Slashdot Mirror
NVIDIA Releases JTX1 ARM Board That Competes With Intel's Skylake i7-6700K (phoronix.com)
An anonymous reader writes: NVIDIA has unveiled the Jetson TX1 development board powered by their Tegra X1 SoC. The Jetson TX1 has a Maxwell GPU capable of 1 TFLOP/s, four 64-bit ARM A57 processors, 4GB of RAM, and 16GB of onboard storage. NVIDIA isn't yet allowing media to publish benchmarks, but the company's reported figures show the graphics and deep learning performance to be comparable to an Intel Core i7-6700K while scoring multiple times better on performance-per-Watt. This development board costs $599 (or $299 for the educational version) and consumes less than 10 Watts.
The "deep learning" benchmark is a GPGPU workload which does practically nothing on CPU.
Nvidia has just made a SoC Chip that has about equally fast iGPU than what Intel has, for a lower energy consumption.
But in CPU performance, the Skylake is MUCH faster.
I guess my question is, what could/would I do with one as a layman with a passing (but growing) interest? Would this be a pricey replacement for a RPi or maybe a controller hub type of thing for a collection of RPis? I do have a project in mind to finally make use of these things - I've even got a half dozen of the RPi still sitting in their boxes (except for one that I opened and poked at) but I'm not exactly sure where to begin. Well, I know where I will begin - I'm just not sure that I should begin there. It's a long story...
"So long and thanks for all the fish."
Meh, they matched the GPU performance of GT2 for twice the $, now let's compare it to gt4 with 128MB of eDRAM...
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
This is just a particular benchmark that happens to run entirely in the GPU.
Just because its low power does not means it have the same performance.
In performance per watt, Intel and ARM are mostly the same .
And I'd like to see actual benchmarks, not "We used CUDA based benchmarks that are designed to run well only on Nvidia GPUs!" As a benchmark, as last I looked Intel had the best performance per watt GPUs around.
The article is silly. Who would buy a i7-6700K purely for the GPU. If you want that kinda gpu power you can get a dedicated graphics card for much less.
Tegra X1 is an embedded chip. What NVIDIA claim it is designed to do is basically make a self-driving car out of it. For this purpose the GPGPU capability would actually be important and also Skylake would not meet as Intel likely don't offer them in industrial/automotive temperature ranges.
In reality the best thing it can do might be a digital signage or laggy infotainment system, but in that ground it should perform better than its competitors.
The problem with this is that the availability of GPGPU code for embedded systems really suck. This while intel makes a lot of effort to optimise Linux, Java and other commonly used software to their hardware.
Freedom of speech? How can a company "allow" or "disallow" journalists to publish benchmarks? Do they have to sign an NDA?
Yeah, four ARM cores can match an i7-6700k in nVidia's fucking dreams. This is pure trolling. For the desktop systems we tend to care about, nobody is even using the Skylake's onboard GPU. That's targeted to allow server systems to expose GPGPU while also having top-of-the-line Intel CPU performance for every other workload. This doesn't compete on that front either. Finding an actual configuration and workload for which this SoC actually competes with the i7-6700k is nigh-on impossible.
what could/would I do with one as a layman with a passing (but growing) interest?
You could buy one and leave it in the box, then post vague questions on Slashdot that don't give any hint as to what your project actually is :p
For some parallel tasks it could be cost effective. A TFLOP of GPU with only 10 watts is nothing to sneer at. It might even be lower watt/flop then an FPGA, which tend to be power hogs. Of course, the 10 watt figure is for the card form factor SOC only, so the power and size is greater for the SOC plugged into the carrier board. And the cost needs to come down quite a bit for their likely market place. Either the price falls by a huge amount or it goes nowhere.
Even so, this could be interesting for some niche markets.
Why is Snark Required?
Until you see the $1495 pricetag!
And I'd like to see actual benchmarks, not "We used CUDA based benchmarks that are designed to run well only on Nvidia GPUs!" As a benchmark, as last I looked Intel had the best performance per watt GPUs around.
And I'd like to see actual benchmarks, not "We used CUDA based benchmarks that are designed to run well only on Nvidia GPUs!" As a benchmark, as last I looked Intel had the best performance per watt GPUs around.
Of course they use benchmarks that run well on CUDA. Some algorithms can't be parallelized effectively over hundreds of GPU cores. Other algorithms can take a hit due to the branching required. However, there are some real world applications that can be effectively parallelized on CUDA that really make sense.
