Nvidia's Jetson TX1 ARM development board | Hacker News Reader

Nvidia's Jetson TX1 ARM development board

phoronix.com

69 pointswillow988655 comments

Another 64-bit ARM board out there: the Gigabyte MP30-AR0 http://b2b.gigabyte.com/products/product-page.aspx?pid=5422#... http://www.cnx-software.com/2015/03/27/gigabyte-mp30-ar0-is-... . Unfortunately this one has limited distribution and $ask pricing.

That really frustrates me. I'm running a Dreamplug [1] as one of the main boxes on my network. The second USB port has died and it's really starting to show its age.

The Gigabyte board would give me enough SATA ports and grunt to run a decent NAS - I think it'd give the C2750 from Intel a run for its money, power wise.

Your linked article has a comment suggesting the Gigabyte board retails for 987 Euros. Jeebus.

[1]: https://www.globalscaletechnologies.com/t-dreamplugdetails.a... ($50+ JTAG boards... yay! /s)

The dual 10GbE onboard is very nice, but it doesn't help to encourage optimism about the price ...

The dreamplug is probably closer to the pogoplug in feature set (and definitely cost!).

http://www.amazon.com/gp/product/B005GM1Q1O/?&tag=bcgpfeed-2...

It's dirt cheap ($10-20), easy to load linux on [1], and it can be a simple file server, it can run minidlna, google cloud print [2], all kinds of stuff. 800MHz is not a screamer, but most simple I/O tasks like A/V streaming and printing are a walk in the park.

[1] http://archlinuxarm.org/platforms/armv5/pogoplug-series-4

[2] http://geekvisit.com/hacking-pogoplug-mobile-adding-airprint...

oh I so forgot about these things. I might scrape a few to leverage some hdd lying around to make some nice cache, server, tiny app, foo.

Funny you should mention that. I've got one on the bench from a failed u-boot upgrade (worked after first reboot, then never again). I have the right tools to fix it, I'm just wondering whether it's worth the time or effort. They're $40-50AUD here as they're not sold locally (?)

The X-Gene 1 processor is made on 40nm process, so it is hardly competitive

I don't see a GPU? as opposed to the Nvidia board

It has a VGA port and what seems to be a verrry modest video system. However, it also has two PCIe3 ×16s, so presumably you can use your own GPU, if the software support is there. There doesn't seem to be an integrated GPU in the AMD dev kit http://www.amd.com/en-us/innovations/software-technologies/s... either, so presumably there is Linux ARM PCI GPU support out there, unless everyone is just SSHing in?

Yet another: the 96Boards[1] DragonBoard 410c[2][3] from Qualcomm has a Snapdragon 410 quad-core Cortex-A53 with a good amount of connectivity and IO interfaces (as per the 96Boards spec) and an Adreno 306 GPU which supports the open source Freedreno[4] GPU drivers (if you have a binary blob allergy). That said, this thing has a whole different performance tier than the TX1, but if you wanted to start poking at 64-bit ARM, this thing might be worth a look, especially with the comparatively low price.

[1] https://www.96boards.org/

[2] https://www.96boards.org/products/ce/dragonboard410c/

[3] https://developer.qualcomm.com/hardware/dragonboard-410c

[4] https://freedreno.github.io/

The price is quite a disappointment particularly when comparing to the Shield pricing (and how long NVidia has taken to present the kit - the X1 was introduced now almost a year ago / Shield on the market since early this year).

Let's see in the next days how performance of the board stacks up once the embargo on publishing test results is over.

Remember that the 1TFlops published is FP16 not FP32.

Dev boards are usually priced much higher than consumer products since you usually don't have the same scale of manufacturing.

I think with the dev kit you're mostly paying for the motherboard-type thing and all the software.

The module itself doesn't seem that expensive, given that you're getting a top-notch ARM SoC. Personally I'd be interested in a group buy when it gets released (early next year, for $299 in 1k quantities).

The only Cortex-A72 chip that I know of is the MediaTek MT8173, and the GPU on that is nowhere close to a Tegra X1. So it's not like this is a previous-generation chip, despite being close to a year old.

I'm disappointed too - the Jetson TK1 was marketed at $192, so this is more than triple the price. Dev kit or no, their costs didn't increase that much.

