Intel and Alibaba, China’s equivalent of Amazon, have gone ahead with a partnership over AI developments despite the friction imposed by the ongoing trade standoff between the US and China. However, this comes as little surprise to followers of both companies, which have been pursuing global strategies knowing that the potential long-term gains outweigh risks and short-term challenges.
It is true that the trade tensions have led Alibaba to step up investment in its own semiconductor business as a hedge against tariffs and constraints, but for now it recognizes the value of Intel hardware for its cloud strategy. Intel meanwhile sees Alibaba with its huge presence in Chinese retail and eCommerce as a major outlet for its chips and technology. In anticipation of this, Intel participated with Alibaba on China’s Singles Day on November 11th, the country’s answer to Black Friday, when the company racked up 213.5 billion yuan, or $30.8bn, in gross merchandise sales.
The partnership announced at CES 2019 involved integrations between Intel hardware and Alibaba’s cloud infrastructure, with Singles Day providing a point of peak demand to prove the concept. For that, Alibaba deployed Intel Xeon Scalable processors along with the associated Optane DC persistent memory in its production environment.
These slotted into a dedicated architecture developed jointly by the two companies called Joint Edge Computing Platform, launched in September 2018. This incorporates various AI accelerators from Intel, including CPUs, field programmable gate arrays (FPGAs) and Movidius vision processing units (VPUs). VPUs are distinct from video processing units dedicated to encoding and also GPUs, being designed to run algorithms such as convolutional neural networks (CNNs) oriented towards machine vision.
The VPUs here typically incorporate direct interfaces to take data from cameras, bypassing off chip buffers, emphasizing on-chip dataflow between multiple parallel execution units. Intel has already been very successful with VPUs in China, having announced in July 2018 that Baidu, China’s Google, was using its Movidius VPUs for its Xeye camera.
A key ingredient for its Alibaba AI partnership is the Intel Optane DC persistent memory, set to become generally available in the first half of 2019, and formerly called 3D Xpoint. This comes with the Xeon Scalable processors with the aim being to make large amounts of memory available close to the chip for rapid access, which Alibaba needs to perform real time analytics at times of huge workload as on Singles Day. It’s similar in concept to SAP’s HANA in-memory databases – you want everything in the very fast RAM, rather than slower storage.
Persistent memory is a relatively recent addition to the storage hierarchy that starts with levels of on-chip cache at the top, through DRAM and onto solid state drives (SSD) and then hard disk drives (HDD). Each successive layer is slower and more expensive per bit but with higher capacity, while only SSD and HDD at the bottom are persistent in the sense that data is retained in the event of power being switched off.
Non-volatile memory emerged as a half-way house by retaining its bits without power but lacking the persistent metadata such as page table entries needed to ensure information could be reconstructed. Persistent memory built on this to incorporate the basis file structure needed to retain data robustly and as such fits in the storage hierarchy between DRAM and the disk SSD or HDD. It gives some of the capacity and lower cost per bit of SSD combined with the higher performance of DRAM, or even on-chip cache.
It is not suitable for all applications and will not replace disk, but will figure for the many use cases involving analytics and monitoring where large amounts of data need to be parsed in real time. Applications where it has already been used include fraud detection, cybersecurity monitoring and financial trading, as well as machine vision. Any process involving massively parallel processing and distribution of execution to multiple devices will benefit from persistent memory.
Intel’s Optane DC persistent memory has two operating modes, App Direct mode and Memory mode. Applications that have been specifically tuned can take advantage of App Direct mode to exploit the full value of the product’s native persistence and larger capacity. In Memory mode, applications running in a supported operating system or virtual environment can use the product like DRAM, taking advantage of the additional system capacity made possible from module sizes of up to 512 GB without needing to rewrite software.
With Alibaba, Intel is developing the first major application of their combined AI platform to enable 3D tracking of athletes’ movement and posture ready for the Tokyo 2020 Olympic Games. In fact, it should be ready before the games, because the aim is to allow coaches to extract biomechanical data to help optimize training and iron out inefficient movements that impede performance. A key design goal is avoiding need for special sensors and to enable standard cameras to be used, located in multiple positions around the event to capture 3D images.
At the same time the technology can be used to create more immersive viewing for the audience through the ability to zoom in on individual athletes and highlight their posture. The role of AI will be to extract meaningful geomechanical data from the images of use both for coaches during training and for broadcasters to analyze, dissect and re-examine highlights during instant replays.
Meanwhile, Intel has also been forging ahead with other initiatives in China, especially around mobile, sufficient to justify its own annual 5G Summit there held in September 2018.