fbpx

5 Startups Building Artificial Intelligence Chips

June 25. 2016. 4 mins read
Table of contents

The first thing we asked when we were turned on to this niche was, what’s an artificial intelligence chip? It’s best to first think about what artificial intelligence software requires which is a great deal of processing speed, then a great deal of power in order to feed that processing speed. However, it’s not just speed and low power that matter, it’s also the way the processor functions. This excerpt from MIT Technology review explains why we can’t just use a high-end Intel processor chip for artificial intelligence:

While a top-of-the-line Intel processor packs more than enough punch to run sprawling financial spreadsheets or corporate operations software, chips optimized for deep learning break particular types of problems—such as understanding voice commands or recognizing images—into millions of bite-size chunks. Because GPUs like Nvidia’s consist of thousands of tiny processor cores crammed together on one slice of silicon, they can handle thousands of these chunks simultaneously. Assigning an Intel processor to such work would be a huge waste of resources, since each of these processors contains a few dozen cores that are designed to run complex algorithms. Deep-learning chips don’t need to do that much thinking to handle all those micro-tasks.

The term used above called “GPU” is an acronym that stands for Graphical Processing Units which are chips that are used in artificial intelligence applications like image recognition and computer vision. The actual way you architect a chip can be optimized for specific artificial intelligence tasks like image recognition, voice recognition, or big data analysis of any kind. In the case of deep learning, you use artificial neural networks which simulate the behavior of the brain by creating simulated neurons. Here are 5 companies that are building chips and hardware solutions that promise to optimize artificial intelligence tasks.

KnuEdge

KnuEdge really isn’t a startup since they’ve been operating in stealth mode for 10 years now. Founded by a former Head of NASA, KnuEdge recently came out of stealth to let the world know that they’ve taken in $100 million from unnamed investors to develop a new “neural chip” which the company describes as follows:

KNUPATH delivers chip technology based on LambaFabric™, which will enable neural computing through an entirely different architecture than the GPUs, CPUs and FPGAs currently on the market. LambdaFabric is inherently designed to scale up to 512,000 devices in the most demanding computing environments, has rack-to-rack latency of only 400ns, and low-wattage, 256-core processors. KNUPATH technology is entirely new and built on biological principles, and will reset the standard for chip/system-level compute in data centers and consumer devices.

That chip technology is supposed to offer a 2X to 6X performance advantage over similar chips and the Company has already generated revenue from sales of their prototype systems. KnuEdge is organizing into 3 separate companies under the “KnuEdge umbrella”. KnuPath is their chip offering. KnuVerse delivers proven military-grade voice recognition and authentication technology. Knurld.io is a public cloud API service that allows developers to simply incorporate voice authentication into their proprietary products. KnuEdge claims that it is now possible to authenticate to computers, web and mobile apps and IoT devices with only a few words spoken into a microphone. How nice would it be to never have to remember a password ever again?

UPDATE: 08/24/2018 – KnuEdge may have gone bust, with speculation being raised that they couldn’t find any more money. You can read more about it in this May 2018 article by the San Diego Business Journal, and the fact that their website no longer functions leads us to believe that KnuEdge may have gone kaput. 

Nervana

Founded in 2014, San Diego based startup Nervana Systems has taken in $24.4 million in funding from 20 different investors, one of which was the highly respected firm of Draper Fisher Jurvetson (DFJ). The Nervana Engine (coming in 2017) is an application specific integrated circuit (ASIC) that is custom-designed and optimized for deep learning. This solution is made possible by a new memory technology called High Bandwidth Memory that is both high-capacity and high-speed, providing 32 GB of on-chip storage and a blazingly fast 8 Tera-bits per second of memory access speed. The Company also presently provides an artificial intelligence service “in the cloud” which they claim is the world’s fastest and which is presently being used by financial services institutions, healthcare providers, and government agencies. The idea here is that their new chip will assure the Nervana Cloud platform remains the fastest in the years to come.

