Benchmarks Of Many ARM Boards From The Raspberry Pi To NVIDIA Jetson TX2

[GraceAsylum]

I think when you get above $50 for a SBC it gets a little murky. You could get a Intel atom with more performance. Libreboot even hints that the Intel ME stuff may not be active in the atom.

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157726

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157729

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157728

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157731

[GraceAsylum]

AMD may wind up in a chromebook and to me that is a good precursor to AMD on a SBC.

https://chromeunboxed.com/amd-powered-chromebooks-they-may-be-closer-than-you-think/

[NotMine999]

I think when you get above $50 for a SBC it gets a little murky. You could get a Intel atom with more performance. Libreboot even hints that the Intel ME stuff may not be active in the atom.

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157726

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157729

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157728

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157731

I have used a few of these devices for some "not quite embedded" applications simply because I could get "up & running" on x86 style code much much faster than struggling through setting up some embedded ARM device. Even the power usage of the devices you mention are getting close to embedded ARM, but they are still many Watts away from being truly competitive.

I have a few "repurposed" retail ARM devices that draw about 5W steady on 120VAC 60Hz "at the wall socket". I only got those ARM devices for a "proof of concept" and got them "up & running" with some ease because other really cool people on the Internet had already created bootable Debian images (or Arch Linux if you wish) and install procedures for them. Otherwise I would have had many hours of work getting those "repurposed" ARM devices working. It is not like those ARM devices were running some obscure ARM creation; they are based on commonly available Marvell Kirkwood SoC chips. When I got those devices working it was a few years ago, in the pre-DTS days of ARM on Linux, and they still work fine to this day.

I agree with those in this thread that running x86 style hardware on embedded devices would be easier than ARM (x86 architecture is much more standardized and information is more widely available), and it would definitely speed up development, but it all comes down to costs. The embedded world is based on "high volume, low margin, lowest possible cost & power consumption, and the absolute barest minimum of features". Why? Embedded devices are used in hundreds of thousands of devices that need to do mundane repetitive tasks on a 24x7 basis as part of a larger product, not as an add-on.

Some of embedded devices have tight power budgets, are battery powered, or have a primary component that draws lots of power leaving minimal remaining power for supporting devices without jumping up to a larger (and perhaps undesireable) electrical connection. So the low power ARM world, and that is still a hallmark of ARM compared to Intel & AMD offerings, still gets a good "look see" along with true embedded & custom silicon offerings.

ARM devices tend to be based on customized designs, and that doesn't help matters much. ARM Holdings creates "intellectual property" that is licensed to actual chip designers than then sort out what they want to build for any given application. Both Intel and AMD do all of that sort of work "in house" based on what Marketing says.

ARM-based devices can save power by only having the compute and add-on features that they really need; unused features still use power and saving power is the key. Intel and AMD are making inroads into turning off power to parts of their CPU and APU devices in an effort to save power, but none of that can be externally managed; it's all based on functions integrated into the overall design of the chips.

I like to think of ARM-based devices as more "application specific" in their designs compared to "general purpose" processors that we all know from Intel and AMD. I think those 2 different design goals are more "divergent" than they are "convergent", and very unlikely to change unless there is a YUGE uptick in demand and sustained consumption for such converged designs & devices.

[jurgel]

Thanks for providing performance per dollar comparison.

[LoveRPi]

Unfortunately I don't have the XU4, that's why it wasn't tested.

We will send you one along with the Firefly RK3399 this week. Will have the HiKey 960 and another board to you in April.

[t.s.]

I think when you get above $50 for a SBC it gets a little murky. You could get a Intel atom with more performance. Libreboot even hints that the Intel ME stuff may not be active in the atom.

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157726

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157729

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157728

https://www.newegg.com/Product/Product.aspx?Item=N82E16813157731

Or you can get something like this:
http://www.gearbest.com/tablet-pcs/pp_268587.html
http://www.gearbest.com/tablet-pcs/pp_256039.html

Seriously, these SBC board anything that > $70 is insane. You can get a complete tablet with that price.

[LoveRPi]

Seriously, these SBC board anything that > $70 is insane. You can get a complete tablet with that price.

The issue is low quantity production and very technical support requests. The development costs in hardware and software for a reference development board that exposes all of the chip's features are quite high. You have to run through a few designs to get all of the quirks out and that cost is not offset by large quantities. It's not quite the same as the Raspberry Pi which can be considered by all intents as a consumer toy.

[t.s.]

The issue is low quantity production and very technical support requests.

What kind of very technical support request? I understand about Low quantity production, but that can be mitigated too with: create a design that future-proof, or can be put to TV Box, or just copy RPI design, among others.

The development costs in hardware and software for a reference development board that exposes all of the chip's features are quite high.

Example?

You have to run through a few designs to get all of the quirks out and that cost is not offset by large quantities. It's not quite the same as the Raspberry Pi which can be considered by all intents as a consumer toy.

There's SBC that intent to be consumer toys too that's quite expensive. For design, they can make a future-proof design, so there's no need to always designing new boards. Hardkernel is one the folks that made it right, IMO. Good product, quite cheap. The downside is that they shipment rate is high: $16 minimum. Or 1/3 price of Odroid C2.

[GruenSein]

that is just the thing, i agree here but AMD already have products for this and it would not be a franken CPU either since it designed to scale from very low power to servers and GCN is already used on embedded(that for some reason are not available easily to buy).

IoT goes from cell phones, routers, industrial to rpi like boards and it uses pretty much the same SoC, it just changes the delivery format(the board itself) and as someone that works with this products and know many people on the industrial low power sector, i can tell you this WE HATE ARM TO THEIR BONES and i mean hate like "wow we are surprised nobody have gone to qualcomm with a machine gun and go crazy" hate. I can tell you ARM is the most unsupported shitty piece of hardware junk you can get your hands on(but is the only thing we have).

If today you decide to create an embedded device for market release in Q1 2018, you will spend prolly 3 or 4 weeks on C/C++ doing the actual code of the product and the rest of the year until 3 hours before release, you will be fighting the SoC everyday, 10h a day, you will refactor your OS at least 7 times, spend 6 months debugging(to find out is driver bugs that won't ever get fixed), 3 months begging your SoC provider for a fix that will never come, at least 3 meetings in this timeframe will end up being you wanting to quit and ipads/laptops flying in the room, until 1 month before release management settle and let you chop half the product features to bypass the miriad of driver bugs you encountered and finally release a product you won't ever want to touch with a 18m pole to update it again just to move to the next nightmare for 2019 release.

Intel and nVidia are IoT developers wet dream but sadly their offers are so out of market that management cannot afford to use them due to costs and here is where AMD can find a very interesting market and gain developer support to go for the mobile market on android later on (be it AArch64 or x86)

You paint quite a picture.

[yaneti]

I'd like to echo the AMD sentiment here. I wish they produced the SOC and found a reliable SBC manufacturer to participate in this segment. Specifically I would really like someone to challenger the likes of UP Squared, which would be the perfect SBC for my needs.