Nvidia Jetson TK1 Reviewed | Hacker News Reader

Nvidia Jetson TK1 Reviewed

drdobbs.com

65 pointsfooqux28 comments

This device's use case is not to pump HDMI video or act as a RasPi replacement. It is to accomplish heavy GPU tasks cheaply in both power and money.

This post on the NVidia blog shows two good use cases: http://devblogs.nvidia.com/parallelforall/low-power-sensing-...

Closest competitor would be the Parallela. http://www.parallella.org/

Price is $192; link to buy is here https://developer.nvidia.com/jetson-tk1

Would be a great device for computer vision tasks. But it's not cheap enough to displace the Raspberry Pi on one hand, and it is unlikely in general to displace Gallileo and other trimmed PCs on the other. (Yes, it is significantly more bang for the buck - but until the ecosystem catches up, it's much less useful)

The tegra could cost[1] $26 with some time and large volume.

At that price, we might see dev boards with a price closer to $100.

[1]the tegra k1 chip is around 135 square milimeter ,at 28nm (which costs $4000 for a 44000 square mm wafer).which comes down to around $13 chip manufacturing costs.

nvidia's gross margin(55%).

For what it's worth, it's also available from newegg at the same price. Would be worth seeing what site offers the best shipping.

http://www.newegg.com/Product/Product.aspx?Item=N82E16813190...

This device really is a GPGPU with co-ordinating processor bolted on. Any notion of it being for the ever-nebulous Internet-of-Things is just marketing nonsense, but if your task relies mainly on GPU compute capacity per dollar this thing looks amazing.

If you haven't played with non-fixed pipeline GPUs then you probably should as your intuition about how much compute capacity they have is probably off by an order of magnitude.

At the recent show in Asia, Nvidia was keen to promote the video input side of these smaller GPU chips - with particular reference to self-driving cars. In their vision, each car is going to need approx. 8 cameras, and each of these is going to require dedicated 3D-analysis on the video in real time (object localization, identification, etc). (#cars * 8) makes #GamerPCs look small...

It's seems like a great device for IoDCT, "Internet-of-Display-Connected-Things", as well as the upcoming IoTTS, "Internet-of-Things-That-See".

(I see now why "IoT" is popular. You can take any established computing application with a boring old name such as "computer vision" and thingify it into an acronym of exciting new cloud frontiers.)

I suddenly imagined a cross function device: I don't see why dropping in a new car stereo should not hold a board like this able to process camera and sensor data for my own use, as well as aggregating for research. I'd love, for example, there be standard physical mounts for cameras and sensors, in front and tail lights, above the rearview, in wheel arches, and a data backplane. One of these, plus a display, would make a nice in car entertainment hub, and barely use the GPU abilities, which could switch to cams, or alert me to tire wear, or even ground condition. If connected to a ECU (regs ahead, sure, but..) you might damp steering input to deal, with camber, or unload the suspension at a corner if a object looked like a puncture risk. Maybe someone should make aftermarket lamp assemblies that accommodate bullet style cams. Would take some handy optical calculations. I think automotive sensing needs a boost for collecting data by affording accommodation to such retrofits, without causing people to do hardware mods, that are likely needing certification or plain ugly, and make a supplier alliance for anonymized data. Opt in, I hope, as nobody's done a good job of instilling trust much, but i'd go for it. As for Internet of Things ideas, driver driven vehicles might be able to broadcast data useful for driverless cars. The need is for retrofitting cars as much as hoping new models will sell, as this is the sort of thing that I imagine would be a option. Make it part of something people want, like strong GPU for mapping, and you might have a deal.

Just a related thought to having GPU and vision input on cars. Get enough data and you can start finding the optimal lines through corners. It's a bit theoretical, to sell as a economy feature, or maybe just a little bit premature, but as above, I really do want more GPU horsepower on my GPS device, beyond what mobile phones can supply.

I think it is UPS trucks that Nokia HERE maps rig with LIDAR scanners. I'd like more of that to happen and become public data. Nokia seem to have the majority of auto manufacturers in deals, Microsoft didn't get to buy that bit of the company clean, and I bet they've thought about how to extend journey data capture. That's potentially a nice ecosystem to have strength in, as qualifying anything that my interact with car control is a substantial barrier to entry.

Which leads mw to wonder how long it will be until, assuming someone gains access to a platform, compiling a Google maps street view dataset is possible on a budget. Maybe that is Nokia's plan...

I am wondering if you can add an external GPU via the mini pcie interface:

http://www.hwtools.net/Adapter/PE4C%20V2.0.html#Web-Shop

When are we going to move away from mechanical fans for cooling? They seem to be the weakest link in every board that has a fan on it and die a horrible noisy death at some point way before you expect and they're noisy even when they're new. Surely they can design better?

