当前位置:首页 > 报告详情

RISC-V ISA 扩展及其在超维计算中的硬件加速.pdf

上传人: c** 编号:955303 2025-10-27 23页 1.39MB

1、RISC-V ISA Extensions with Hardware Acceleration for Hyperdimensional ComputingRocco Martino,PhD Candidaterocco.martinouniroma1.itRISC-V Summit Europe 2025.Paris,12-15 May.Outline2RISC-V Summit Europe 2025.Paris,12-15 May.RISC-V ISA Extension with Hardware Acceleration for HDCDIGITAL VLSICIRCUITS AN

2、D SYSTEMSRESEARCH GROUPIntroduction Hyperdimensional Computing Paradigm Why a custom extensionProposed Solution Klessydra T03 RISC-V Core The HDCUResults Hardware resource usage Speed up performanceMauro OlivieriOrdinary ProfessorFrancesco MenichelliAssistant ProfessorAntonio MastrandreaResearch Fel

3、lowAbdallah CheickResearch FellowMarcello BarbirottaResearch FellowSaeid JamiliPhD CandidateMarco AngioliPhD CandidateAndrea MarcelliPhD CandidateRocco MartinoPhD CandidateMarco PisaniPhD CandidateThe Hyperdimensional Computing paradigm for learning taskHyperdimensional Computing(HDC)1 is a consolid

4、ated computing paradigm that encodeinformation through distributed high-dimensional representations called hypervectors(HVs).3The mathematical space where HVs are manipulated is characterized by a very small set ofarithmetic vector operations:BindingBundlingPermutationSimilarity2DIGITAL VLSICIRCUITS

5、 AND SYSTEMSRESEARCH GROUPRISC-V Summit Europe 2025.Paris,12-15 May.RISC-V ISA Extension with Hardware Acceleration for HDC1 Kanerva,P.(2009).Hyperdimensional computing:An introduction to computing in distributed representation with high-dimensional random vectors.Cognitive Computation,1(2),139159.h

6、ttps:/doi.org/10.1007/s12559-009-9009-8By appropriately combining these operations it is possible to perform various learning taskssuch as classification,clustering and regression.4DIGITAL VLSICIRCUITS AND SYSTEMSRESEARCH GROUPRISC-V Summit Europe 2025.Paris,12-15 May.RISC-V ISA Extension with Hardw

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
根据文章内容,以下是全文关键点的概括: 1. **研究背景**:超维计算(HDC)是一种通过高维表示(超向量)进行信息编码的并行计算范式,适用于资源受限系统上的AI算法。 2. **RISC-V扩展**:为了加速HDC任务,研究人员设计了一种名为HDCU的RISC-V指令集扩展,集成到Klessydra T03核心中。 3. **HDCU特性**:HDCU具有自定义指令集、专用功能单元、本地内存和可配置性,可加速HDC的主要操作。 4. **硬件资源使用**:在FPGA上实现HDCU,分析了硬件资源使用随SIMD宽度增加的扩展性。 5. **性能提升**:HDCU在绑定、捆绑、排列、裁剪、相似性和关联搜索等操作上实现了2.70倍到3844.09倍的加速。 6. **结论**:HDCU是一个可配置的通用协处理器,通过扩展RISC-V指令集,为程序员提供了易于使用的加速器,显著提升了HDC任务的性能。
"RISC-V加速HDC,效率翻倍?" "HDC学习任务,RISC-V如何助力?" "RISC-V新扩展,AI加速新篇章?"
客服
商务合作
小程序
服务号
折叠