资源描述
,AI-Optimized Chipsets,Part III: Key Opportunities & TrendsAug 2018,An Introduction,Previously in Part I, we reviewed the ADAC loop and key factors driving innovation for AI- optimized chipsets.In Part II, we review the shift in performance focus computing from general application neural nets and how this is driving demand for high performance computing. To this end, some startups are adopting alternative, novel approaches and this is expected to pave the way for other AI-optimized chipsets.In this instalment, we review the training and inference chipset markets, assess the dominance of tech giants, as well as the startups adopting cloud-first or edge-first approaches to AI-optimized chipsets.,The training chipset market is dominated by 5 firms, while the inferencechipset market is more diverse with 40 players,Source: UBSThe training chipset market is dominated by 5 firms which have developed massively parallel architectures, well-suited for deep learning algorithms. NVIDIA is the most prominent with its GPU technology stack.The early leaders in this market are likely to maintain their lead, but the inference market is large and easily accessible to all, including tech giants and startups. Given the significant market for inference chipsets, many defensible market niches based on high-system speed, low power and/or low Total Cost of Ownership (TCO) products are likely to emerge.At least 20 startups and tech giants appear to be developing products for the inference market. According to UBS, within todays datacenter (DC) market, the CPU in particular, appears to be the de facto choice for most inference workloads.,Source: UBS,Source: The Verge | Tech Crunch | Nvidia | UBS,Almost all training happens in the cloud/DC, while inference is conducted in the cloud and at the edge,Self Driving Cars,Drones,Training,CloudData Center (DC)Cloud | Data Centre,SurveillanceIntelligent CamerasMachinesEdge,InferenceThe market for AI-optimized chipsets can be broadly segmented into cloud/DC and edge (including automotive, mobile devices, smart cameras and drones).The edge computing market is expected to represent more than 75% of the total market opportunity, with the balance being in cloud/DC environments.The training chipset market is smaller than that of inference because almost all training is done in DCs. The market for inferencechipsets is expected to be significantly larger due to the impact of edge applications for deep learning.With the exception of Google that created the Tensor Processing Unit (TPU) specifically for deep learning. Other training chipset players entered the deep learning market by repurposing existing architectures that were originally used for other businesses (e.g. graphics rendering for high-end PC video games and supercomputing).While a handful of startups plan to commercialize new deep learning training platforms, the vast majority of new players seem focused on inference for large markets (e.g. computer vision). The multitude and diversity of deep learning applications have led many firms to pursue niche strategies in the inference market.All training chipset players have developed high performance inference platforms to support sizable end markets in DCs, automotive, and supercomputing. Dozens of startups, many of which are based in China, are developing new designs to capture specific niche markets.,Creating an interesting landscape of tech giants and startups staking theirbets across chipset types, in the cloud and at the edge,Cloud | DC (training/inference)Edge (predominantly inference),Key ObservationsAt least 45 startups are working on chipsets purpose-built for AI tasksAt least 5 of them have raised more than USD 100M from investorsAccording to CB Insights, VCs invested more than USD 1.5B in chipset startups in 2017, nearly doubling the investments made 2 years ago,Most startups seem to be focusing on ASIC chipsets at the edge and in the cloud/DC,FPGAs and other architectures also appear attractive to chipset startupsStart-ups,Note: Several startups are in stealth mode and company information may not be publicly available.Source: CrunchBase | IT Juzi | CB Insights | What the drivers of AI industry, China vs. US by ZhenFund | Back to the Edge: AI Will Force Distributed Intelligence Everywhere by Azeem | icons8 | UBS,Source: Nvidia | UBS,Most training chipsets are deployed in DCs. The training market isdominated by Nvidias GPU,NVIDIA was first to market, by a wide margin, with semiconductors and products specifically designed for deep learning. It has forward integrated all the way into the developer tools stage of the value chain, with its NVIDIA GPU Cloud offering (which is a container registry, rather than competitor to AWS or Azure), stopping just shy of creating deep learning applications that