We have better tech, but it's not cheap. Fans are cheap.

Given that the same SoC is used in the Nvidia Shield tablet, it definitely can be cooled without a fan. I guess that wasn't their priority for the developer board, or it would have made it harder to fit in the various I/O interfaces.

Given its power requirements and lack of anything wireless, I have a hard time buying the IoT moniker. I could see people using this as the guts of a MAME cabinet or an extra-beefy XBMC client, although a $35 raspberry pi would probably suffice just fine for both tasks.

I wonder if you could underclock the K1 enough to go fanless. Though given that the SoC is soldered directly to the board and is probably impossible to find a replacement anyway, messing it up basically makes the entire $200 board worthless. Kinda risky just to reduce noise and/or eliminate moving parts.

Nice board: GPGPU, SATA, mini PCI-E socket, SD card, 16GB of onboard flash, 2GB of ram, gigabit ethernet, power header for your hard drive so it doesn't need it's own power supply. The single USB header is a bit weird though.

I can see this board making for a great little server: downclock the CPU to go fanless & add an SSD and you've got no failure prone moving parts to worry about.

Why would you be buying a board with 192 CUDA cores to run a server? This is the strongest point of this entire concept - desktop grade GPU to do GPGPU work, and yet you would use it as a server? Why?

> Why would you be buying a board with 192 CUDA cores to run a server? This is the strongest point of this entire concept - desktop grade GPU to do GPGPU work, and yet you would use it as a server? Why?

GPGPU isn't limited to desktop computing, a server heavily utilizing GPU for whatever needs would benefit from this board all the same.

Well if you compare it to the old OMAP4 based Pandaboard its quite an improvement. And with that graphics horsepower it probably runs Unity fairly well. I agree that the "Internet of Things" mention is a buzzword hit not a feature. Same price point as the Arndale board (with the Samsung Exynos SOC). I would enjoy one of these with as much I/O available as the Beaglebone Black.

[1] http://www.arndaleboard.org/wiki/index.php/Main_Page

Unfortunately it doesn't look like Unity can target Linux ARM yet, which is a real shame as I could have justified purchasing at least a couple of these for a current project. I can't see it being too difficult for them to add this in as you can target Android ARM.

http://feedback.unity3d.com/suggestions/platform-support-for...

This is pretty much exactly the kind of device I've been looking for. I wanted a tiny system with a GPGPU that runs CUDA from the get-go. I have an nVidia desktop, but I'm running Windows for work and my other hobbies, so Python + Theano + CUDA are a little tricky. This is pretty much perfect because I can set it up as a Linux device on my network, SSH into it when I need to play with GPGPU stuff, get it working, then push the data to Amazon EC2 when it's ready for prime time. I bought this within minutes of hearing about it.

A 192-core GPU? Am I out of touch with this area of technology or is that a typo?

A high end gaming card has thousands of those. If you read the CUDA documentation you can find out more about the hardware. Physically a high end card has several streaming multiprocessors each with their own caches and access to a global level 2 cache, instructions are sorted into categories (integer math, floating point math, load/store/special functions) each of those correspond to functional units on the chip that execute in lock-step. The current generation has 32/64 of those (called the warp size, notice 192/32 = 6). High end cards have several of those streaming multiprocessors, for example the Titan 30 * 192 = 5760. A warp scheduler on each chip decides what to run, so for all 192 "cores" to be in use you would have to have independent floating point / integer and special function instructions to do. It is the same marketing trick that AMD uses to sell two integer functional units in a four core design as 8 'cores'.

Overall cuda models the underlying hardware fairly faithfully, so if you are interested in the hardware design you should read the CUDA documentation.

It's not actually the same marketing trick as AMD's new CPUs. The two cores in each unit of those CPUs are genuine, full-fat integer cores with their own control flow that can execute unrelated instructions from two different threads or processes at the same time; only the floating-point unit is shared between them. CUDA "cores" are more like individual lanes within the SIMD/SSE unit of a CPU - the control flow and instruction decoding are shared amongst an entire "warp" of 16/32/62 "cores" and all the "cores" execute exactly the same instruction in lockstep on different data.

They are not "real cores". GPUs are massively parallel, so these are 192 computing cores, but they share everything else.

192 cores is right. It's now Nvidia's smallest 'slab', with the workstation/server class chips having multiples of that.

For comparison, Nvidia's highest level gaming GPU, the Titan Z[1] has 5760 cores... but it costs about $3000

[1]: http://www.geforce.com/whats-new/articles/announcing-the-gef...

If it were cheaper, there's a whole pile of applications I can imagine for it. $200 is a bit steep. Really interesting for a first gen- maybe it will lead to a range of models with different specs.

Reply on news.ycombinator.com