Research Blog
Exploring neural-symbolic systems, hardwired LLM inference, 3D chiplet integration, and diffusion-oriented accelerators. Early-stage concepts at the intersection of algorithms, architecture, and silicon.
AI Hardware Weekly Digest: WorldKV for Video World Models, Gated DeltaNet-2 Linear Attention, and HRM-Text Efficient Pretraining
AI Hardware Weekly Digest: WorldKV for Video World Models, Gated DeltaNet-2 Linear Attention, and HRM-Text Efficient Pretraining
AI Hardware Weekly Digest: Anthropic-Microsoft Maia 200 Deal Talks, LlamaWeb Browser Inference, and InnerQ Hardware-Aware Quantization
AI Hardware Weekly Digest: Anthropic-Microsoft Maia 200 Deal Talks, LlamaWeb Browser Inference, and InnerQ Hardware-Aware Quantization
AI Hardware Weekly Digest: Runtime-Certified Quantized Attention, BrainChip AKD1500 Mass Production, and Qwen3.7-Max 1M Context
AI Hardware Weekly Digest: Runtime-Certified Quantized Attention, BrainChip AKD1500 Mass Production, and Qwen3.7-Max 1M Context
AI Hardware Weekly Digest: Exciton-Polariton All-Optical Computing, Samsung Strike Deal, and Custom AI ASIC Landscape
AI Hardware Weekly Digest: Exciton-Polariton All-Optical Computing, Samsung Strike Deal, and Custom AI ASIC Landscape
AI Hardware Weekly Digest: TurboQuant 4.6× KV Cache Compression, RePlaid Continuous Diffusion Scaling, and CSIRO Vetra Edge AI
AI Hardware Weekly Digest: TurboQuant 4.6× KV Cache Compression, RePlaid Continuous Diffusion Scaling, and CSIRO Vetra Edge AI
AI Hardware Weekly Digest: TriAxialKV Mixed-Precision Quantization, KVDrive Multi-Tier Cache, and NVIDIA $106B Earnings
AI Hardware Weekly Digest: TriAxialKV Mixed-Precision Quantization, KVDrive Multi-Tier Cache, and NVIDIA $106B Earnings
AI Hardware Weekly Digest: SANA-WM Efficient World Model, Samsung HBM for Mobile, and Google I/O 2026
AI Hardware Weekly Digest: SANA-WM Efficient World Model, Samsung HBM for Mobile, and Google I/O 2026
AI Hardware Weekly Digest: CXMT 719% Revenue Surge, Memristor CIM Breakthrough, and Google I/O 2026
AI Hardware Weekly Digest: CXMT 719% Revenue Surge, Memristor CIM Breakthrough, and Google I/O 2026
AI Hardware Weekly Digest: KV-RM Static-Graph Serving, Quantization Security Risks, and Samsung Strike Threat
AI Hardware Weekly Digest: KV-RM Static-Graph Serving, Quantization Security Risks, and Samsung Strike Threat
