Axis ARTPEC-7 Chip Release Examined

Ethan do you know which contractor is actually building the chipset in Taiwan? Perhaps they are attending this event: http://www.semicontaiwan.org/en/home

Reactions:

Undisclosed #2

"With security concerns and nation-state cyberspying fears increasing, promoting that one develops their own is a differentiator".

this is not a serious claim IMO, you have no way of knowing how able and professional their firmware engineers are, they could be just as sloppy as hikvision or any other company.

i greatly dislike the "between the lines racism" of this sentence implying that since the firmware is done by a Swede it will be more secure than firmware coded in China. if Axis programmers are sloppy, the chance of a back door exists just the same.

Reactions:

(7)

(14)

(1)

(5)

Undisclosed Integrator #3

In reply to Undisclosed #2

I think the quote is more geared toward ensuring that you know and are responsible for every part of production. Not just "the firmware was programmed by a Swede so it's great." It is more of a this chip was built by us for our device and not manufactured by a company that can be heavily influenced by the Chinese government. It would be the same as Aston Martin coming out with a new car and ignoring all of the questions about the drive train because they are using off the shelf parts from Zhongxing.

Reactions:

(7)

Morten Tor Nielsen

In reply to Undisclosed #2

prescienta.com

The vulnerability is rarely in the chip, and rolling your own is almost certainly more risky than using an established platform. And I'm guessing that the ARTPEC-7 (like the predecesor) is actually based on ARM with some extra peripherals to do image processing in HW. How this makes things "more secure" is not clear to me.

http://lup.lub.lu.se/luur/download?func=downloadFile&recordOId=8871498&fileOId=8871501

In terms of security, isn't Axis' APAC more of a risk? If the camera credentials are somehow compromised, it's possible to install and run clandestine software on a machine (the camera) in the network, and not be subjected to company wide anti-malware policies. I'm not saying it's a huge risk, but I'd argue that this is a bigger risk than the chip itself being compromised.

I'm curious as to the AI aspect as the trend seems to be to move the intelligence to the cloud (or at least, offload the processing from the camera - think of Ring and Nest). It might be that next year your wonderful new ARTPEC-7 based camera comes across as a demented fool compared to a cheap non-name camera fed into a Jetson Nano based board (or just fed into a decent PC).

Reactions:

(2)

John Honovich

IPVM

I'm curious as to the AI aspect as the trend seems to be to move the intelligence to the cloud (or at least, offload the processing from the camera - think of Ring and Nest)

We are equally seeing the trend moving towards the edge. It is not just Axis. Look at the big push Intel is making with Movidius (e.g., we tested their Myriad 2 here). For example, Avigilon is going with that for their H5 analytic cameras. And Qualcomm is pushing into smart IP cameras. And Ambarella has pivoted from video streaming SoCs to computer vision SoCs (see Ambarella on Computer Vision).

Morten, what is your view of what Intel, Qualcomm, and Ambarella are doing?

Reactions:

(2)

(1)

Morten Tor Nielsen

In reply to John Honovich

prescienta.com

I think there's a market for that sort of thing. But I think the biggest value is to the manufacturers of these chips, and that many customers should forget about all this AI hype, and get something more cost-efficient and flexible.

I have long been a proponent for on-edge processing, but only up to a point, and I think we're past that point now.

We also have to be a little careful with the term "edge". NVIDIA for example, considers PC's outside of the cloud as "edge", whereas this industry often considers the cameras "the edge". Maybe I am interpreting the diagram in the deepstream SDK page the wrong way, but it seems as if cameras are just providing video, nothing more.

My impression is that 90% of all installations do not need facial detection at all. If you're in an office, a simple, crude detector is enough, if something moves there's 99% chance that there's a face on the top 1/4th of the thing moving. If you're looking at a parking lot, you're not going to have enough pixels to recognize people and so on. Don't get me wrong, it's a cool toy, but the value is questionable in most cases. In housing, where you may have a single camera at 20 entrances, you can hook it up so that a single PC does the face detection when someone rings the bell.

So this leaves airports, powerplants and other high-profile assets you want to protect. In that case it comes down to this : should we place 100 expensive "AI" cameras around our infrastructure, only to find that 2 years from now, YOLOv5 is 100x more efficient and accurate, but your cameras don't have the chips to handle that. Or should you get cheapo cameras and a few PC's that can be upgraded, and that offer 100x more flexibility (want to do facial detection on another camera? just point the algo to that camera instead).

