150 sample test set, ThermLeT achieves an RMSE of 3.16°C and an MAE of 2.24°C, improving accuracy by 35% over ThermEDGe and by 71% over HotSpot, while running approximately 6x and 11x lower inference latencies respectively. Ablation studies confirm robust generalization to unseen floorplans. Code and dataset available at https://github.com/MicroSystemsDesignLab/ThermLeT

0.43
3

Clock Distribution Network with Backside Clock Mesh in BSPDN Architecture

cs.AR ·2026-11-16 ·Seohyun Kim, Youngjin Chae, Shilong Zhang, Youngsoo Shin
top venue
Abstract · Accepted to ASP-DAC 2027

We propose a clock distribution network (CDN) in which the clock mesh is located on backside metal (BM) layers, while the rest of the network remains on the frontside. Since the backside power distribution network (BSPDN) also uses the BM layers, the clock mesh and power network must share them. We identify an important problem of placing mesh receivers while accounting for keep-out zones (KOZs) between clock- and power-nTSVs. We show that this problem can be solved using a potential-field method (PFM). Machine learning models are developed to place the mesh receivers: a U-Net predicts their initial positions, and an MLP predicts force weights for PFM-guided position refinement. Both models are trained through self-supervised learning by optimizing a differentiable objective that accounts for clock latency, IR-drop, and KOZ violations. Experimental results show that, on average, the proposed CDN reduces maximum local clock latency by 26%, global clock skew by 40%, and mesh switching power by 29% compared with a frontside CDN. On the same backside CDN, ML-guided PFM satisfies KOZ constraints while lowering IR-drop by 33% and requiring 85% fewer iterations than standard PFM.

0.40
4

Do Apple M-Series Processors Enforce the Compound Memory Model? Sequential and Parallel Composition of Arm and TSO

cs.AR cs.PL ·2026-11-16 ·Vijay Nagarajan
code/data
Abstract · 26 pages, 6 figures, 11 tables. Artifact: https://github.com/vijay-nagarajan/apple-tso-litmus

Apple's M-series processors can run each thread under the Arm memory model or under a TSO model, and threads of both kinds can share memory. This raises two questions of composition that Apple has not answered. Across threads, how does the machine behave as a whole? Within a thread, which Arm instructions play the roles of TSO's loads, stores and atomic read-modify-writes, and how are they ordered between themselves and across other Arm instructions? We propose that both answers are compositions of preserved program order. Across threads, each thread keeps the preserved program order of its own model: a parallel composition. Within a TSO thread, every Arm instruction keeps TSO's order and adds Arm's, and every atomic memory operation (AMO) is locked, as x86's read-modify-writes are: a sequential composition. We test both on an Apple M1 and an Apple M2 running Linux, with 936 litmus tests chosen from the two models' preserved program order, each run 6x10^8 times on each machine. Arm-mode threads follow Arm's official model, TSO-mode threads follow x86-TSO and the sequential composition, and programs that mix the two modes follow the compound model: neither machine shows an outcome these models forbid. Both machines also preserve three orders that these models do not require, and a compare-and-swap whose comparison fails behaves as if it writes back the value it read, which the Arm model does not require. All three orders held in runs up to 40 times longer.

0.40
5

Gateway-Free Kafka Streaming from FPGA-Based Detectors: A Bare-Metal Pipelined Producer on Zynq-7000

cs.NI cs.AR ·2026-11-17 ·Fernanda Zapata Bascuñán, Matías Hampel
published
Abstract · Accepted for pucblishing at IEEE Explore proceedings of ARGENCON 2026

