Movidius (company)

Movidius is a company based in San Mateo, California that designs specialised low-power processor chips for computer vision and deep-learning. It was announced that the company was to be acquired by Intel in September 2016.

Company History

Movidius was co-founded in Dublin in 2005 by Sean Mitchell and Dr. David Moloney, with Dr. Valentin Muresan heading up the Timisoara (Romania) design-centre and Martin Mellody as VP of SW Engineering.^[1][2]

David and Sean's vision for the company when they set it up was nothing less than to revolutionise the fabless semiconductor industry which until that point was focused on building processor based SoCs (System on Chip) with hardware acceleration to do most of the intensive processing which couldn't be handled by the CPU.

Fergal Connor (SHAVE architect) and Cormac Brick (now VP of SW Engineering) joined the company in late 2007 and Brendan Barry (VP of HW Engineering) joined in early 2008 to head up Movidius SoC development. John Bourke joined the company as CFO in 2008 and spearheaded Movidius fundrainsing activities along with Sean Mitchell. Paul Costigan joined the company as COO in 2009 and ran the Movidius Hong Kong operation for the duration of his tenure in Movidius.

Between 2006 and 2016, it raised nearly $90 million in capital funding.^[3]

In May, 2013 the company appointed Remi El-Ouazzane as CEO.^[4]

In January, 2016 the company announced a partnership with Google.^[5] Movidius has been active in the Google Project Tango project.

Movidius announced a planned acquisition by Intel in September 2016.^[6]

Products

ISAAC

After the first year of feasibility study on accelerating game-physics, Movidius joined the Hothouse incubation programme in DIT Dublin Docklands centre and started to ramp up the IC design team to get the first test chip “ISAAC” off the ground. The ISAAC design was already very advanced by the time the company took on the first external funding with investment from Enterprise Ireland in August 2007 followed by an investment round from angel investors in November 2007. ISAAC was a fully functional SoC with all of the interfaces necessary to fully test the core communications and computation subsystem and contained eight SHAVE v1.0 processors and a LEON3 RISC as system controller with peripherals and caches. The ISAAC test chip was taped out at the end of 2007 and we had working silicon back on 65nm HCMOS LP process technology from TSMC by March 2008. While the chip design was going on, the software team were working to build development tools and the game physics engine software. At the time it was calculated that the ISAAC testchip had cost around $1M to build starting from a blank sheet of paper including the SHAVE processor and CMX multiported memory subsystem and software tools. During the timeframe much more straightforward conventional SoCs were being built using licensed IP in competing fabless companies for on the order of $10M.

Myriad 1

Using the ISAAC testchip Movidius started to test the market mainly with customers like Samsung and LG in Korea who started the company more towards imaging and video processing applications which in turn fed into the product definition of the production version of ISAAC code-named “SABRE” which eventually became known as Myriad 1. Myriad1 was manufactured in the same TSMC 65nm HCMOS LP process as ISAAC. The entire system ran off a 180MHz clock generated by an onboard PLL as in the case of ISAAC.

The first foray into image and video processing came when the company, together with Toshiba, developed a Myriad1 version in 2010 with a 512Mb stacked DRAM in package and began the development of the first Image Signal Processing (ISP) pipelines and 3D rectification processing for stereoscopic capture from the 2 camera interfaces of Myriad 1 via a pair of MIPI to parallel converter chips from Toshiba which were integrated into a SiP (System in Package) device.

Movidius CTO David Moloney gave the first information on the Myriad1 architecture at HotChips in 2011 which generated a lot of interest in the platform. This interest resulted in Movidius being introduced to Motorola's ATAP division via Kartik Venkataraman, CTO of Pelican Imaging. The relationship with ATAP blossomed, championed by Victor Vedovato and Johnny Lee within Regina Dugan's DARPA-inspired team aimed at building rapid prototypes, in this case using the depth solution developed dual camera module developed using Myriad1.

Myriad 2

With input from the team at ATAP Movidius began to define a new architecture aimed at high performance computer vision applications codenamed “Fragrak” (Myriad 2), targeting 28nm technology. The engagement with ATAP was formalised in late 2012 around a computer vision project for object tracking called “Project Pink”. Project Pink eventually became the Google Project “Tango” when the ATAP team were transferred into Google proper.

The strong interest in Myriad2 from Google ATAP and other customers with the skillful negotiation of Sean Mitchell lead to renewed VC interest in funding the company which had been struggling financially and ultimately lead to a reboot of the company with new VCs in 2013 and with the reboot came a focus on the US market and a Silicon Valley HQ and new CEO.

As of February 2016, Movidius's latest Myriad 2 chip is an always-on manycore Vision processing unit that can function on power constrained devices.^[5] It is a heterogeneous architecture, combining twelve SHAVE (Streaming Hybrid Architecture Vector Engine) 128bit VLIW SIMD processors connected to a multiported Scratchpad memory, a pair of LEON4 UltraSPARC ISA processors for control, and a number of fixed function units to accelerate specific video processing tasks (such as small Convolutions and color conversion lookups). It includes camera interface hardware, bypassing the need for external memory buffers when handling realtime image inputs. In terms of software, a Visual programming language allows workflows to be devised, and there is support for OpenCL.

Fathom

Fathom is a USB stick containing a Myriad 2 processor, allowing a vision accelerator to be easily added to devices using ARM processors including PCs, drones, robots, IoT devices and video surveillance for tasks such as identifying people or objects. It can run between 80 and 150 GFLOPS performance at below 1W of power. The company switched from a previous 65nm process to a 28nm one to increase its chip’s efficiency by 20-30x. The Fathom is expected to cost under $100 per unit.^[7]

See also

• Vision processing unit
• MPSoC
• Coprocessor
• Convolutional neural network

References