AI Hardware Weekly Digest: World Action Models, Multi-Agent Coordination, and Cerebras IPO Pricing
AI Hardware Weekly Digest: World Action Models, Multi-Agent Coordination, and Cerebras IPO Pricing
AI Hardware Weekly Digest: FibQuant KV Cache, KV-Fold Recurrence, and Jensen Huang's China Trip
AI Hardware Weekly Digest: FibQuant KV Cache, KV-Fold Recurrence, and Jensen Huang’s China Trip
AI Hardware Weekly Digest: int4 KV Cache Beats fp16 on Apple Silicon, Federation of Experts, and Cerebras IPO
AI Hardware Weekly Digest: int4 KV Cache Beats fp16 on Apple Silicon, Federation of Experts, and Cerebras IPO
AI Hardware Weekly Digest: LaProx KV Cache, SpikingBrain, Cola DLM, and Intel's Neuromorphic Bet
AI Hardware Weekly Digest: LaProx KV Cache, SpikingBrain, Cola DLM, and Intel’s Neuromorphic Bet
AI 硬件研究周报（2026.05.11）：EA-WM 事件感知生成世界模型、RecursiveMAS 递归多智能体系统、机器人世界模型综述
AI 硬件研究周报（2026.05.11）：EA-WM 事件感知生成世界模型、RecursiveMAS 递归多智能体系统、机器人世界模型综述
AI 硬件研究周报（2026.05.10）：GYAN 神经符号语言模型、Embody4D 4D 世界模型、PV-VAE 预测性视频生成、ParoQuant 旋转量化
AI 硬件研究周报（2026.05.10）：GYAN 神经符号语言模型、Embody4D 4D 世界模型、PV-VAE 预测性视频生成、ParoQuant 旋转量化
AI 硬件研究周报（2026.05.07）：OpenAI 机器人硬件分拆上市、Broadcom 10GW 定制加速器、Flow Matching ODE 求解器硬件优化
AI 硬件研究周报（2026.05.07）：OpenAI 机器人硬件分拆上市、Broadcom 10GW 定制加速器、Flow Matching ODE 求解器硬件优化
AI 硬件研究周报（2026.05.06）：NEURON 神经符号临床系统、无 DRAM AI 推理芯片（Fractile）、WindowQuant VLM KV Cache 量化、SNN 无反向传播学习
AI 硬件研究周报（2026.05.06）：NEURON 神经符号临床系统、无 DRAM AI 推理芯片（Fractile）、WindowQuant VLM KV Cache 量化、SNN 无反向传播学习
AI 硬件研究周报（2026.05.05）：视觉生成五层范式演进、Dual-Blade 边缘 KV Cache 卸载、RISC-V 成为 AI 硬件开放基础
AI 硬件研究周报（2026.05.05）：视觉生成五层范式演进、Dual-Blade 边缘 KV Cache 卸载、RISC-V 成为 AI 硬件开放基础
AI 硬件研究周报（2026.05.04）：KV Cache 三维优化（DepthKV/PolyKV/CacheFlow）、HBM-PIM 张量加速、World-R1 几何一致性世界模型
AI 硬件研究周报（2026.05.04）：KV Cache 三维优化（DepthKV/PolyKV/CacheFlow）、HBM-PIM 张量加速、World-R1 几何一致性世界模型
AI 硬件研究周报（2026.05.03）：图世界模型统一范式、EdgeSpike 超低功耗 SNN 框架、HfO₂ 忆阻突触降低 70% 能耗
AI 硬件研究周报（2026.05.03）：图世界模型统一范式、EdgeSpike 超低功耗 SNN 框架、HfO₂ 忆阻突触降低 70% 能耗
AI 硬件研究周报（2026.05.02）：普林斯顿3D生物混合神经芯片、腾讯HY-Embodied具身模型、LLM递归自我改进的数学不可能性证明
AI 硬件研究周报（2026.05.02）：普林斯顿3D生物混合神经芯片、腾讯HY-Embodied具身模型、LLM递归自我改进的数学不可能性证明
AI 硬件研究周报（2026.05.01）：具身 AI 的 3D 生成综述、脉冲神经元逻辑电路、Motubrain 世界动作模型
AI 硬件研究周报（2026.05.01）：具身 AI 的 3D 生成综述、脉冲神经元逻辑电路、Motubrain 世界动作模型
Agentic Harness Engineering: 可观测性驱动的编码智能体 Harness 自动进化