Unfortunately they're pretty much the only game in town for this kind of hardware. There's a few alternatives (Parallella, FPGAs, etc) but nothing with the kind of ecosystem that CUDA provides.

Also remember that NVIDIA loves saying that this is as fast as the ASCI RED Supercomputer from 2000... but that was FP64 1TFLOP machine. This NVIDIA chip gets a total of 16 GFLOPs of FP64... a whopping 1/64 the power of the ASCI computer they try to compare it to.

Very impressive, but it feels expensive to me. I'm cheap though.

The linked article was pretty light on the detail, and weirdly written. It never states how many cores the CPU has; I had to check Nvidia's page (http://www.nvidia.com/object/jetson-tx1-dev-kit.html) to find out: it's a quad-core so four cores.

Some pretty nice low-level embedded-style I/O on there too (GPIOs, I2C, I2S, SPI, TTL UART).

According to Wikipedia it's 8 cores (big.LITTLE, so depending on the kernel and microarchitecture they might not all be usable at the same time):

https://en.wikipedia.org/wiki/Tegra#Tegra_X1

Right ... that's weird, why isn't Nvidia's own Jetson-page more clear on this? Confusing.

~~You only get 4 usable at any given time. Either the big core runs (A57) or the LITTLE core (A53) runs.~~

Sounds like I'm mistaken.

Is this specific to Nvidia chips? Because that's not how it works on big.LITTLE SoCs from other manufacturers. In those you can use all cores at the same time if you want to.

Samsung 5410 worked like that, but apparently that was a bug in that SoC (fixed in the 5420 so you can use all 8 cores).

It's basically a stock Tegra X1 SoC built out into a single-board computer. so it's best to just look and see what has been previously written about the X1.

As such, I don't understand the embargo on performance measurements - barring any major surprises, I think Tegra X1's performance is pretty much known at this point.

The X1 has actually got 4xA53 and 4xA57 cores in a big.LITTLE style arrangement.

I hoped that it might be affordable like the TK1 dev board. But the card sized one is 300$/1k and the dev board 600$... 300$ for students. http://www.anandtech.com/show/9779/nvidia-announces-jetson-t...

I'm sure you'll see a market in 3rd-party carrier boards & packages with the X1 module on it.

Love the fact that it's got a proper x4 PCIe slot on the dev kit.

Hopefully someone will sell the modules with a MOQ of less than 1k though!!

Yeah, I'm really hoping someone offers carrier boards much like SparkFun did with their blocks[1] for the Intel Edison which passed the Hirose DF40 connector through to allow stacking.

[1] https://learn.sparkfun.com/tutorials/general-guide-to-sparkf...

That is non starter I think. High performance expensive dev kits don't have good track record at facinating the market.

It's astonishingly beefy, with games-console (or more) levels of computing power. Very few people need that on an embedded board.

This isn't an "embedded board".. It's got a GPU, it's got PCIe, it's 64-bit with 4 cores.

Yes it is: embedded is a use category, not a performance specification. It's certainly not a "desktop" or "consumer" product as it doesn't even have a case. It's designed to be embedded within another product or system.

Last-gen game consoles levels though, which are more than 10 years old now.

This is exactly what I need for my deep learning research. Iv been abusing my raspberry pi 2 with heavy neuro nets. Shame about the prise though. I will wait for the price to drop(hopefully).

Heavy neural nets on a rasp-pi? How does that work out?

https://www.bitpi.co/2015/03/31/basic-deep-learning-raspberr...

Dunno, depending on what you're doing the Shield TV costs less and does about the same. People are running Ubuntu on these today, and you get the video, XHCI USB, GigE ethernet and WiFi working fine. Of course having a more "normal" platform is nice, but does a UART really cost 300 dollars more? (The ShieldTV has no physical serial port, at least not one I found anywhere).

Also see https://github.com/andreiw/shieldTV_demo

nVidia...are you listening? Uncripple your firmware so booting custom images is not a song-and-dance (you broke it in 1.4!) and at least TELL us where the UART pads are on the motherboard. If you're really cool put together an "official" Ubuntu image that runs on the TX1 and the Shield.

Yep, I'm hoping that NVIDIA's Linux For Tegra[1] 23.1 release is installable on the Shield TV since it'll be adding AArch64 support for the TX1.

