Nvidia For Video Surveillance Examined

By Brian Karas, Published Aug 23, 2017, 11:23am EDT

Nvidia is making a big push into the security market, claiming more than 50 partners, investing heavily in industry events, positioning their GPUs as the workhorses of deep learning applications. 

However, Nvidia offers multiple product platforms, including Tesla, Jetson, etc., based on different GPU architectures such as Pascal or Maxwell. It can be difficult to understand how these components are used in video surveillance systems (e.g., what to use in a camera vs server vs the cloud, etc.).

In this report, IPVM reviews Nvidia's products related to surveillance, including where the various Nvidia product lines are most applicable, how they compare to alternative products, and their limitations.

******** **. *********

*** ******* ******** ******* on ***** ** *** ******* tasks *** *** ** ***** used ***:

  1. ******** - ********* * set ** ****** ** learn ************* ** *********** among **** 
  2. ********* - ********* * given ***** (** ******* of ** *****) ** learned **** ** ******** ** object 

******** ** **** ********* *** time-consuming **** ********* ** *** typical ******** ***********, *** will **** ** *** the **** ********** *****. In ******, *** ***** GPU *** ** **** for ****** ********, ****** training ** **** *************** intensive *** * *** developed *** ******** ************ would ****** ** ******** for *********.

Product ******

*** ********* ***** **** ****** ***** how **** *** *********** their ******** *** ************ applications:

**** **** ** *****, the ***** ***** ******** of ********** ******* *****, which **** ***** ** correlate ** ****** ****, higher ***** *****, *** larger *********.

GPU *************

******* ** *** ***** categorizes ********** ** ***** names/families (*.*.: *******, ****, Xeon), ****** *** ***** * similar ********. ****** *** ***** GPU ************* ******** **** (ranked **** ******/**** ******** to ******/***** ********):

***** ******* ********* ***** come ** ******** **** use ********* *** *************. And ***** **** *** generally not **** ** ******** components, *** *** ***** sells ****, *** ******* tend ** **** ** part ** * ****** circuit ******** ** ***** product, ********* ** *** intended *** ****.

Product ********* - **** ******** **** ***** **********

*** ***-***** ******** **** ********** ** video ************ ***:

  • *****- *** **** *** servers
  • ****** - ****-***** *** *** cameras
  • ******* *** - ***-******* ******** *****
  • ******- ****** *********** *** graphics *****
  • ***- **** ***** ******** ******** for ********* **********

**** ****** ** **** of ***** ** ********* below.

Tesla - *******/****

*************** ** ********* *** most ****** ****** *** surveillance *****. ** ** ********* sold as * **** **** to ** ********* ** a ******, ** ** NVRs ******** ** *********** a *** ****, ****** a **** ****** **** interfaces ** ******'* ****** ***, ***** ****** ****** interconnect ********* ** **** available *** *********** ************.

****** *** ** **** *** **** inference *** ********, ********* on *** *****. *** most ****** ****** ********* marketed ** ************ ************ are (****** ********* *** platform, *=*****, *=******):

  • ****- ******* **********, ******* power, ******** ** ******** R&D ************ ********* *******-********* training. ~$**,***+
  • **** - ******** *** ******** use, ~$*,***-$*,***, ********* ** model/interface
  • ***- **** ******* ********/*********. Used ** **** ***** servers, ******* **** ***** do ******* ******** ** well ** *********. ~$*,***
  • **- ********* ************, **** common ** *** ******* or ********** ******** ****-******** applications (*.*.: *********** ******, advanced *********). ~$*,***

***** *** ********* ******* *** ***** ****** *********, ****** ***** *** less ******** **** ** video ************, ************ *** Kepler ****, ***** **** positioned **** ** ******* compute ************ **** ******* learning/deep ********.

*******, ***** ***** ** the **** ****** ****** now, ****** ** *********** the ************* ** **** learning ** ** *******.

Jetson - *********** *******

******** ******'* ****-************ ****, ******** for ******* ******* **-***** deep-learning *********. ***** **** in *** ******, *** TX1 (******* ************) *** the *** (****** ************).

***** *** ****** ****** is ****** ********, ** is ****** ($*** - $500 ***** ********* ** volume), *** ****** ***** (50 ** * ** mm / ~*" * 3.5") ******** ** *** electronics ********* ***** ****** modern ************ *******. ******** seen ** ******'* ***** **** demos ***** *** ****** modules **** ********** ****** than ******* *******. *** cost **** ******* *** entire *** *** **** some **** ******** *******, making ** * **** costly ********* ** ******* in * *****-****** ******.

GTX - ******** *****

*** ******** *********** **** **** graphics *****, *** *** not ******** *** ****-******** use ********, ** ****** **** not **** ***** *** the **** *********** *** environmental ***** ** *** Tesla *****. *******, ***** their ***-**** (********* ***** $500), *** *** **** they *** *** **** GPUs ** **** ** the ***** *****, **** manufacturers *** ***** *** cards ** *******.  

Quadro - ******** ******** *****

****** ***** *** * step ** **** *** cards. **** *** *********** intended *** ******* *********** applications, **** *** ******* market ***** *** ***** like *** ********* ** 3D ********. *******, **** are *** ******** *** more ********* *****, *** have * ****** ******** than *** *** *****, making **** **** ******** for ******* **** *********** closer ** * ***** is ********, *** ***** cost ** ***. ****** cards ********* ***** **** $500-$1,000, *** *** ** higher *** ******** ******.

******** ********** ****** ***** for ***** ********** ****** **********.

DGX - **** ****** / ******** *&*

******-* ***** ** **** *** training **** ******* ******* of ******* **** **** to ** ********. ****** range **** ~$***,***-$***,***, ****** it **** ********** ** organizations ******* ******* ** deep-learning, *** ********** ******** models *** **** ******** than **** ** ********* seen ** *** ******** industry. ** ** *********** an ***** ****-***** ****** with ** ** * Tesla **** *******.

