NEWS: 公告在東京證券交易所JASDAQ標準市場新上市

表紙
市場調查報告書

AI 軟件優化解決方案(2021):Kisaco 領導力圖

Kisaco Leadership Chart on AI Software Optimization Solutions 2021

出版商 Kisaco Research Limited 商品編碼 1007521
出版日期 內容資訊 英文 49 Pages
商品交期: 最快1-2個工作天內
價格
AI 軟件優化解決方案(2021):Kisaco 領導力圖 Kisaco Leadership Chart on AI Software Optimization Solutions 2021
出版日期: 2021年02月24日內容資訊: 英文 49 Pages
簡介

我們今天見證的人工智能 (AI) 的革命或戲劇性演變始於可以移植深度學習神經網絡的硬件加速器的出現。過去需要數月的學習時間現在可以在 Nvidia GPU 上減少到數天或數小時。今天,人工智能市場正在引入新的優化形式,除了加速之外還有許多功能,在純軟件基礎的軟件級別運行,機器學習 (ML) 技術堆棧。許多 AI 軟件優化 (AISO) 產品來自相對較新的初創企業。

本報告在 Kisaco 領導力圖表 (KLC) 上比較了 AISO 市場中的領先公司。介紹AISO的特點,大公司分析,各廠商簡介。

目錄

  • Kisaco Research 的觀點
  • 動機
  • 主要調查結果
  • AI 軟件優化 (AISO):市場和技術形勢
  • AISO的定義
  • AISO 產品適用於 ML 堆棧的多層
  • 硬件層
  • 運行層
  • 模型推斷
  • 模型開發和培訓
  • AutoML/垂直特定應用
  • 混合優化
  • AISO 解決方案市場是一個新興市場
  • 市場觀點
  • ML 生命週期管理的作用
  • 解決方案分析:供應商比較
  • Kisaco 人工智能軟件優化解決方案 (KLC) 領導力圖表
  • 人工智能軟件優化解決方案廠商對比
  • KLC 用於 AI 軟件優化解決方案
  • 供應商分析
  • 應用大腦研究,Kisaco 評級:領導者
  • Kisaco 的評估
  • 代碼播放
  • Deci AI、Kisaco 評級:領導者
  • Kisaco 的評估
  • 深度人工智能
  • Deeplite、Kisaco 評級:領導者
  • Kisaco 的評估
  • Eta Compute、Kisaco 評級:初創企業
  • Kisaco 的評估
  • Mipsology,Kisaco 評級:競爭者
  • Kisaco 的評估
  • OctoML、Kisaco 評級:競爭者
  • Kisaco 的評估
  • SigOpt,英特爾公司,Kisaco 評級:領導者
  • Kisaco 的評估
  • 附錄
  • 供應商解決方案選擇
  • 選擇標準
  • 排除標準
  • 調查方法
  • KLC 的定義
  • Kisaco 研究評估
  • 參考資料
  • 謝謝
  • 作者
  • Kisaco Research 的分析網絡
  • 版權聲明/免責聲明
目錄

Motivation

The current revolution or dramatic evolution in artificial intelligence (AI) we are witnessing was sparked by the arrival of hardware accelerators onto which deep learning neural networks were ported: training times that took months ran in days or hours on Nvidia GPUs. This gave rise to the explosion in AI hardware accelerator chips competing to take a share of the large and still growing accelerator market. Now a new form of optimization, that encompasses a host of features beyond and inclusive of acceleration, has appeared in the AI market, purely software based: meaning that they operate at the software level in the machine learning (ML) technology stack. Many of the AI software optimization (AISO) products have emerged from relatively recent startups. These products can optimize ML models that run on just central processing units (CPUs) or enhance performance on standard AI accelerators: graphics processing units (GPUs), field-programmable gate arrays (FPGAs), and digital signal processors (DSPs). AISO products also compete with the newer breed of AI chips (which we label as domain specific architectures, DSAs), making the whole AI field a lot more nuanced and competitive. For both enterprise users and product manufacturers there are now wider options in choosing the best combination of software and hardware for their AI applications and products requirements.

In this report we feature the leading players in the AISO market, compared side by side in our Kisaco Leadership Chart (KLC). We explain what this technology has to offer, reveal our analysis of the top players, and profile each of these vendors.