Theres no point in comparing algorithms poorly suited for GPUs. NVidia might as well throw in the towel now for those applications. However theres a reason why OpenCV contains so many CUDA implementations of algorithms that have already been written for CPUs. I guarantee it's not because programmers get off on writing CUDA versions(although it's possible some do). It's because these CUDA versions actually provide speedups.
Given that the X1 can be used in embedded systems, you must understand the architecture and your algorithm to decide if the X1 is well suited for your application.
I wouldn't think 10 watts produces much heat...
To compare with intel boards, we also need to check which drivers nvidia provide and if there are open source.
being equally fast as intels graphics is like crowing about beating a legless man in a foot race.
So it doesn't run a mainline Linux kernel? Or does someone know otherwise. I couldn't find anything on the nvidia web site. Nor do I see how to buy one at the educational price.
Well... It's a bit long but I'll try to make it short.
I want to install a gate at the end of my driveway.
Yes, they could just walk around the gate or even drive through this non-existent gate (it's not there yet).
I want a sensor that will trigger a quiet alarm and/or send a push notification to my phone.
I want this gate to open for certain people's cars - I can put an RFID chip in a box with a magnet - I hope. (This is where I'd start - I know nothing about RFID)
I want to those people to have the gate automatically open (and send me a notification).
I want to be able to send a message to open the gate from afar - via cell.
I want to be able to view the gate by IP connected camera.
I should also be able to speak to the people at the gate.
I should have a "leave me alone" button that keeps the gate from opening unless I opt to open it (and a switch on the alarm to shut it off).
I should probably put a timer on it so that it needs a manual override to open - just in case someone finds it online and I've a security flaw. A flailing gate would be funny but only for a little while.
I'd like to be able to configure the gate to open automatically for certain cars at certain times.
Now, some of this, I know how to do. I'm thinking a Pi might not have enough oomf to run the software needed at the gate itself. I'm curious if this would be a better spot to start than to try it on a Pi.
Also, I know nothing about the RFID but I do have some (bad) programming skills and think I can make the rest of this work - even if I've never done it before. The worst it will do is keep me amused and occupied for a while - which is probably a good thing. I'm probably missing a few things but that's the gist of it. I can do most of this, not quickly and not well. It's not at all about security, it's just to geek out for a while. You could, quite literally, drive around any sensible gate that I could put there.
KGIII - Posted AC 'cause it's that time of night again where it says resource not available. :/ Well, morning.
This thing is for when you're doing something that can benefit from GPGPU, and a R-Pi isn't providing enough CPU power. The obvious example is machine vision, and I'm pretty sure that's the prime example that nVidia actually gave when announcing the thing: robotics. It's got a tiny little power footprint, which is the advantage over something from intel.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
http://techgage.com/article/intels-skylake-core-i7-6700k-a-performance-look/
http://www.anandtech.com/show/...
I am confused - none of the benchmark articles above mentioned anything about the impressive "GPU power" of Core i7-6700K from intel
Could you kindly provide us with proofs that Intel's Core i7-6700K comes with impressive native GPU performances?
Thanks !
If nVidia is truly serious about this thing they can put it out with a loss-leading price point at $99 to the marketplace
That will sure create a typhoon like rush in the developer community and within a few months lots of very interesting result will be posted all over the Net
Unfortunately, ,many Chinese owned companies such as VIA and nVidia are never known to cater to the developer communities ...
Think back to https://en.wikipedia.org/wiki/... like ideas. Can the math be spread over a lot of cores, new gpu's and then work out quick, better, sooner, with less heat?
If yes, great. If no, buy into a different CPU for the calculations.
Domestic spying is now "Benign Information Gathering"
Thanks. I'm logged into my VPN and it's happily eating JS and won't let me post when logged in. No, I do not know why. It has been doing this daily for about two weeks now. *sighs*
Just above your post, I posted what I'm going to attempt. I talked it over a little with another /. user and I'm gonna go for it. I'll probably work on it over the winter and start the project rollout in the spring. You can take a gander at that reply, if you'd like. I think it will make an excellent "starter" project. I'm not really a beginner so to speak but I'll approach it like one. First, I must learn about RFID chips. :/ I think they'll be my best solution but that's another topic for another day.
Also, I should sleep at some point.
KGIII (AC 'cause something is broken and I'm too lazy to fix it.)
these only have value to a subset of computations and skew the comparison with the cpu or even the bigger gpu brethren.