We present a bare-metal Apache Kafka producer implemented on a Zynq-7000 System-on-Chip (SoC) for event-driven data acquisition. Events are generated in Programmable Logic (PL), transferred to the Processing System (PS) through AXI4-Stream and AXI DMA, and published to a standard Kafka broker without an intermediate gateway computer. The PS constructs ProduceRequest messages and transmits them through the lwIP raw TCP/IP API. A configurable pipeline limits the number of outstanding Kafka requests. Pipeline depths of D = 1, 2, 4, and 8 were evaluated with two independent DMA sources. For each depth, 20 independent runs of 5,000 events were acquired, for 400,000 event records in total. Each event carries a 32-bit frame counter and a 32-bit PL timestamp. PS-side instrumentation records DMA start and completion, publication, successful local tcp_write() acceptance, and reception of the corresponding Kafka acknowledgement. Measurement records are exported to a separate Kafka topic after the timed event stream has completed. Each of the 80 runs contained 5,000 records with frame values from 0 to 4,999, with no sequence discontinuities in the exported metric records and no recorded Kafka errors. Mean operational throughput was approximately 5.3 kevents/s at D = 1, 9.4-9.6 kevents/s at D = 2, 16.2 kevents/s at D = 4, and 16.8-17.5 kevents/s at D = 8. The median publish-to-acknowledgement interval was 1.62-1.64 ms at D = 1, 0.98 ms at D = 2, 0.66 ms at D = 4, and 0.83-0.86 ms at D = 8. At D = 4, the corresponding P95 values were 0.81-0.83 ms.

0.40
6

From Token Changes to Task Failures: Bit Flips in a Shared KV-Cache Block

cs.CR cs.AR ·2026-11-17 ·Yuji Yamamoto, Satoshi Matsuura
Abstract · 19 pages, 4 figures. Extended version of arXiv:2604.17249 (SECRYPT 2026, 23rd International Conference on Security and Cryptography). Submit…

In LLM serving systems with prefix caching, each KV block of a shared system prompt is held in GPU memory as a single physical copy, which every request with that prompt reads. Our earlier work measured change and showed that one bit flip in such a block changes many later responses without making them visibly degenerate. This paper extends that work by measuring how often the changed responses are wrong. Using software fault injection into vLLM serving Qwen3-8B, we flip the most significant mantissa bit or the sign bit of a value in such a block. We score responses on four tasks: code generation, short-answer questions, multiple-choice reasoning, and refusal under a policy. The fraction of changes that turned a correct response wrong depended strongly on the task: 0.023 on MBPP+ and 0.59 on held-out TriviaQA at the mantissa bit, and 0.019 and 0.60 at the sign bit. Harm fell on requests whose sampled answers are sometimes wrong, so the damage of a flip depends on how common such requests are. Harmed responses looked like ordinary errors, so a flip should be detected by checking cached blocks rather than by monitoring outputs.

0.40
7

Computed Permutation Trellis Coding: Near-Optimal Quantization from Two Multiplications and a 64-Byte Table

cs.IT cs.AIcs.ARcs.LG ·2026-11-17 ·Muhammed Fatih Balın
Abstract · 38 pages, 6 figures, 27 tables

We propose Computed Permutation Trellis Coding (CPTC), a low-bit quantization format for tensors written once and read many times. Formats that round each value on its own, such as INT4 and FP4, waste a large part of their bits against the information-theoretic ideal. Trellis codes recover this loss by encoding values jointly, but prior decoders either need a large lookup table for their best quality or compute values with a hash that restricts what they can decode. CPTC's decoder multiplies its state by two coefficients and adds the top bits of the products to index a 64-byte table. Because that sum is not uniformly distributed, the rarely used ends of a small table can hold values far out in the tails. One decoder serves every bit width from 1 to 6, reproduces the standard integer and floating-point formats exactly, and comes within 0.07 bits of the ideal at 4 bits. It matches a random-permutation trellis code whose table is thousands of times larger, and exceeds every QTIP codebook. In LLM weight quantization without fine-tuning, CPTC has the lowest KL divergence to the unquantized model at every tested bit width on Llama 3.1 8B and 70B Instruct against QTIP's codebooks (by up to 27%). The decoder has been built into the AWS Trainium 4 AI accelerator, where it runs at the throughput of the native FP4 and FP8 formats for a negligible fraction of the systolic array's area. Fused encoding kernels quantize an 8B model's linear layers in under two minutes on one GPU and, in cycle-level simulation, on one Trainium 4 chip.