What you will learn:

  • We define and explain the new AI software optimization solution market.
  • We explain where in the ML technology stack AISO products operate.
  • We delve into the different types of optimization techniques, from standard operations to the advanced techniques available from the AISO vendors.
  • The relevance of ML lifecycle management and how AISO vendors offer lifecycle functionality.
  • We compare seven leading players in the AISO space side by side in our KLC assessment.
  • We provide heatmaps to visually see what key features the KLC vendor products have to offer.
  • We provide a profile on each of the KLC participating vendors together with three strengths and three weaknesses.

Table of Contents

  • Kisaco Research View
  • Motivation
  • Key findings
  • AI software optimization: market and technology landscape
  • Defining AISO
  • AISO products operate across multiple layers of the ML stack
  • Hardware layer
  • Runtime layer
  • Model inference
  • Model development and training
  • AutoML and vertical specific applications
  • Hybrid optimization
  • The AISO solution market is an emerging one
  • Market view
  • The role for ML lifecycle management
  • Solution analysis: vendor comparisons
  • Kisaco Leadership Chart on AI software optimization solutions 2021
  • AI software optimization solution vendor comparisons
  • The KLC chart for AI software optimization solutions
  • Vendor analysis
  • Applied Brain Research, Kisaco evaluation: Leader
  • Kisaco Assessment
  • Codeplay
  • Deci AI, Kisaco evaluation: Leader
  • Kisaco Assessment
  • Deep-AI
  • Deeplite, Kisaco evaluation: Leader
  • Kisaco Assessment
  • Eta Compute, Kisaco evaluation: Emerging Player
  • Kisaco Assessment
  • Mipsology, Kisaco evaluation: Contender
  • Kisaco Assessment
  • OctoML, Kisaco evaluation: Contender
  • Kisaco Assessment
  • SigOpt, an Intel Company, Kisaco evaluation: Leader
  • Kisaco Assessment
  • Appendix
  • Vendor solution selection
  • Inclusion criteria
  • Exclusion criteria
  • Methodology
  • Definition of the KLC
  • Kisaco Research ratings
  • Further reading
  • Acknowledgements
  • Author
  • Kisaco Research Analysis Network
  • Copyright notice and disclaimer

Figures

  • Figure 1: The ML technology stack and the optimizations possible at each level.
  • Figure 2: Comparing different bit precision formats.
  • Figure 3: Where Sycl and OpenCL sit in the technology stack.
  • Figure 4: Comparing number of supported Nvidia GPU operations by different interfaces.
  • Figure 5: Apache TVM: working at intermediate representation (IR) level (TVM in grey).
  • Figure 6: ML lifecycle in the ML tool eco-system (covering DataOps, AutoML, MLOps).
  • Figure 7: AISO features available in the participating vendor products.
  • Figure 8: Heat map analysis of participating vendor solution technical features.
  • Figure 9: Kisaco Leadership Chart on AI software optimization solutions 2021.
  • Figure 10: Kisaco Leadership Chart on AI software optimization 2021: ranking of vendors
  • Figure 11: ABR's temporal dithering.
  • Figure 12: ABR's LMU characteristics versus the well-known LSTM neural network model.
  • Figure 13: ABR LMU models for keyword spotting. Note x-axis goes from large to small models left to right.
  • Figure 14: An AI horizontal AI hardware organization.
  • Figure 15: Working on top of open industry standards.
  • Figure 16: The Deci Platform.
  • Figure 17: The Deci Platform: dashboard.
  • Figure 18: The Deci Platform Insights screen.
  • Figure 19: Deep-AI training and inference flow.
  • Figure 20: Deep-AI software stack.
  • Figure 21: Deeplite Edge AI solution.
  • Figure 22: Deeplite automated model optimization.
  • Figure 23: Deeplite Neutrino profiler: results readout.
  • Figure 24: Eta Compute Tensai platform.
  • Figure 25: Mipsology Zebra software stack.
  • Figure 26: Mipsology Zebra performance metric efficiency: comparisons.
  • Figure 27: Where TVM (in blue) sits in the ML technology stack.
  • Figure 28: AutoTVM overview.
  • Figure 29: Where SigOpt sits in the ML technology stack
  • Figure 30: SigOpt features for management of the training process.