Intel Myriad

In 2017-2018 Machine learning comes out research laboratories in practice and makes the way in the user devices as now speak, "on border". At the suggestion of ARM company of such solution call edge machine learning (edge is translated as an edge, an edge, border).

The technologies of the Californian company Movidius purchased by Intel opened for her the road on the market of deep machine learning and computer vision, and the Movidius Myriad X processor released in 2017 allowed corporation to enter an edge machine learning segment.

Movidius by the time of acquisition from Intel made more than ten years intellectual property (IP) in the form of documentation for production of processors ISAAC, Myriad 1, Myriad 2. Its last achievement before transition to the structure of Intel was connected with the device in Fathom USB flash card format with the Myriad 2 processor capable to work together with the ARM processor running Linux.

Fathom costs only $79 and is used in drones, robots, IoT devices and video surveillances.

Before acquaintance to Movidius it should be noted that intentions of this company were in details described in the document filed in 2016 in the request for receiving a grant from European Commission on the Horizon 2020 PROGRAMME project.

Chronology of developments of Movidius

2018

Release of Windows Machine Learning for Intel Movidius Myriad X

On March 7, 2018 on Windows Developer Day the Microsoft corporation announced the new version of Windows 10 - Windows Machine Learning (ML). It as well as possible corresponds to the Intel Movidius Myriad X VPU system integrated on the chip. They are integrated by the concept "output" or inference - so machine learning specialists call practical use of ready neural networks as the tool at the level of edge machine learning. About a ratio of processes of a training of network and an output see here.

Windows ML ensures functioning of previously trained networks on any devices supported by Windows 10 and, thus, allows the developer to apply the trained network in some specific application. Till 2017-2018 for an output local or cloud CPU or GPU were, as a rule, used that extremely inefficiently and significantly limits a scope, for example, in those self-contained mobile devices where work with the image is required.

Without tough binding to machine vision the similar problem of autonomism of processes of an output is solved based on the ARM processor in the Trillium project, and Intel offers qualitatively new processor for work only with images Intel Vision Processing Units (VPU) with a proper name of Movidius Myriad X.

This specialized chip serves for acceleration of work of neural networks at an output stage therefore marketing refers it to AI though the intellectuality is still very conditional. This chip reasonable to recognize difficult, it is just smart.

2017

Myriad X release

If Myriad 2 VPU which is released in 2012 (the cm is more detailed below), was initially conceived for problems of machine vision, and already then is successfully adapted for machine learning, then the Myriad X processor is initially adapted directly to this application class. Therefore in Myriad X to 16 the number of SHAVE processors is increased and there was absolutely new Neural Compute Engine accelerator that, according to the representative of the company, brought to 10 to multiple performance improvement on transactions with a floating point. This statement sounds a little strange as for acceleration of processes of an output on neural networks work speed on integer operations and transactions with the fixed point, and with the lowered accuracy is more critical. Most likely the Neural Compute Engine accelerator which device is not disclosed is focused on these transactions. Only the fact that Neural Compute Engine is a hardware accelerator for deep neural networks in which the intelligent interface of memory will be exempted from the well-known bottle throat arising at data exchange between memory and the processor is known. Neural Compute Engine differs in low energy consumption.

One more accelerator as a part of Myriad X VPU getting to category of specialized iron (fixed-function hardware) is intended for work with stereoscopic effects and depth of the image.

Architecture of Myriad X VPU

Emergence of Myriad X VPU devices is interesting that they open qualitatively new opportunities. It is known that about 90% of information of people obtain using sight, about 9% — using hearing and only 1% using other sense organs. Most likely something similar will be fair also for the computers entering interaction with the world around. Myriad X VPU – one of the first steps in this direction.

Packet of Neural Compute SDK and the processor for image processing of Myriad 2

The packet of Neural Compute SDK (NCSDK) provides compilation and execution of the neural networks created using frameworks of Caffe or TensorFlow on the Intel Movidius Neural Compute Stick device (Intel Movidius NCS).

The basis of Intel Movidius NCS is formed by the processor for image processing (Vision Processing Unit, VPU) Intel Movidiu Myriad 2.

VPU Myriad 2 includes 4 Gbit of LPDDR3 DRAM, accelerators for image processing and an array from 12 vector processors with architecture of a superlong instruction word (Very Long Instruction Word, VLIW) called in this case SHAVE. These processors allow to parallelize work, increasing thus work speed.

Flowchart Myriad 2 VPU

Movidius NCS is connected to the processor of applications (Application Processor, AP) as which it is possible to use, for example, Raspberry Pi and UP Squared board connected through the USB interface.

The LEON CPU processor constructed on architecture of SPARC which is a part of Myriad 2 VPU is initially empty. At the beginning of work in it it is stitched managing software. It occurs at connection of Myriad 2 VPU to AP on USB. AP interacts with Myriad VPU via Neural Compute API interface, AP opens the device and pumps over managing software (firmware) from Neural Compute SDK in Movidius Neural Compute Stick. Then the host computer sees Movidius NCS as the normal device. This device connected to USB 2.0 or USB 3.0 can load into itself columns of neural networks and the commands managing work of neural networks.

At last the graph is loaded into memory of DRAM on API, at the same time the LEON processor coordinates process of receiving the graph and his distribution by the SHAVE processors serving as accelerators of neural networks. The result of work of networks is returned via the same interface. Besides LEON controls physics of work of Movidius Neural Compute Stick.

NCSDK include means for preparation of neural networks on a host - it is mvNCCompile, mvNCCheck and mvNCProfile.

Can be an example of use of Myriad 2 VPU:

• The quadcopters of new generation capable to distinguish landing conditions and to submit to management by means of gestures

• Smart cameras Google Clips

Google Clips

Before entry into Intel in 2016

Movidius company (nee Movidia) generally Irish, and so Irish that, speaking at its 10-year anniversary in 2015, the prime minister of this country told that Ireland should become the leader in computer vision. Such innocent exaggeration of the state opportunities is pardonable to the leader of the small, but proud country.

Scientific employees-mathematicians from college of the Holy Trinity in Dublin David Meloni, Andy Nisbett and Xing Mitiel and also Valentin Murosan from the Timishoara university (Romania) became founding fathers of Movidius. From 1994 to 2005 their work was limited to researches.

Success of Movidius considerably reminds triumph of Nvidia. The processors which brought to these two companies, though different in the scales, glory and money in both cases thought not absolutely for those purposes to which they began to serve. Movidius company was it is created long before a boom of neural networks. At first creation of the processor allowing to accelerate solving of tasks of game physics was the purpose. So call simulation of laws of Newton in virtual environments.

Timeline Movidus

"Flash card" of Neural Compute Stick

The core of the last independent developments is the multinuclear vector processor SHAVE with low power consumption. In 2007 the first experimental ISAAC chip, in a year – the second SABRE, in 2010 the first serial processor, in 2012 - Myriad 2 was made. The Neural Compute Stick Deep Learning Development Kit and "fleshka" set became the last independent development.

Robotics