0.40
8

blaris: Windowless Random-Access LZSS for Constrained Microcontrollers

cs.AR cs.IT ·2026-11-17 ·Ihor Lukianov
code/data
Abstract · 4 pages, 1 figure, 2 tables. Source code and benchmark suite available at https://github.com/lis05/blaris

Embedded LZ77 decompressors (such as heatshrink and tamp) require an in-RAM sliding window that couples memory usage to compression ratio and incurs high latency on random-access reads. We present blaris, a lightweight, windowless LZSS decompressor for execute-in-place (XIP) random access on constrained microcontrollers. By partitioning tokens into bidirectionally traversable streams with power-of-two checkpoints, blaris resolves back-references directly in the compressed payload, achieving compression ratios comparable to high-window heatshrink and tamp configurations without allocating a RAM window. Targeting compress-on-host, decompress-on-device workflows, blaris assumes trusted build-time payloads and omits runtime error and bounds checks. On RV32IMC, it requires just 548-784 B of Flash and 104-112 B of total RAM, achieving orders-of-magnitude faster small-slice random reads than ratio-matched, block-indexed heatshrink and tamp.

0.30
9

LASH-PAT: Deadlock-Free Adaptive Shortest-Path Routing for Inter-Chiplet Interconnects

cs.AR cs.NI ·2026-11-17 ·Patrick Iff, Torsten Hoefler
Abstract

With the slowdown of transistor scaling, chipletbased systems, and in particular chiplet-based wafer-scale integration, have emerged as a promising approach to sustain growth in computing capability per cost for modern processors and accelerators. In recent years, many specialized inter-chiplet interconnect topologies have been proposed. In these topologies, the simple, well-established, and deadlock-free dimension-order routing scheme is not applicable due to its reliance on regular mesh structures. In this work, we study routing in these recently proposed inter-chiplet interconnect topologies by categorizing and evaluating nine deadlock-free routing functions for irregular topologies known from data center networking in this new context. This reveals a trade-off between turn-based routing, which enables path diversity, and layered routing, which ensures shortest-path delivery. To resolve this trade-off, we propose LASH-PAT, a routing function that extends layered shortestpath routing with adaptive path diversity. LASH-PAT supports multiple shortest paths and dynamically selects among them based on local congestion while preserving deadlock freedom. When compared to the best-performing baseline routing function for each combination of topology, traffic pattern, and evaluated metric, LASH-PAT improves throughput by 21.15% and reduces energy per packet by 12.77%, while providing latency comparable to the lowest-latency baseline routing function.

0.30
10

Mixed-Size Placement Initialization via Stiefel-Manifold Sequential Subspace Optimization

cs.AR ·2026-11-17 ·Pengwen Chen, Chung-Kuan Cheng, Shao-Yun Fang, Chester Holtz, Sheng-Tan Huang +3 more
Abstract · 14 pages, 6 figures, 5 tables. Submitted to IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD)