GPU ***** - ***** *********

****** ** **** ********* a *** *****, ***** ********* ******** of ** ****** *** ******* (other **** *** ****** devices) ** * ******, where ********** *** *** applications **** ******* **** GPU ************. **** ******'* EC2 *****, ** ****** server ******* ********, ***** can ******** ****** *** ********* for *****-**** ** ****-**** needs. *** *******, * developer ******** ** *********** designed ** *** ** a ****** ****** ***** use *** ***** *** the ********* ******** *******, and **** **** *** output ** *** ******** done ** *** ***** to *** ** *** Jetson ****** *** *********. Developers ***** **** ***** cloud ********* ********* **** analyze ***** ***** **** cameras *** ****** ********* or ***** ********, ********** the **** *** ********* GPUs ** *** ******, this ** *** ******** that****** ****. 

Rival GPU *******

*** [**** ** ****** available], ******** ********** ** both ****** *** ***** **** has ********* * **** of ***'*, ****** ***** offering ** **** **** limited **** ******'*. *** primarily ****** * ******** of *** ****** ******* [**** no ****** *********] *** **** in * **** **** factor, ****** ** ********* suited *** *********** ********** applications *** ******* ********** to Nvidia ***** *****. *** has **** * ****** competitor ** ****** ** *** desktop ****** *** ***-***** video *****, *** ** have *** **** *** tests **** *** *** two ********* ***-******** ***** against **** ***** *** analytics ************.

*****'* ******** *** ******** *** ****-***** processing ******, ********** ***-***** GPU's *** *** ** smart *******, ******, *** similar *******. *** ******** chips **** ******** ************ for ***** ************, *** contain ***** ************ ***** at ******** *****, ***** the ****** ****** **** ****** is * **** ******* purpose ******. ***** ***** and ***** ******** ********* make *** ******** ***** a ********* ****** ********** to ******.

****** **********

****** **** *** ************ ****** as ** *********** ** sell ***** ******** **** previously ******** ********. *******, ****** is **** **** ** being **** ** ******* high ****** *** ***** products, ***** ** ********* to ** ** ********. When *** ***** ** a ** **** ***** is ******* ***** ** the **** ** *** **** of *** ******** *** the ****** ** ** installed **. *****'* ******* ** the ******** ******** **** relies ** ****-******** ********* finally ****** ******** **** have **** ********* **** over *** *****, ***** the ******* *** ****** direct ******* ****. 

Comments (8)

Great article and very technical as always.  I was not even aware of the DGX. Do you feel there is a future in deep learning for surveillance or is mostly hype at this point?  How far out do you think it is before something truly useful is available? 

First, I do think there are "truly useful" products already available. Avigilon's Appearance Search, Platesmart's LPR, Umbo's consumer camera, etc. None of these have been massive disruptors in the industry, but they are all "useful" for their target applications.

Second, the best applications are likely still to come, in the form of very low cost/main stream products, or capabilities that we have not yet seen. Deep learning itself should help to solve many of the object classification problems (sorting out interesting stuff in the image from the non-interesting). Because deep learning is gaining popularity in so many other industries and applications security/surveillance products are able to ride that development wave and benefit from it, but it is still somewhat of a trickle-down effect.

Behind the scenes there are also companies working on tagging datasets (e.g: this is a picture of a car, this is a picture of a circus bear, this is a picture of a cactus), and selling those datasets to other companies for commercial applications. Couple this with the general "Moore's Law" kind of advancements in GPUs, and we could start to see GPU's that come pre-trained for common surveillance applications. Making it possible for companies with very little analytics-dedicated R&D people to roll out products that rival anything currently available for a fraction of the cost. Much like how today you can manufacture a camera from a collection of already-developed components that will have very good performance specs.

OpenALPR has been optimized to run entirely on NVIDIA's newer GPUs. We have spent the last few months doing the heavy lifting to eliminate CPU bottlenecks. Our code runs on Ubuntu 16.04  supporting TX1, TX2, Tesla P100, Quadro M620, GTX 1050 and newer NVIDIA GPUs (https://developer.nvidia.com/cuda-gpus).

Surprised no one is mentioning this alongside the Nvidia/HikVision Supercomputing partnership, or HikVision's new "Deep Learning" NVRs.

If HikVision becomes capable of delivering better analytics at scale it could be another game changer for the monitoring side of the security industry.  

---

https://ipvm.com/reports/hikvision-nvidia

https://ipvm.com/reports/hik-deep

Hikvision is using the NVIDIA Tesla P4 in their deep learning NVR, according to Nvidia. Given the price of that part, even with volume discounts, the "at scale" part may be challenging, especially when you consider Hikvision had led primarily on a low-cost approach, and the Tesla P4 adds a significant cost to the unit, relatively speaking.

The Nvidia/Hikvision partnership is less significant overall, the DGX-1 that Hikvision bought can be used to accelerate training for DNN's, but Hikvision still has to do the core software development, feed in the training data, etc. Much like their Cisco "partnership" did not inherently make their devices more secure, the Nvidia partnership does not automatically make their devices smarter, they still need to do the heavy lifting.

I fundamentally agree with you. HikVision would have to do a lot of the heavy lifting before this even gets off the ground. That being said it does seem to put them in a good place for the future. 

Thank you # 2

Does anybody know if Hikvision iVMS software utilizes GPU at all? We're about to buy several agent computers, and I'm wondering if I should pick the discrete graphics card or rely on embedded intel one

Read this IPVM report for free.

This article is part of IPVM's 6,738 reports, 909 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

Already a member? Login here | Join now
Loading Related Reports