Agentic Harness Engineering: 可观测性驱动的编码智能体 Harness 自动进化
AI 硬件研究周报（2026.04.30）：代理世界模型分类学、行为克隆缩放定律、信号折叠神经形态硬件
AI 硬件研究周报（2026.04.30）：代理世界模型分类学、行为克隆缩放定律、信号折叠神经形态硬件
AI 硬件研究周报（2026.04.29）：LingBot-Map 流式 3D 重建、DeepSeek V4 混合注意力架构、MOMO 机器人技能学习
AI 硬件研究周报（2026.04.29）：LingBot-Map 流式 3D 重建、DeepSeek V4 混合注意力架构、MOMO 机器人技能学习
AI 硬件研究周报（2026.04.28）：概率计算处理器、信号折叠神经形态硬件、低功耗计算机视觉挑战
AI 硬件研究周报（2026.04.28）：概率计算处理器、信号折叠神经形态硬件、低功耗计算机视觉挑战
AI 硬件研究周报（2026.04.26）：边缘 LLM 推理的 KV Cache 优化、CPU-GPU 混合注意力、跨数据中心 Prefill 服务
AI 硬件研究周报（2026.04.26）：边缘 LLM 推理的 KV Cache 优化、CPU-GPU 混合注意力、跨数据中心 Prefill 服务
Design Conductor: AI Agent 12小时自主设计1.5GHz RISC-V CPU
Design Conductor: AI Agent 12小时自主设计1.5GHz RISC-V CPU
AI 硬件研究周报（2026.04.18-04.25）：世界模型用于机器人训练、分子忆阻器神经形态硬件、NSF NeuroAI 路线图
AI 硬件研究周报（2026.04.18-04.25）：世界模型用于机器人训练、分子忆阻器神经形态硬件、NSF NeuroAI 路线图
AI 硬件研究周报（2026.04.18-04.24）：LLM 生成硬件的表示瓶颈、概率 Ising 机并行加速、KV Cache 神经垃圾回收
AI 硬件研究周报（2026.04.18-04.24）：LLM 生成硬件的表示瓶颈、概率 Ising 机并行加速、KV Cache 神经垃圾回收
Hardware-Efficient Neuro-Symbolic Networks with Exp-Minus-Log Operator
原文: arXiv:2604.13871 核心贡献: 提出 DNN-EML 架构，使用单一硬件可实现的 Sheffer 算子实现神经符号网络
The Price Is Not Right: Neuro-Symbolic AI Outperforms VLAs with 100x Lower Energy
原文: arXiv:2602.19260 作者: Timothy Duggan, Pierrick Lorang, Hong Lu, Matthias Scheutz 机构: Tufts University 核心贡献: 神经符号方法在结构化长视野操作任务上超越 VLA，能耗降低 100 倍
Switch-Centric In-Network Architecture for Accelerating LLM Inference
Switch-Centric In-Network Architecture for Accelerating LLM Inference
Neuromorphic Computing for Low-Power Artificial Intelligence
Neuromorphic Computing for Low-Power Artificial Intelligence
HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents
HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents
EPAC: The Last Dance - 欧洲RISC-V HPC加速器芯片的全栈实践
EPAC: The Last Dance - 欧洲RISC-V HPC加速器芯片的全栈实践
CUTEv2: 面向多样化CPU架构的统一可配置矩阵扩展
CUTEv2: 面向多样化CPU架构的统一可配置矩阵扩展
AI Hardware Weekly Digest - April 14, 2026
AI Hardware Weekly Digest - April 14, 2026
Build on Priors: 视觉-语言引导的神经符号模仿学习实现数据高效的机器人操作
Build on Priors: 视觉-语言引导的神经符号模仿学习实现数据高效的机器人操作
EvoSkills: 通过协同进化验证实现智能体技能的自我进化
EvoSkills: 通过协同进化验证实现智能体技能的自我进化
AI 硬件加速前沿：从 3D 堆叠内存到 LLM 解码优化
AI 硬件加速前沿：从 3D 堆叠内存到 LLM 解码优化
MicroScopiQ: 通过异常值感知微缩放量化加速基础模型
原文: arXiv:2411.05282 | PDF 会议: ISCA 2025, Tokyo, Japan 作者: Akshat Ramachandran, Souvik Kundu, Tushar Krishna 机构: Georgia Institute of Technology, Intel La...