Mixed-size global placement simultaneously optimizes the locations of large macros and millions of standard cells in modern very-large-scale integration (VLSI) physical design. State-of-the-art analytical placers achieve strong scalability, but their nonlinear objectives and density handling make them sensitive to the initial placement. This paper presents a geometry-driven framework that formulates placement initialization as a Quadratically Constrained Quadratic Program (QCQP) with first-order centroid and second-order spreading constraints. Explicit mixed-size spreading targets are derived from the die geometry and object sizes, to obtain stable constraint values even under extreme size variation. A utilization-aware rescaling of the constraint matrix prevents pathological spreading when fixed objects occupy significant die area. The QCQP is then transformed into a Stiefel-manifold optimization through metric equalization and centering projection, making the underlying geometric structure explicit. The resulting Stiefel-constrained quadratic objective admits efficient large-scale optimization via a Riemannian sequential subspace method (SSM). An angle-sorted cycle sparsification for multi-pin nets periodically keeps the Laplacian approximation sparse yet geometry-consistent. As a drop-in initializer for existing analytical placers, our method yields improvements in weighted half-perimeter wirelength (wHPWL): 13.8% average under RePlAce on the four TILOS benchmarks supported by its flow, 3.5% under DREAMPlace on all 20 benchmarks, and 4.8% under AutoDMP on 19 benchmarks with converged baselines. Post-route evaluation shows that these pre-route improvements carry through to routed designs: routed wirelength decreases by 1.1 to 2.3% on all three benchmarks. Standard-cell area and power also decrease on all three. Design-rule-check (DRC) violations decrease on two of the three.

0.30
11

QuarkDB: A Hybrid Database System for Quantum Data Management

cs.DB cs.AR ·2026-11-16 ·Wenrui Zhang, Yu Wang, Xiao Zhang
Abstract · 14 pages, 12 figures, Submitted to ICDE2027

Emerging quantum applications like quantum communication generate large volumes of high-dimensional quantum data, requiring efficient data analysis and storage. However, classical database systems often struggle to meet requirements such as quantum processing or version management. This paper presents QuarkDB designed for efficient quantum data management. QuarkDB establishes a unified object model for relational and quantum data and defines the syntax of Quantum Structured Query Language. To support efficient quantum data management, QuarkDB integrates three key components: (1) a Hybrid Query Engine with cost-based hybrid plan selection over classical and quantum operators, (2) an Execution Engine with a three-level instructions intermediate representation, QuarkDB-IR, to support quantum execution and reduce quantum processing costs, and (3) a Storage Engine supporting quantum data version management, state reconstruction and data reuse. Compared with classical databases such as SQLite and PostgreSQL or quantum databases like Qute, our QuarkDB achieves query-latency speedups ranging from 2.14x to 25.86x speedup under extensive real-world workloads.

0.28
12

PAISF: A Partially Adaptive In-Order Selection Function for Routing in On-Chip, Inter-Chiplet, and Wafer-Scale Networks

cs.AR cs.NI ·2026-11-17 ·Patrick Iff, Luca Oertig, Andrin Weber, Luca Benini, Torsten Hoefler
Abstract

The growing number of compute elements per chip, along with advances such as chiplet systems and wafer-scale integration, drives the need for scalable and efficient routing algorithms for irregular topologies. A routing algorithm consists of a routing function (RF) that ensures deadlock freedom and a selection function (SF) that avoids congestion and, in some cases, guarantees in-order packet delivery. We propose two novel SFs for irregular topologies, ISF and PAISF, which ensure in-order delivery by hashing a packet's entropy value to determine its path. Packets with the same entropy follow the same path and arrive in order. Compared to a deterministic RF, a deterministic SF, and Multi, three alternative in-order delivery strategies, (PA)ISF improves throughput by 15.2% and reduces energy per packet and latency under load by 12.1% and 11.3%, at the cost of a zero-load latency increase by 3.3% on average across workloads and topologies.

0.28
13

KD-DSE: A Knowledge-Driven LLM Framework for Fast Microarchitecture Design Space Exploration

cs.AR ·2026-11-17 ·Siqing Fu, Xin Liu, Yuhan Tang, Hanqing Li, Feng Guo
Abstract · 4 pages, 4 figures

