This market research report was originally published at Tractica's website. It is reprinted here with the permission of Tractica.
Intel held its Analyst Event for European analysts during June. Some of Tractica’s observations are discussed below.
First impressions and highlights of the event:
The recent Analyst Event in London provided a good overview of Intel’s overall strategy. It also provided analysts with ample opportunity to have breakout discussions with the company’s key people. Intel had multiple sessions across 2 days focused on different areas, including memory, interconnect, data centers, edge computing, quantum computing, AI, and software. Tractica’s focus was on better understanding the company’s AI strategy and products.
Intel’s strategy was laid out by Raja Koduri, chief architect and senior vice president of Intel Architecture, Software and Graphics. In his presentation, which was entitled “Think Exponential,” Koduri made the case for how Intel is now spread across six key pillars: packaging, software, security, interconnect, process, and architecture. Intel’s executives believe the company has the components to tackle industry challenges related to moving from the data-centric era to the intelligence era, which is defined by AI. Therefore, AI is a key theme for Intel and is central to how the company sees the value of hardware compute shifting over the next decade or so. With the ability to pull multiple levers, from process to memory and packaging, Intel is confident it has what it takes to win in this new intelligence-driven, exponential paradigm.
One big highlight is the fact that Intel will have AI acceleration in every product from here on. Its Scalar Vector Matrix Spatial (SVMS) strategy targets scalar (CPU), vector (GPU), matrix (ASIC), and spatial (FPGA) computing platforms. Intel has the Golden Cove CPU architecture coming up in 2021, and it will have AI performance improvements as well as 5G support capabilities. Also, Cascade Lake and Ice Lake CPUs now come with deep learning Boost acceleration in the software, which has provided as much as a 28x increase in performance. Altera (FPGA) and Movidius (ASIC) already provide acceleration, and both target inference workloads. Nervana will add to the ASIC portfolio later in 2019, targeting both inference and training in separate chips. In 2021, Intel is also launching a GPU chip that will be on the new 7 nm process node and will target AI and high performance computing (HPC) workloads.
The second highlight is related to Intel’s OneAPI software strategy, which allows developers to develop applications across the SVMS platforms covering CPUs, GPUs, FPGAs, and ASICs. This is a powerful strategy and a big differentiator in a market that is currently dominated by NVIDIA’s GPU-focused CUDA. The ability for AI developers to target not just GPUs but many other types of compute architectures opens up the market and provides much needed competition against NVIDIA.
Interestingly, the Movidius team at Intel was initially skeptical of OneAPI because it meant having to adopt a new software base for the company’s chips. However, the team has since become a champion for the strategy. It allows Intel’s engineers to not worry about individual framework support, as OneAPI can handle all the major frameworks. But OneAPI doesn’t get launched commercially until later in 2019.
Intel is seeing demand within the AI edge hardware space, as it is focused on AI edge opportunities through its Movidius team. The company is seeing these opportunities emerge across multiple industries, including smart cities, drones, manufacturing, retail, and financial services, among others. However, as Intel characterized it, the market is still in the early stages, with some sectors like video surveillance being much further ahead than others. Most other markets are seeing pockets of activity while video surveillance is already mature and accelerating ahead. This makes sense, as Movidius is largely focused on vision applications – and vision is by far one of the largest markets in terms of AI activity. When asked about voice recognition support, the Movidius team was confident of supporting that workload, but so far, it hasn’t gotten any traction in that area.
Another observation is around the application of AI with edge servers, which Intel said was mostly a video surveillance phenomenon through network video recorders (NVRs). This assessment validates Tractica’s own view of the AI edge market. Also, Intel made it clear that while it has exited the mobile handset market, it is still very much active in 5G wireless infrastructure like base stations, core networks, etc. However, there wasn’t any specific mention of 5G and AI edge and what new opportunities they might bring for Movidius or Intel.
What should customers expect from Intel over the next year?
Intel is preparing for the long haul rather than the short haul. Thus, changes in its strategy are unlikely to bear fruit in the next 12 months. The one area that will start to have an impact on the market over the next year will be OneAPI. Its success will depend on whether it can offer credible competition to NVIDIA and its CUDA developer ecosystem. Most CUDA developers are interested in the high end performance capabilities of GPUs. Until we see how Nervana’s NNP ASICs (available 2H 2019) stack up against NVIDIA’s GPUs, it will be hard to estimate how much of a dent OneAPI can make in NVIDIA’s domination.
Unfortunately, customers need to wait another couple of years for Intel’s GPU product. Beginning in 2020, customers who are on CUDA will be able to compare the performance of their models to Intel’s FPGA, ASIC, and CPU chips. The ability to test inference workloads – especially across all the various offerings – is a powerful one. This ability is something that Intel’s customers would value and see as a major differentiator from NVIDIA or many of the other AI chipset startups that have entered the space.
Research Director, Tractica