TIE Scheduler: Uncertainty-Aware Output Length Prediction for Efficient LLM Inference Scheduling
TIE Scheduler: Uncertainty-Aware Output Length Prediction for Efficient LLM Inference Scheduling
SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning
SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning
Integer-State Dynamics of Quantized Spiking Neural Networks for Efficient Hardware Acceleration
Integer-State Dynamics of Quantized Spiking Neural Networks for Efficient Hardware Acceleration
SALS: 潜在空间稀疏注意力实现 KV Cache 压缩
原文链接: arXiv:2510.24273 PDF
Video Generation Models as World Models: Efficient Paradigms, Architectures and Algorithms Survey
Video Generation Models as World Models: Efficient Paradigms, Architectures and Algorithms Survey
GPU-FPGA Heterogeneous Systems for Disaggregated LLM Inference: Memory Processing Pipeline Acceleration
GPU-FPGA Heterogeneous Systems for Disaggregated LLM Inference: Memory Processing Pipeline Acceleration
EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards
EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards
SPINIC: Programmable Superconducting Neuron with In-Memory Computation for Ultra-Efficient Neuromorphic Computing
SPINIC: Programmable Superconducting Neuron with In-Memory Computation for Ultra-Efficient Neuromorphic Computing
RoboStereo: Dual-Tower 4D Embodied World Models for Unified Policy Optimization
RoboStereo: Dual-Tower 4D Embodied World Models for Unified Policy Optimization
GPU-FPGA 异构系统加速 LLM 推理中的内存处理
GPU-FPGA 异构系统加速 LLM 推理中的内存处理
AS2: Attention-Based Soft Answer Sets for End-to-End Differentiable Neuro-Soft-Symbolic Reasoning
AS2: Attention-Based Soft Answer Sets for End-to-End Differentiable Neuro-Soft-Symbolic Reasoning
Beyond GEMM-Centric NPUs: 高效扩散 LLM 采样架构
原文: arXiv:2601.20706 | PDF 作者: Binglei Lou, Haoran Wu, Yao Lai, et al. (Imperial College London, University of Cambridge) 核心贡献: 针对扩散 LLM 采样优化 NPU 架构，提出 d-...
异构计算：AI Agent 推理的未来关键
原文: arXiv:2601.22001 | PDF 作者: Aaron Zhao (Imperial College London), Junyi Liu (Microsoft Research) 核心贡献: 提出系统级异构计算是 AI Agent 推理的关键，识别”内存容量墙”问题
KernelCraft: 面向新兴硬件的Agentic底层内核生成基准测试
原文: arXiv:2603.08721 | PDF 作者: Jiayi Nie, Haoran Wu, et al. (University of Cambridge, Imperial College London, AMD, University of Edinburgh) 核心贡献: 首个评估 LL...
EQ-ViT: Algorithm-Hardware Co-Design for Real-Time Vision Transformer Acceleration on Versal ACAP
原文: EQ-ViT: Algorithm-Hardware Co-Design for End-to-End Acceleration of Real-Time Vision Transformer Inference on Versal ACAP Architecture 会议: ESWEEK 2024...
Neuromorphic Computing Roadmap: Scaling Brain-Inspired AI to Production
Neuromorphic Computing Roadmap: Scaling Brain-Inspired AI to Production
MediaTek Genio Pro: 50+ TOPS Edge AI Chip for Robotics and Embodied Intelligence
MediaTek Genio Pro: 50+ TOPS Edge AI Chip for Robotics and Embodied Intelligence
Innatera Synfire: Unifying the Neuromorphic Ecosystem for Edge AI
Innatera Synfire: Unifying the Neuromorphic Ecosystem for Edge AI
Google TurboQuant: 6x KV Cache Compression with Near-Optimal Distortion Rate
TurboQuant: Online Vector Quantization with Near-Optimal Distortion Rate
UCV: 通过软件原生优化普及和加速硬件验证