Domain-specific architectures have made design space exploration (DSE) a bottleneck of processor design, yet each candidate costs a full simulation, so exhaustive search is out of reach. Surrogate models need thousands of simulations before they predict reliably, and LLM agents that search directly hallucinate, lack cold-start architectural knowledge, and follow a fixed per-round policy. We present KD-DSE, a knowledge- driven LLM framework for fast microarchitecture DSE. KD- DSE couples a five-state constrained agent with real gem5- McPAT evaluation, retrieves architect-curated guidance from an Architecture Knowledge Base conditioned on the current search state and bottleneck diagnosis, and allocates its budget with a diagnosis-guided two-phase exploration strategy. On a 29- parameter out-of-order core with a budget of 18 simulations, KD-DSE achieves about 4.8% higher hypervolume and about 21.7% lower ADRS than the state-of-the-art MicroEvo, and the ablation attributes the gain chiefly to the Knowledge Base. These results indicate that constraining LLM reasoning and grounding it in architectural knowledge can drive fast and effective microarchitecture DSE. The code will be open-sourced upon acceptance.

0.27
14

StreamingDPD: A 1.28-GSa/s Instantaneous-Power-Gated Depthwise TCN Accelerator for Real-Time 320-MHz Digital Predistortion

eess.SP cs.AR ·2026-11-17 ·Yizhuo Wu, Xinling Wang, Chang Gao
Abstract

State feedback between successive samples limits the streaming throughput of recurrent neural networks for digital predistortion (DPD). We present StreamingDPD, an accelerator that processes 320-MHz signals at 1.28 GSa/s. Its 207-parameter instantaneous-power-gated depthwise temporal convolutional network (IPG-DWTCN) uses $|x[n]|^2$ to gate feature channels and causal dilated depthwise convolutions to model power amplifier memory. The network maps to a spatially unrolled fixed-point pipeline with shift-based quadratic activations and layer-wise mixed precision, sustaining one complex sample per cycle per lane. Its algorithmic cost is 554 floating-point operations per complex sample, 30.1% fewer than the recurrent baseline with the best linearization performance. In RF playback measurements on three power amplifiers driven by 40- to 320-MHz Wi-Fi 6/7 and 5G NR waveforms, IPG-DWTCN achieves the lowest mean adjacent-channel power ratio (ACPR) and ranks first or ties for first in mean error vector magnitude (EVM) against five neural baselines of comparable size. Four-lane FPGA hardware-in-the-loop tests on a 320-MHz 4096-QAM Wi-Fi 7 waveform achieve -45.82 dBc ACPR and -42.94 dB EVM at 1.28 GSa/s. At the same sample rate, post-layout evaluation of a single-lane 12-nm FinFET ASIC yields a 0.10 mm$^2$ core, a 52-cycle (40.63 ns) latency, and an estimated typical-corner power of 142.21 mW (111.10 pJ/sample). Relative to the prior 22-nm recurrent DPD ASIC, the design provides 5.1 times the throughput at about one-seventh the energy per sample.

0.27
15

Architectural Principles for Parameter-Efficient Unitary Transformations in Coherent Optical Neural Networks

cs.LG cs.ARphysics.app-phphysics.optics ·2026-11-16 ·Kazuo Aoyama, Hiroshi Sawada, Shota Kita, Akihiko Shinya, Masaya Notomi
Abstract · 28 pages, 28 figures

Coherent optical neural networks based on interferometric architectures offer a promising route toward high-throughput, energy-efficient computation by exploiting the multiple degrees of freedom of light, including spatial, temporal, and spectral modes. However, practical implementations are constrained by the number of tunable parameters due to fabrication complexity, device footprint, and control overhead. In such parameter-constrained settings, achieving high expressive power remains a central challenge. We identify two key architectural requirements for maintaining high expressive power under these constraints: (i) statistical uniformity in the implemented transformations and (ii) strong mixing capability per parameter, which enables rapid and unbiased redistribution of information across modes. To realize these requirements, we develop a logarithmic-depth butterfly-topology architecture whose fundamental building block is a single phase-shifter--beam-splitter pair, in contrast to conventional rectangular mesh architectures based on Mach-Zehnder interferometers. Computational evaluations show that, particularly at restricted parameter counts, the proposed architecture achieves lower matrix-reconstruction errors and higher classification accuracies than conventional meshes in both feedforward and recurrent networks. We further quantify these design principles using graph-theoretic measures, expected squared moduli of transformation-matrix entries, and matrix-reconstruction error. The resulting framework provides a quantitative and physically interpretable basis for parameter-efficient optical neural networks.