would compete with its own customersMost training chipsets are deployed in DCs. The training market is dominated by Nvidias GPU which has massively parallel architectures, very strong first-to-market positioning, ecosystem and platform integrationGPUs have hundreds of specialised “cores”, all working in parallelNvidia delivers GPU acceleration for both training (Nvidia DGX SYSTEMS for Data Center, Nvidia GPU Cloud) and inference (NvidiaTitan V for PC, Nvidia Drive PX2 for Self-Driving Cars, NVIDIA Jetson for intelligent machines)In the DC market, training and inference are performed on the same semiconductor device. Most DCs use a combination of GPU andCPU for deep learningThe CPU is not well suited for training, but much better suited for inference where code execution is more serial than parallel and low- precision & fixed point logic are still popular,Titan V,Source: Market Realist | Intel and Xilinx SEC Filings | The Rise of Artificial Intelligence is Creating New Variety in the Chip Market, and Trouble for Intel by The Economist | Deloitte,Microsoft and Amazon both use FPGA instances for large scale inferencing- implying that FPGAs may have a chance in the DC inference market,By end of 2018, Deloitte estimates that FPGAs and ASICs will account for25% of all chips used to accelerate machine learning in the DC,At the same time, some tech giants have also started building proprietarychipsets in the cloud/DC,Alibaba, Amazon, Baidu, Facebook, Google, Microsoft and Tencent (i.e. the “Super 7 Hyperscalers”) exhibit a high degree of vertical integration in the deep learning value chain, capable of building/designing proprietary accelerator(s) by virtue of self-administered hyperscale DC expertiseIf data is the new oil, these firms are emerging as the vertically integrated oil companies of tomorrows data driven global economy. Semiconductor firms must develop strategies to maintain their market power and remain relevantThe Super 7 hyperscalers, as well as IBM and Apple, have been very active in each stage of the value chain, including algorithmic research, semiconductors, hardware and end-user applications,Source: UBS,Source: Google Rattles with Tech World with a New AI Chip for All by Wired; Googles Next-Generation AI training system is monstrously fast by the Verge; Googles Second AI Chip Crashes Nvidias Party by Forbes; Microsoft Announces Project Brainwave To Take On Googles AI Hardware Lead by Forbes; Building an AI Chip Saved Google from Building a Dozen New Data Centers | Google,Googles ASIC chipset the cloud TPU is a case in point,Instead of selling the TPU directly, Google offers access via its new cloud serviceGoogles original TPU had helped power DeepMinds AlphaGo victory over Lee SedolIt is making cloud TPUs available for use in Google Cloud PlatformThe TPU is designed to be flexible enough to for many different kinds of neural network models,Cloud TPU (2017),TPUv1 (2015),TPUv3 Pod (2018),Inference only,Training and inferenceTPUv2 Pod (2017),As more compute is pushed to the edge, tech giants are increasingly developing or acquiring solutions in this space,As these trained models shift to real-world implementation, more compute will be pushed to the edgeOther edge applications that will contain specialized deep learning silicon include drones, kiosks, smart surveillance cameras, digital signage and PCs,Tech Companies,Chipset Makers,Source: UBS,Where Chipset Makers are Investing in Private Markets (2015-2017)Intel Capital is the most active corporate entity. Close behind is Qualcomm VenturesOver 79 deals went to IoT companies which is to be expected, given their relevance to embedded computing with small chipsets15 of these deals went to drone-specific companiesAR/VR companies also attracted 25 deals50 deals went to AI companies, many of which were cutting edge computer vision and auto tech playsChipset makers also invested in horizontal AI platforms as well as visual recognition API maker,Driving a deal-making galore in the chipset industry,Source: Where Major Chip Companies Are Investing In AI, AR/VR, And IoT by CB Insights,At the same time, there is a proliferation of new entrants, mostly ASIC- based startups including Graphcore, Wave Computing,SolutionHighlightsFinancing,Source: What Sort of Silicon Brain Do You Need for Artificial Intelligence? by The Register | Venture Beat | Suffering Ceepie-Geepies! Do We Need a New Processor Architecture? By The Register,Source: PitchBook | CB Insights | Forbes,KnuEdge, Gyrfalcon Technology,Solution,Highlights,Financing,Source: PitchBook | CB Insights | Forbes | Cerebras | Groq | Tenstorrent,As well as Cerebras, Groq, Tenstorrent - startups with potential offerings in the cloud/DC,SolutionHighlightsFinancing,Source: 12 AI Hardware Startups Building New AI Chips by Nanalyse | AI Chip Boom: This Stealthy AI Hardware Startup Is Worth Almost A Billion by Forbes | Chip Startups were almost toxic before the AI boom now investors are plowing money in to them | Pitchbook | Startup Unveils Graph Processor at Hot Chips by EETimes,At the edge, there are also ASIC offerings by Horizon Robotics, Mythic,SolutionHighlightsFinancing,Source: PitchBook | CB Insights,Cambricon, Kneron,Solution,Highlights,Financing,Source: 12 AI Hardware Startups Building New AI Chips by Nanalyse | AI Chip Boom: This Stealthy AI Hardware Startup Is Worth Almost A Billion by Forbes | Chip Startups were almost toxic before the AI boom now investors are plowing money in to them | Pitchbook | Startup Unveils Graph Processor at Hot Chips by EETimes | Hailo,Thinci, Hailo.,Solution,Highlights,Financing,Source: Syntiant | PitchBook,As well as Syntiant,Located: US Founded: 2017 Stage: Series A ASIC,Solution,Highlights,Financing,Source: PitchBook | CB Insights | CrunchBase | Efinix | Achronix,Efinix and Achronix - FPGA startups offering cloud/DC and edge solutions,SolutionHighlightsFinancing,While ThinkForce, Adapteva offer alternative chipset architectures fromCPUs, GPUs, FPGAs and ASICs in their solution(s),SolutionHighlightsFinancing,Source: PitchBook | CB Insights | CrunchBase | ThinkForce | Adapteva,Notwithstanding this positive outlook, challenges abound for many ofthese startups,Source: AI Chip Boom: This Stealthy AI Hardware Startup Is Worth Almost A Billion by Forbes | The Race to Power AIs Silicon Brains by MIT Technology Review | Lux Capital,Conclusion,In Part I, we note that deep learning technology has primarily been a software play to date. The rise of new applications (e.g. autonomous driving) is expected to create substantial demand for computing.Existing processors were not originally designed for these new applications, hence the need to develop AI-optimized chipsets. We review the ADAC loop and key factors driving innovation for AI-optimized chipsets.In Part II, we explore how AI-led computing demands are powering these trends. To this end, some startups are adopting alternative, novel approaches and this is expected to pave the way for other AI-optimized chipsets.This is the end of Part III, where we took a high-level look at training and inference chipset markets, the activities of tech giants in this space, coupled with disruptive startups adopting cloud-first or edge-first approaches to AI-optimized chips.Evidently, this is a super-saturated industry with many players. It is unclear where the exits will be. Tech giants are either moving forward with their roadmaps (i.e. Nvidia), building their own chipsets (e.g. Alibaba, Amazon, Apple, FaceBook etc) or have already made acquisitions (e.g. Intel).In Part IV, we look at other emerging technologies including neuromorphic chips and quantum computing systems, to explore their promise as alternative AI-optimized chipsets.Finally, we are most grateful to Omar Dessouky (Equity Research Analyst, UBS) for his many insights on the semiconductor industry.Do let us know if you would like to subscribe to future Vertex Perspectives.,Thanks for reading!,DisclaimerThis presentation has been compiled for informational purposes only. It does not constitute a recommendation to any party. The presentation relies on data and insights from a wide range of sources including public and private companies, market research firms, government agencies and industry professionals. We cite specific sources where information is public. The presentation is also informed by non-public information and insights.Information provided by third parties may not have been independently verified. Vertex Holdings believes such information to be reliable and adequately comprehensive but does not represent that such information is in all respects accurate or complete. Vertex Holdings shall not be held liable for any information provided.Any information or opinions provided in this report are as of the date of the report and Vertex Holdings is under no obligation to update the information or communicate that any updates have been made.,About Vertex VenturesVertex Ventures is a global network of operator-investors who manage portfolios in the U.S., China, Israel, India and Southeast Asia.Vertex teams combine firsthand experience in transformational technologies; on-the-ground knowledge in the worlds majorinnovation centers; and global context, connections and customers.,About the Authors,Emanuel TIMOR General Partner Vertex Ventures Israelemanuelvertexventures,Sandeep BHADRAPartnerVertex Ventures US sandeepvertexventures,Brian TOH Executive Director Vertex Holdingsbtohvertexholdings,Tracy JINDirectorVertex Holdings tjinvertexholdings,XIA Zhi JinPartnerVertex Ventures China xiazjvertexventures,24,Thanks,
展开阅读全文