论文: Democratizing and Accelerating Hardware Verification with Software-Native Optimization 会议: ISCA 2026 核心贡献: UnityChip Verification (UCV) - 软件原生硬件验证平台
VMXDOTP: RISC-V Vector ISA Extension for Efficient Microscaling (MX) Format Acceleration
VMXDOTP: RISC-V Vector ISA Extension for Efficient Microscaling (MX) Format Acceleration
UCV: 软件原生硬件验证平台 - ISCA 2026
UCV: 通过软件原生优化实现硬件验证的民主化与加速
Programmable Superconducting Neuron for Ultra-Efficient Neuromorphic Computing
Programmable Superconducting Neuron for Ultra-Efficient Neuromorphic Computing
ReNN-RV: Run-time PE Reconfiguration for DNN Inference Acceleration with Custom RISC-V ISA
ReNN-RV: Run-time PE Reconfiguration for DNN Inference Acceleration with Custom RISC-V ISA
LeWorldModel: Stable End-to-End JEPA World Models from Pixels
LeWorldModel: Stable End-to-End JEPA World Models from Pixels
Helios: Hardware-Software Co-design for 3D-DRAM-based LLM Serving Accelerator
Helios: Hardware-Software Co-design for 3D-DRAM-based LLM Serving Accelerator
ChatNeuroSim: LLM Agent Framework for Automated CIM Accelerator Deployment
ChatNeuroSim: LLM Agent Framework for Automated CIM Accelerator Deployment and Optimization
AA-DiT: Algorithm-Architecture Co-Design for Diffusion Transformer Acceleration
AA-DiT: Algorithm-Architecture Co-Design for Diffusion Transformer Acceleration
PRISM: Photonic Similarity Engine for KV Cache Block Selection in Long-Context LLM Inference
PRISM: Photonic Similarity Engine for KV Cache Block Selection in Long-Context LLM Inference
PdNeuRAM: Forming-Free Multi-Bit ReRAM for Energy-Efficient Neuromorphic Computing
PdNeuRAM: Forming-Free, Multi-bit Pd/HfO₂ ReRAM for Energy-Efficient Neuromorphic Computing
Neuro-Symbolic AI Survey: Task-Directed Advances in the Black-Box Era
Neuro-Symbolic Artificial Intelligence: A Task-Directed Survey in the Black-Box Models Era
LeWorldModel: Stable End-to-End JEPA from Pixels for Embodied AI
LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels
Hummingbird+: Advancing FPGA-based LLM Deployment from Research Prototype to Edge Product
Hummingbird+: Advancing FPGA-based LLM Deployment from Research Prototype to Edge Product
Hummingbird: A Smaller and Faster Large Language Model Accelerator on Embedded FPGA
Hummingbird: A Smaller and Faster Large Language Model Accelerator on Embedded FPGA
ME-ViT: Memory-Efficient FPGA Accelerator for Vision Transformers
ME-ViT: A Single-Load Memory-Efficient FPGA Accelerator for Vision Transformers
Daily Research: Neuromorphic Computing & Spiking Neural Networks
🔍 Today’s Research Focus
Daily Research Roundup: LLM Hardware Acceleration & World Models
🔍 Today’s Research Focus
Understanding Bottlenecks for Efficiently Serving LLM Inference With KV Offloading
Understanding Bottlenecks for Efficiently Serving LLM Inference With KV Offloading
Speculating Experts Accelerates Inference for Mixture-of-Experts: 通过专家预取加速 MoE 推理
Speculating Experts Accelerates Inference for Mixture-of-Experts: 通过专家预取加速 MoE 推理
The Price Is Not Right: Neuro-Symbolic Methods Outperform VLAs on Structured Long-Horizon Manipulation Tasks
The Price Is Not Right: Neuro-Symbolic Methods Outperform VLAs on Structured Long-Horizon Manipulation Tasks
VLA-Perf: VLA 推理性能全景分析——NVIDIA 首个系统性研究