0.25
16

PVTSizing 2.0: A Fast and Reliable Analog Circuit Sizing Operator for Agentic EDA Workflows

cs.AR cs.LG ·2026-11-17 ·Zichen Kong, Minghuan Song, Xiyuan Tang, Ziming Wei, Yi Zhong +1 more
Abstract · 10 pages, 4 figures, 5 tables

Analog circuit design automation requires both coordinated tool execution and efficient numerical search. Large language models (LLMs) offer new ways to organize tool execution, but finding one sizing that satisfies specifications across process, voltage, and temperature (PVT) corners remains costly. PVTSizing 2.0 addresses these needs through a human-agent workflow built around an independently usable numerical optimizer. An agent carries the designer's approved task through execution and result review. The optimizer reduces unnecessary training through loss-guided updates and uses critic-guided local refinement to explore additional candidates while accounting for cross-corner conflicts. An optional LLM adviser supplies parameter relationships to focus numerical search on circuit-informed directions, while the optimizer selects adjustment amplitudes and verifies their effects through simulation. Across four circuit benchmarks, the standalone optimizer completes all 40 runs and achieves a 7.3x speedup in overall mean runtime over the original PVTSizing, with a comparable mean sample count. In a separate study, the best tested knowledge-injection configuration further reduces mean runtime by 39.9% and simulation samples by 35.4% relative to the standalone optimizer.

0.24
17

Accelerating 3D Gaussian Splatting Training through Selective Freezing and Computation Caching

cs.AR ·2026-11-17 ·He Zhu, Zhiheng Dai, Zihan Liu, Ruyang Li, Hui Wei +7 more
hot direction
Abstract

3D Gaussian Splatting (3DGS) becomes a key primitive for immersive 3D experiences. However, its applicability is limited by its long training overhead. Although prior work has shown that gradient accumulation is the dominant bottleneck in training, existing works either have limited speedup or degrade reconstruction quality. To address this issue, we present ElsaGS, an algorithm-hardware co-designed framework that accelerates 3DGS training with negligible accuracy degradation. Our key observation is that most Gaussians converge early, while later iterations only refine a small subset of Gaussians. Thus, we propose a technique, progressive Gaussian freezing, which dynamically freezes well-converged Gaussians and caches their cumulative contributions, thereby reducing computations in both forward rasterization and backward gradient accumulation. To make caching efficient, we further propose an architecture, ElsaGS core, which integrates a pixel-parallel rasterization array with a set-associative frozen Gaussian cache, FGcache. Across three 3DGS datasets, ElsaGS freezes up to 90% of Gaussians with negligible quality loss and achieves up to 20.6x end-to-end speedup and 99.4% energy reduction over GPUs.

0.22
18

How well does graph superoptimization transfer? Adapting tensor-graph superoptimizer to a Triton-based NPU

cs.PF cs.ARcs.PL ·2026-11-17 ·Weronika Hryniewska-Guzik, Mateusz Comblik, Daniel Murawski, Jakub Schmiegel, Maciej Świetlik +2 more
published
Abstract · Accepted for FlagOS Forum at MICRO 2026 - OASIS-AI workshop: Open Compute Architecture and Software Infrastructure for Heterogeneous AI Syst…

For automatic optimization systems, performance portability across increasingly diverse AI accelerators remains a challenge. We investigate how well graph-level superoptimization transfers across architectures by adapting Mirage from its CUDA-oriented design to the Huawei Ascend 910B3 through a Triton-based compilation stack. We evaluate the adapted system on transformer-relevant workloads, including matrix multiplication, softmax, RMSNorm+MatMul, gated MLP, FNN, and attention. The results show that graph-level optimization can successfully transfer and outperform available reference implementations by up to 87.0% on fused workloads. However, performance is highly dependent on tensor shape and workload structure, with substantial regressions for other configurations. We identify target-specific tensor partitioning, execution-resource modeling, backend lowering and fusion, and search and verification costs as key limitations. Our findings demonstrate that a portable graph representation and programming model do not imply a portable optimization space, motivating modular target-specific mechanisms within otherwise reusable compiler infrastructure.

