[content] => 
    [params] => Array
            [0] => /forum/index.php?threads/chips-for-machine-learning-part-ii.8252/

    [addOns] => Array
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021171
            [XFI] => 1050270

    [wordpress] => /var/www/html

Chips for Machine Learning (Part II)

Al Gharakhanian

New member
I covered the Machine Learning (ML) product offerings from Alphabet, Nervana, and NVIDIA in my last week’s post and I intend to touch on the offerings from Wave Computing, Movidius, and Greenwave Technologies in this post.
Let me begin by indicating that the material presented today will be obsolete and incomplete very quickly since there are several other un-announced products in the works.

Wave Computing
Wave Computing (Wave Computing) is a venture-backed company led by Derek Meyer formed six years ago and just recently came out of the stealth mode. They have built a massively parallel data flow-processing chip (DPU) targeting Deep Learning applications. The company seems to have a very clear and concise product strategy revolving around providing native runtime support for both TensorFlow and CNTK. CNTK to Microsoft is what Tensorflow is to Alphabet, they are both open source ML development frameworks. Wave’s chip is intended for datacenter applications and the company is also building 1U servers based on their chip. It is not clear whether the company will be selling chips, or servers, or cloud services. They have raised around 16M is VC funding as well as $9M in debt financing. Additionally they have received a substantial NRE from a strategic early adopter.

Movidius is a company that has taken a very different approach from the others. They have opted to address the machine vision market and have built a product (Myriad 2) that is intended to be deployed at the edge of the network (vs. the datacenter servers). Myriad 2 seems to be suitable for VR/AR goggles, wearables, robotics, and surveillance cameras. The company was formed in 2006 and has raised roughly $87M in five rounds. They have operations in Silicon Valley, Ireland, and Romania. The Myriad 2 VPU (Vision Processing Unit) is the world's first dedicated ultra-low power vision processing engine with ML capabilities. Designed for the next generation of smart devices, the Myriad 2 enables advanced computer vision and is low power enough to be used in portable battery-operated gear.

Greenwaves Technology
Greenwaves Technology (Internet Of Things (IoT) processor - GreenWaves Technologies) is a very young company based in France developing an eight-core chip called GAP-8. Similar to Movidius they are betting that certain IoT applications will require substantial computing power at the edge of the network and not on cloud servers. Greenwaves’ SoC, dubbed as GAP-8; is more general purpose than Myriad 2 in the sense that it can be used in non-vision applications as well. The company claims that the chip can deliver the highest computational horsepower (GFLOPs) per watt among peers. The company is tackling the power dissipation from multiple angles one of which is offering soft implementation of wireless OFDM modem functionality that can reduce chip count, power, and cost. They are also addressing the “dirt cheap” cost mandate of IoT by using a mature process node (55 nm).
The chip provisions a dedicated core for supporting TensorFlow computations natively. GAP-8 is architected based on RISK-V PULPino architecture. PULPino (Parallel Ultra-Low Power Platform) is an opensource processor co-developed by several European Universities.

It is also noteworthy to mention that similar to NVIDIA, AMD’s GPUs have also been used in ML applications. As for IP vendors, the offering from CEVA is noteworthy.