Powering Artificial Intelligence
23 October 2019
Bottom line:
There is no single solution to power AI, as different chipsets respond to different requirements, like intensive computation, power consumption, and low latency.
While the cloud infrastructure has already been developing for some time, AI at the edge is still in its infancy.
The Edge represents the sweet spot for AI, as it is where AI de facto will happen and will get into our lives, driven by IoT and other smart devices. It is also the segment that is experiencing the fastest pace of growth, and where most of the investable ideas are to be found.
We play the AI at the edge theme through our exposure to Lattice Semiconductor (LSCC:US).
AI: different tasks, different requirements
- AI has become a buzz word being used in any context, ranging from Agricultural applications, to finance, social media or governmental institutions
- AI often appears as a black box that “somehow” enhances human capabilities, optimizes processes, or accelerates decision making.
- Inside that black box, though, there are recognizable elements that address specific tasks.
- Like children, machines need to learn (training). They create their learning model with external help (supervised learning) or without it (unsupervised learning).
- Once the learning process is completed, the machine is ready to prove what it is capable of (inference).
- Training and inference are two distinct steps and may occur in different physical places: either on the cloud or on the device itself (edge computing) but also in different kinds of processors (chipsets).
- This can happen for different reasons: privacy, latency, need for computing power, etc.
- Training requires a lot of data and a lot of computing power.
- Inference is driven by latency, which means that given an input (for instance, a pedestrian crossing the road), the machine needs to make a decision (stopping the car) in the shortest possible time.
.png)
Deep learning
.png)
No chip to rule them all
- High throughput and low latency carry different requirements that derive from the different purposes they have been designed for: there is no one-solutionfits-all.
- As computing needs present various challenges, different types of chipsets have been developed to provide appropriate solutions.
- Each one is presenting pros and cons, as listed in the attached table.
.png)
.png)
Two different playgrounds, and one is more appealing
- The slowdown in silicon performance (end of Moore’s law) has resulted in the development of an ecosystem of different architectures and chipsets on the cloud and at the edge. These two markets are growing at different paces.
- The cloud ecosystem is dominated by large firms: Intel (INTC:US), AMD (AMD:US), NVIDIA (NVDA:US), and FPGA leaders, such as Xilinx (XLNX:US). In 2018 chips sales accounted for $4.2bn.
- The edge landscape is still very fragmented and accounted for $1.9bn last year, but is growing fast (expected CAGR of 31% over 2018-2025). Players in this field are Lattice Semiconductor (LSCC:US), Ambarella (AMBA:US), ARM (owned by Softbank, 9984:JP), Thinci (not listed) and Graphcore (not listed) each one developing chips for specific applications or markets.
- Among these sectors, those where AI edge chipset are expected to dominate are:
- Advanced driver-assistance systems (ADAS) – have already started to deploy dedicated edge chipsets, used in cars, as they become increasingly autonomous.
- Smart appliances, smart homes, and smart cities – all using embedded AI chipsets to satisfy the requirements of low latency and low power.
- Robotics and other industrial devices – represent a high volume market for edge chipsets that will deploy on-premises the models trained on the cloud.
.png)
Catalysts:
- IoT (edge inference). Expected to be the largest field of use for low power FPGAs and ASIC, as they offer a complementary solution to common problems such as low power requirements, latency, and in situ operation for privacy issues.
- 5G infrastructure. Crucial in the growth of AI chips, especially for FPGA, which is already widely used, and in the medium to long future for the ASIC.
- Datacenters. A significant driver for all kinds of chipsets, as AI will be deployed across every type of IT infrastructure.
- Autonomous Vehicles. Inherently stringent safety requirements will trigger the adoption of one chip technology (likely ASIC) in the long term.
Risks:
- High development costs. Likely to hinder the take-off of certain types of chipsets, notably FPGA and ASIC.
- The software stack remains obscure. Lack of understanding is delaying the adoption of certain chipsets, like FPGA, despite efforts to create an AI framework by the leading producers (Vitis for Xilinx and SensAI for Lattice Semiconductor),
- Oligopolistic market. Concentration in the hands of a few big tech companies may hinder the innovation and development of new players.
Sources:
https://erpinnews.com/fog-computing-vs-edge-computing
https://blogs.nvidia.com/blog/2016/08/22/difference-deep-learning-training-inference-ai/
https://towardsdatascience.com/benchmarking-hardware-for-cnn-inference-in-2018-1d58268de12a
https://m.eet.com/media/1313554/InferenceandTrainingRevenueChart.JPG
Explore:
Disclaimer
This report has been produced by the organizational unit responsible for investment research (Research unit) of atonra Partners and sent to you by the company sales representatives.
As an internationally active company, atonra Partners SA may be subject to a number of provisions in drawing up and distributing its investment research documents. These regulations include the Directives on the Independence of Financial Research issued by the Swiss Bankers Association. Although atonra Partners SA believes that the information provided in this document is based on reliable sources, it cannot assume responsibility for the quality, correctness, timeliness or completeness of the information contained in this report.
The information contained in these publications is exclusively intended for a client base consisting of professionals or qualified investors. It is sent to you by way of information and cannot be divulged to a third party without the prior consent of atonra Partners. While all reasonable effort has been made to ensure that the information contained is not untrue or misleading at the time of publication, no representation is made as to its accuracy or completeness and it should not be relied upon as such.
Past performance is not indicative or a guarantee of future results. Investment losses may occur, and investors could lose some or all of their investment. Any indices cited herein are provided only as examples of general market performance and no index is directly comparable to the past or future performance of the Certificate.
It should not be assumed that the Certificate will invest in any specific securities that comprise any index, nor should it be understood to mean that there is a correlation between the Certificate’s returns and any index returns.
Any material provided to you is intended only for discussion purposes and is not intended as an offer or solicitation with respect to the purchase or sale of any security and should not be relied upon by you in evaluating the merits of investing inany securities.
Back to research