0.20
19

Building Fully On-Device GUI Agents with Mobile NPUs

cs.AR ·2026-11-16 ·Zhe Dong, Guanyu Cai, Yuchen Jiang, Jiliang Wang
Abstract

Mobile GUI agents are increasingly used to automate everyday smartphone tasks, but most existing systems still rely on cloud-hosted models or remote inference, introducing repeated communication latency, privacy exposure, and dependence on network availability. Modern smartphones integrate neural processing units (NPUs) that provide an attractive substrate for fully on-device LLM inference. However, directly deploying conventional GUI-agent pipelines on mobile NPUs is inefficient. GUI-agent inference is fundamentally a long-input, short-output workload, yet conventional agents spend substantial time on autoregressive reasoning and action generation, while mobile NPUs provide much higher throughput for prompt prefill than token-by-token decoding. Meanwhile, dynamically varying GUI-agent contexts conflict with the fixed context profiles required by statically compiled NPU runtimes. We present DroidPrefill, a prefill-centric inference architecture that reshapes the GUI-agent workload around mobile-NPU execution. DroidPrefill combines Prefill-Centric Inference to reduce unnecessary autoregressive decoding with Efficient Context Management to preserve decision-critical information within a fixed compiled context budget. Experiments on commercial smartphones show that, in controlled comparisons with the autoregressive T3A+ baseline, DroidPrefill reduces per-step inference latency by up to 85.0%, while improving task success rate by up to 24.6%.

0.20
20

Integrated GPU Core for Mamba-3

cs.AR cs.LG ·2026-11-16 ·Shiva Raja, William Desueza, Ajay Joshi
Abstract

Mamba-3, the latest Mamba state-space model (SSM), enables efficient long-context modeling, matching Mamba-2 perplexity with half the state size and increasing decoding arithmetic intensity. However, its multi-input multi-output (MIMO) computation, exponential-trapezoidal discretization, and rotary positional embeddings (RoPE) introduce dense, element-wise, and sequential operations. Despite kernel fusion, conventional GPUs execute these operations inefficiently. Our characterization of Mamba-3 hybrid models reveals numerous kernel launches, substantial data movement, and increased register pressure that reduces warp occupancy and memory-latency hiding. We propose RescuCore, a specialized GPU execution unit for Mamba-3 inference that retains other hybrid model layers on existing GPU resources. Integrated alongside CUDA Cores and Tensor Cores within each Streaming Multiprocessor (SM) sub-partition, RescuCore shares the register file, shared memory, and warp schedulers. Dense projections remain on conventional GPU resources, while RescuCore pipelines discretization, normalization, RoPE, MIMO computation, recurrent state updates, and output gating to reduce kernel fragmentation and intermediate data movement. We select BF16 after evaluating multiple numerical formats, preserving model accuracy with an estimated area overhead of 5.9% of an A100 die for state size 32 and MIMO rank 4. We evaluate RescuCore using cycle-accurate Accel-Sim simulation, RTL synthesis, and synthesis-informed AccelWattch power modeling. Compared with conventional GPU execution, RescuCore achieves 1.4-9.4x speedup and 1.4-5.0x energy reduction during prefill, and 1.3-2.8x speedup and 1.2-2.6x energy reduction during autoregressive decoding. These results demonstrate the potential of GPU-integrated specialized execution for SSM-based hybrid models.

0.20