It's clear to me that image classification is impressive, but it's not at a point where I'd call it "mature". It will improve and evolve, but I am pretty sure the transistor count increase dramatically before we get plug'n play AI where the algo doesn't mistake a snowbank for a bathtub.

What about vandalism? Your precious AI cameras are exposed to the "elements", whereas a Jetson based box is cold and cozy in the locked down server room far away from vandals. If your cheapo cameras break, just get another cheap replacement, there's no vendor lock-in.

I guess it comes down to cost (like all things). These 1st generation cameras are going to be sold at a premium (I guess), but I'm hard pressed to find a scenario where I'd go for this sort of thing vs. a dumb camera and a decent PC or Jetson nano board.

I think that's why the Ring and Nest cameras are super-stupid and cheap. Amazon and Google (and NVIDIA) knows that it makes no sense to add all this silicon to the cameras at this point in time. It's much better to just stream back to the data-center, and then upgrade the software and silicon there.

Reactions:

(3)

John Honovich

IPVM

I think that's why the Ring and Nest cameras are super-stupid and cheap. Amazon and Google (and NVIDIA) knows that it makes no sense to add all this silicon to the cameras at this point in time.

For Ring and Nest, they are cloud first offerings so adding more cloud is perfectly sensible. The problem for commercial / industrial / government video surveillance is that a large part of the market will not allow their video to be sent to the cloud, so manufacturers like Axis, Hikvision, etc. need to choose between on camera, on recorder, or on server but all on premise.

Your precious AI cameras are exposed to the "elements", whereas a Jetson based box is cold and cozy in the locked down server room far away from vandals. If your cheapo cameras break, just get another cheap replacement, there's no vendor lock-in.

What is funny here is this is the same contest / debate about where encoding should be done - i.e., cheap analog cameras with centralized encoding or expensive IP cameras with distributed encoding.

Will see how it all works out but there is no doubt a lot of powerful players are pushing analytics on the camera.

Reactions:

(1)

Jacob Hengel

YourSix

As a cloud-first advocate, I have to respectfully disagree with quite a few points here. If the camera itself can adjust to the scenario before streaming over the WAN we are looking at a large possibility for decreased bit-rate which is obviously critical in cloud scenarios.

We also have the highly critical weak-point of solutions like Nest where the camera has little/no survivability when internet connectivity is impacted/impaired. AXIS utilizes SD cards inside of their cameras when delivering cloud solutions as a place to offer local storage for higher resolution and higher frame-rate streams that can later be retrieved from cloud but that also work even when the internet is down. Having these more powerful chips inside of the camera itself makes the entire solution vastly more reliable and geared toward business and enterprise customers. When cloud connectivity is down, it is up to the device itself to render the highest quality storage available and ARTPEC-7 is a massive improvement here. Relying solely on cloud is a dangerous move for a business to make; AXIS is covering both fronts with this new chipset. Allow the vast benefits of cloud when it is available/connected and plan for business continuity for when it isn't.

Reactions:

(1)

Morten Tor Nielsen

https://techcrunch.com/2019/01/10/amazon-ring-privacy-concerns/

prescienta.com

In reply to Jacob Hengel

You're disagreeing with a straw-man :)

At no point am I proposing that you should stream to a server in the cloud. In fact I totally agree that a pure cloud recording solution (a'la Ring/Nest) is not a substitute for a local VMS. It's a very different discussion, that probably belongs in a different thread.

Consider a customer that gets 100 cheapo (I recommend Hikua) cameras and a semi-powerful NVIDIA based PC that they place on-site (perhaps someone will release a Jetson based box). As the algos improve, the customer simply updates the PC and reaps the benefits of the improved accuracy and speed. If someone vandalizes a camera, he can replace it with any other cheap camera - perhaps even an Axis one if they feel wealthy. If the user is unhappy with the performance of the software on the PC, pick another vendor, get someone else to do the analysis.

Contrast to another customer that gets 100 expensive AI cameras and a cheap(er) PC. Most of the time, there's nothing going on, and so you have all this power that just sits idle. If one of the cameras break, you need to go get another clever AI camera (at what I presume will cost a premium). In practice there's only one vendor who can upgrade the cameras, if they do (they won't, they'll offer a new camera instead), you'll have to update 100 devices.