VLA-Perf: VLA 推理性能全景分析——NVIDIA 首个系统性研究
DS2SC-Agent: 从数据手册到 SystemC 模型的多智能体自动化生成流水线
DS2SC-Agent: 从数据手册到 SystemC 模型的多智能体自动化生成流水线
ZipServ: Fast and Memory-Efficient LLM Inference with Hardware-Aware Lossless Compression
原文链接: arXiv PDF
MINISA: Minimal Instruction Set Architecture for Next-gen Reconfigurable Inference Accelerator
原文链接: arXiv PDF
Large Video Planner: 基于视频生成的通用机器人控制新范式
Large Video Planner: 基于视频生成的通用机器人控制新范式
Large Video Planner: 用视频生成实现通用机器人控制
Large Video Planner: 用视频生成实现通用机器人控制
History-Guided Video Diffusion: 用历史引导实现超长视频生成
History-Guided Video Diffusion: 用历史引导实现超长视频生成
模型够聪明之后，工程师该做什么：Harness Engineering 实战指南
模型够聪明之后，工程师该做什么：Harness Engineering 实战指南
Design Conductor: AI 自主构建 1.5GHz RISC-V CPU 的突破性进展
Design Conductor: AI 自主构建 1.5GHz RISC-V CPU 的突破性进展
ZipServ: 硬件感知的无损压缩加速 LLM 推理
ZipServ: Fast and Memory-Efficient LLM Inference with Hardware-Aware Lossless Compression
Mozart: Modularized and Efficient MoE Training on 3.5D Wafer-Scale Chiplet Architectures
Mozart: Modularized and Efficient MoE Training on 3.5D Wafer-Scale Chiplet Architectures
HyperOffload: 图驱动的分层内存管理让大模型突破显存限制
HyperOffload: 图驱动的分层内存管理让大模型突破显存限制
PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation
PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation
Orion: 苹果神经引擎 (ANE) 上的 LLM 训练与推理系统
Orion: Characterizing and Programming Apple’s Neural Engine for LLM Training and Inference
History-Guided Video Diffusion: 用历史引导实现超长视频生成
History-Guided Video Diffusion: 用历史引导实现超长视频生成
GOMA: 通过解析建模实现空间加速器的几何最优映射
GOMA: Geometrically Optimal Mapping via Analytical Modeling for Spatial Accelerators
Synthesis-in-the-Loop Evaluation of LLMs for RTL Generation: Quality, Reliability, and Failure Modes
Synthesis-in-the-Loop Evaluation of LLMs for RTL Generation: Quality, Reliability, and Failure Modes
TOM: 三元只读存储器加速器赋能边缘智能大模型
TOM: 三元只读存储器加速器赋能边缘智能大模型
SNAP-V: 面向小型脉冲神经网络的可配置神经形态 RISC-V SoC
SNAP-V: 面向小型脉冲神经网络的可配置神经形态 RISC-V SoC
ROMA: 基于只读存储器的 QLoRA 边缘设备 LLM 加速器
ROMA: 基于只读存储器的 QLoRA 边缘设备 LLM 加速器
MedBayes-Lite: 临床 Transformer 的轻量级贝叶斯不确定性量化框架
MedBayes-Lite: 临床 Transformer 的轻量级贝叶斯不确定性量化框架
LLM 推理硬件的挑战与研究方向：内存与互连是核心瓶颈
LLM 推理硬件的挑战与研究方向：内存与互连是核心瓶颈
LEGOSim: 多芯片异构集成的统一并行仿真框架
LEGOSim: 多芯片异构集成的统一并行仿真框架
Taalas: 模型专用硬件 - 将AI模型转化为硅芯片
Taalas: 模型专用硬件 - 将AI模型转化为硅芯片
Taalas: Model-Specialized Hardware - Turning AI Models into Silicon
ROMA: 基于ROM的QLoRA边缘设备LLM加速器
ROMA: 基于ROM的QLoRA边缘设备LLM加速器
Neural-Symbolic AI Hardware: Unifying Pattern Learning and Logic
Why this direction matters
Hardwired LLM Accelerators: From Programmable Kernels to Fixed-Flow Inference
Motivation
Diffusion Model Accelerators: Efficient Sampling Beyond Brute-Force Denoising
Problem framing
3D Chiplet Systems for AI: Bandwidth-Centric Compute Integration
Why chiplets for AI now
HSCO-Bench: 首个端到端硬件软件协同设计基准测试
HSCO-Bench: An Agent-Driven End-to-End Hardware-Software Co-design Benchmark for Systems-on-Chip
CPPL: 面向 LLM 的电路提示编程语言
CPPL: A Circuit Prompt Programming Language
DeepStack: 分布式3D堆叠AI加速器的设计空间探索框架
DeepStack: 分布式3D堆叠AI加速器的设计空间探索框架
A Scalable Approach to Probabilistic Neuro-Symbolic Robustness Verification
A Scalable Approach to Probabilistic Neuro-Symbolic Robustness Verification
MicroScopiQ: 通过异常值感知微缩放量化加速基础模型
MicroScopiQ: 通过异常值感知微缩放量化加速基础模型