[1] https://developer.nvidia.com/embedded/linux-tegra

Wish there is an affordable ARM SOC board for building DIY Spark clusters, which has a 64bit CPU, 8 cores, gigabit ethernet, SATA, USB3, and 4GB memory (8GB memory would be even better), and under $100.

Spec-wise, Odroid xu4 [1] from Hardkernel is very close to meet this requirement, though still lacking in cpu and memory(only 32bit and 2GB memory for xu4).

[1] http://www.hardkernel.com/main/products/prdt_info.php

What would be the usage? Spark for?

I would use it for personal projects at home. Just for fun.

Though I would envision there might be a niche market for affordable Spark clusters as appliances, say, one 1U box that contains 20 boards with total 160cores, 160GB memory, etc and only 200watts power consumption.

I'm curious: How are you supposed to use their card-sized module outside of the developer kit? Is the 400-pin connector something that's standardized?

Also would you need to add your own cooling here? The developer board looks like it has a beefy fan-based cooler whereas I don't see one on the module.

There's a bit more about the module and devkit here: http://devblogs.nvidia.com/parallelforall/nvidia-jetson-tx1-...

Frustratingly the details for the connector are probably in the Embedded developer centre, which requires a registration. https://developer.nvidia.com/embedded-computing

The SoC (card sized) has everything you'd need, not just the connectors to the several I/O and networking interfaces you may need. You will hook in connectors are you like with the 400-pin interface on the SoC and get going.

I am looking for something that can replace RPi2/Edison in my humanoid robot for real-time computer vision. Is Jetson TX1 a good candidate? Is it a drop-in replacement with a similar power envelope as RPi2/Edison, i.e. can be powered by a small LiPo battery?

It is an excellent candidate if money is no object.

It's more power hungry than RPi2. Don't know much about Edison.

For sure: "Jetson TX1 draws as little as 1 watt of power or lower while idle, around 8-10 watts under typical CUDA load, and up to 15 watts TDP when the module is fully utilized, for example during gameplay and the most demanding vision routines."[1]

Anecdotally, the RPi2 has been measured[2] around 4-6 watts under load and the Edison was measured[3] around 1 watt.

[1] http://devblogs.nvidia.com/parallelforall/nvidia-jetson-tx1-...

[2] https://github.com/geerlingguy/raspberry-pi-dramble/wiki/Pow...

[3] https://www.openice.info/2015/07/24/three-days-with-intel-ed...

Interesting that they seem to have dropped Project Denver in favor of A57 cores. I guess the issue is that big.LITTLE has big power advantages?

No, it's more that Denver sucked. A57 is just faster.

Got any references?

Just look at the various reviews of the Nexus 9 (for example http://www.anandtech.com/show/8701/the-google-nexus-9-review... ) and compare it to the Shield Tablet which has A15s, not even A57s.

If Denver's JIT happens to do well the device flies, but if you're not a benchmark or the benchmark is too large to be easily JIT'ed, then the performance falls off a cliff.

I don't think those are very reliable benchmarks because they throw the entire 64/32-bit browser engine and Android stack (which at the time was the very first spin of Android 5) into the mix.

Then there's the mix of clockspeeds and process tech.

Note that even on those it wins a lot of benchmarks. So I see little evidence Denver sucked.

I am trying to come up with a target market for these boards at these prices.

The development board could $500 or could be $1500 that is understandable, the big question is where is the market for the actual devices with the $300 board inside.

Some sort of gaming kiosks or maybe some sort of industrial use?

Most embedded users do not need that sort of GPU performance.

The people doing GPU computing would be using desktop hardware

Keep in mind the TX1 is just an early version trying to encourage adoption. The same chip is shipping in $200 tablets, so obviously the cpu can be pretty cheap.

Seems well within the cost of say your average pro/enthusiast level quad rotor if it could enable dynamic path finding, obstacle avoidance, and following a target.

Current systems can't for instance follow a bicyclist through a forest without hitting bushes/trees.

Current systems can't fly indoor automatically and avoid furniture, people, walls, etc.

The justification for a chip like the X1 is that it has enough CPU/GPU power for realtime vision type applications.

While I don't have any use for it, my systems programming geek soul loves the spec list, given the board size.

good replacement for TK1. Hope we start seeing tablets with X1 soon. And looking forward to Pascal next year!

Reply on news.ycombinator.com