All of it is done on the LAN, but the "interpretation" of the video is done by a PC instead of the camera.

The first guy/gal is the wise user who makes his CFO happy, and stays in charge of the system. Gets a raise. Takes the family on a nice trip to Disney world.

The other is the schmuck who got had by a slick dude in a big booth at a tradeshow. Next year, he'll have to explain why they are dealing with 200 false alarms per week, and why they can't switch vendors. He'll be working late, while the wife is flirting with random guys on Tinder.

What I read, is that Instead of H.265, you can get Zipstream (like Mobotix's gave you MxPEG instead of H.264?), but I don't see anything in the chip that makes it dramatically better at working as a gateway to the cloud. Certainly not something that can't be accomplished today.

Isn't "Geared towards enterprise" is just a euphemism for "expensive"?

Just need to add this about Ring...

NOTICE: This comment has been moved to its own discussion: AI - On The Camera Or Not?

Reactions:

(1)

Jacob Hengel

YourSix

I believe much of your post is intended as satire so I'm going to get back on track and reiterate my original intent.

Chipset matters for an undebatable variety of reasons (whether in coordination with server or cloud) and what I have seen from ARTPEC-7 means that I will be refreshing even several of the devices on my home. Technology evolves and this is a great evolution.

Reactions:

Morten Tor Nielsen

prescienta.com

In reply to Jacob Hengel

~20% satire.

Reactions:

Undisclosed #5

IPVMU Certified

The vulnerability is rarely in the chip...

who is even looking “in the chip” for vulnerabilities?

have you reviewed HiSilicon’s microcode?

Reactions:

(1)

(2)

(1)

John Honovich

In reply to Undisclosed #2

IPVM

i greatly dislike the "between the lines racism" of this sentence implying that since the firmware is done by a Swede it will be more secure than firmware coded in China.

Your statement is racist. You are equivocating the China Communist Party with Chinese.

However, there are tens of millions of free Chinese living in democratic nations such as Canada, Taiwan, and the United States. The issue is not with the Chinese, it is with the China Communist Party.

For example, why does Taiwan have a ban on Huawei and ZTE network equipment? Are they 'racist' against themselves or do they distrust the China Communist Party?

And why is no one, including the US, trying to ban Taiwanese companies? Are they not Chinese?

The issue is not 'sloppy code', it is the maliciousness of the China Communist Party that controls mainland China.

You are a smart person, you must be able to see the difference between these things, yes/no?

Reactions:

(5)

(2)

Undisclosed Integrator #4

In reply to John Honovich

Come on John! You know how popular it is to be offended these days. If you don't like a Takata Airbag in your car, and you state this publicly, then you must despise the Japanese people. And if you feel nervous boarding a Boeing Airmax, it's just you that deep-rooted bigot inside you. If I don't trust the "autopilot" in a Tesla like I don't trust Hikua, does that make me Anti-American?

Reactions:

(1)

(3)

Undisclosed Integrator #4

Aside from Avigilon, how does the "Signed firmware/secure boot and Trusted Platform Module (TPM) that is FIPS 140-2 level 2 certified," compare to the rest of the industry? Vivotek is doing something with TrendeNet, and Panasonic has "Secure Communication," but does anyone have any tangible progress beside a fancy 10ft banner claiming so?

Reactions:

(1)

George Bentinck

In reply to Undisclosed Integrator #4

Disclosure - I work for Cisco Meraki.

All models of the Cisco Meraki MV12, MV22, & MV72 along with all future camera models make use of a Trusted Platform Module. This is used as a hardware trust anchor for secure manufacturing, secure boot, and to ensure only signed firmware can run on the hardware. This has been in production since February 2018. No user configuration is required.

Reactions:

(3)

Undisclosed #5

IPVMU Certified

Both Dahua and Hikvision have talked about developing their own chips though we have not seen them publicly declare that they have actually finished one.

strictly speaking Dahua has developed its own SOC, though it’s an HD Analog chip, not IP like the Axis chip.

still, just having that capability and experience from CVI SOC production would come in handy if they decide to go IP as well...

you can barely make out the @hua logo in this picture:

Reactions:

Jason Crist

IPVMU Certified

Did I miss the mention of edge capabilities for VMS solutions like Orchid or is that implied in “AI”? Or is this processor not geared towards that?

Reactions:

John Honovich