Horizon Robotics

Click for company websiteFounded in 2015, Chinese based startup Horizon Robotics has taken in an undisclosed amount of seed funding from investors that included Sequoia and legendary venture capitalist Yuri Milner. They are setting out to build a one-stop-shop AI solution which they described as “define the “brain”of things, for convenience, safety and fun”. We’re not sure if there’s a bit of Engrish going on here but the value proposition sounds pretty compelling considering that some very big names have placed bets on it.

krtkl

Click for company websiteFounded in 2015, krtkl wants to create “a tiny wireless computer for creating something truly different“. Techies will drool over the Snickerdoodle, a dual-core ARM processor, FPGA, Wi-Fi, Bluetooth, and starting at just $65, “the smallest, baddest, most affordable platform for powering things like robots, drones, and computer vision”. The product is actually being funded by a crowdfunding campaign that raised over $160,000. We’ve warned readers about crowdfunding before. In the case of krtkl, their latest update says that they’ve received the alpha versions of Snickerdoodle and will be shipping them out shortly

Eyeriss

Click for company websiteEyeriss actually isn’t a startup yet but we couldn’t exclude it from this list given that it’s being developed by MIT and receiving extensive media coverage. Eyeriss is actually a piece of hardware that is an energy-efficient deep convolutional neural network (CNN) accelerator. Here’s a diagram for all you nerds out there that actually know what that means:

Eyeriss's DCNN Accelerator schematic diagram

Conclusion

The main takeaway here is the hardware depicted above is more efficient and 10X faster and consumes 10X less power than currently available graphical processing units GPUs. Partially funded by Darpa, the MIT team led by Vivienne Sze unveiled the chip at a conference this year which was the first time that a state-of-the-art neural network was demonstrated on a custom chip.

Share

Leave a Reply

Your email address will not be published.

  1. What small company produces the mini microphones used in AI devices? What an exciting time this is for us all in this age!

  2. C.P.U Programming:
    Frequency based die’s work better with a (+)Code and (-)Code – based on a 4 bit structure applying (2)Cycles as one code through the C.P.U. Using Voltage Control In adjusting the Frequency jump rates in the die’s. you can control the C.P.U you created and it is better as an Ai’s.. Your Voltage and Amp’s need to be controlled well in the C.P.U to do this.
    The Die’s you have create on the C.P.U.
    Running one cycle through a C.P.U is slow at the moment.

    I think Running Two Cycles As One Is Better Then One Single Cycle

  3. Core Based language’s in CPU’s are hard to do.
    Unless you can figure out the Cycles in Hz’s.
    I used the On’s & Off’s in the Unit of volts (too control the core’s as it moves through the CPU.)
    0= (-1) (So Eg – 10010100 code would be +1 -1 -1 +1 -1 +1 -1 -1
    1= (+1)
    These +1 &-1 are the base for the code to run through the CPU. As Normal BUT as a Single Unit code in controlling every Peace Off the 8-Bit Code you enter into the CPU.
    In every core based +1 or -1 bit there is a Micro Volt and a Micro amp as well. That you can control.
    I call My software Frequency Base mono-pole coding.
    I’ve Know about this for 2years.

    1. You lost us at “Core Based language’s in CPU’s are hard to do” but we’ve left your comment in hopes that someone can vet the technical accuracy of what you’re trying to convey here 🙂

  4. Core Based language’s… I understand in corporate turbo Boost elements of intel’s cpu’s. The processor’s unit dynamic frequency scaling up frequency to automatically raise certain versions of its operating frequency when demanding task are need to run harder and faster.
    This is what you use, that I used to make the Ai in development. Now… In frequency based language in the the cpu. I used Micro Volts and Micro Amps in the single volt of an (+1volt) or an (-1volt) But being the same Unit or bit or cycle Of Hz’s A hole binary bit… just broken up into 50 bits of the Cycle of 50hz’s… each unit of the cycle is an (+1 or -1) Micro Volt with Micro Amps in it. This give you many values as Frequency bit and bits to control what you want… In the Cpu.
    Now
    Depending on the the value of the 8bits. That you want to control. you then control the cpu’s jump point or as you guys may know already the die’s in the cpu. The speed is the key to the Ai working correctly in the cpu. It has many value’s in the Turbo Tech you can use…. Kindest Regards Tim.
    Hope that explains it for you guys. Better.

    I’m Hopping some one know’s this technology now better… Good luck & Happy Building…