The nVidia part is designed to do lots - and I mean lots - of SIMD or MIMD instructions very quickly, similar to a GPU/DSP, and also be a CPU. Got a multi-layered hidden 1K wide markov model that does 30-60 fps synthetic viewpoint stereoscopic imaging at 1280*1024*32bit pixels? Or sensor fusion combining SWIR, LIDAR, and normal vision while compensating for turbulence and contaminants? The Intel part is the other way around - designed to be a CPU with some DSP/GPU capability built in. The two parts are about the same price in quantity, which is why they were compared. The Intel part is an absolute corgi or pomeranian as a number cruncher, but is okay as a general CPU. The nVidia part is a greyhound on amphetamines as a number cruncher, but okay as a general CPU. Different niche markets, chuckles.
Oh, and Freedom of Speech only applies to the government. Companies can quite legally say "if you publish this before it is allowed, we will never give you another story lead or review freebie again."
At 1 TFLOP for 10 watts, I want one. And they are only around $700. $1495 if you want a corporate development support.
Dude, an Arduino Uno has enough "oomf" to do all that nonsense for crying out loud.
being equally fast as intels graphics is like crowing about beating a legless man in a foot race.
The only ones you'll hear complaining about Intel's built-in graphics are the PC gamers and benchmarking sites. I'm actually quite happy downgrading from a Core i3-3227U to a Pentium N3700.
The new A9X in the new iPad leaves the X1 in the dust. The A9X scores 80 in Manhatten test, while X1 only scores 65
I'm assuming SteamOS and the games it supports would not run on this unless everything was compiled for ARM, yes/no?
I had to struggle to think of applications for this board, and then for every single one of them, this board immediately drops to the bottom of the list the moment I look at how much it costs. "Anything you can do, I can do faster and cheaper," I say to this thing.
yep. my exact thoughts when i looked at the article.
it's either entirely cpu on the skylake and/or using the crappy igpu...
likely pairing it with a mid, maybe even less dgou wiuld blow that arm out of the water as nvidia is still using the shitastic arm designed blocks, and i doubt that nvidia will be anle to help much given what qualcomm and apple manage inhouse.
My Timex wristwatch has enough oomf to do all of that shit.
My Timex Sinclair ZX-81 has enough oomf...
The Jetson TK1 sold for $192.
I was really looking forward to a Tegra X1 version of the Jetson, but not at $599 and not at 6+ months after the chipset started appearing in consumer products at a significantly lower price.
(The Jetson TK1 was the first K1 device to launch and was priced similar to or below fully assembled consumer products like the SHIELD Tablet.)
retrorocket.o not found, launch anyway?
In a race to the feet, the legless man always wins. And runs Linux while running Crysis in a Wine while in a Beowulf cluster of itself.
My virtual 8-bit CPU in my Minecraft world has enough oomf...
I'd like to see Nvidia get into making x86 CPUs. This could be a three horse race.
Of course it does. I'd imagine a wristwatch has something along the lines of an Epson S1C88349 CPU running at a few KHz. By comparison, your ZX-81 is a supercomputer.
> The faster they become more like x86, the better.
"Jane! Get me off this crazy architecture!!"
This kind of thing is basically a toy. Stick a bunch of them in a cabinet and build your own bitcoin miner-like thing. You're not going to run games on it.
During my undergrad program we had sixteen teams that used the Jetson TK1, and in every single case they had to switch to a more capable platform. The Jetson TK1 has extremely limited GPIO, poor documentation, buggy drivers, and greatly reduced functionality compared to what's on the spec sheet.
On paper it the TK1 was a great product, in practice it was worthless. NVidia never put much effort into solving these problems and expected the TK1 user base to do all the debugging and fixing, but given the complexity of the board that never happened.
I'd wait a few years to see if the TX1 is ready for prime-time or not. NVidia seems to make a practice of shipping out hardware with incomplete software and limited documentation which means you'll spend more time debugging their system than developing your own applications.
Except for the "Speaking' and "Camera" parts
The dangers of excessive individualism are nothing compared to the oppressiveness of excessive collectivism
Who the hell cares? Seriously, it's a graphics card!
The dangers of excessive individualism are nothing compared to the oppressiveness of excessive collectivism
or in other words, the only ones you DON'T hear complaining about intel graphics are those that don't actually use them.
I think I might get one, then. Thanks. This would be an area where there some maths - I posted as an AC earlier. My VPN is still being screwy so I just logged out.
It'll give me an excuse to brush up on my C and learn about the whole RFID methods. I've been meaning to do both for a while now. If you're curious or inclined to opine the AC post is above. I identify myself.
"So long and thanks for all the fish."
That's what I'm thinking. I need something that can push and, maybe, compress video and sound. It's probably also going to have storage attached to record something like snapshots at 3 second intervals or the likes. I don't want to "make do" with something. I want to just make it, learn about it, and forget about it - until I need to repair or update and realize that I should have followed good documentation methods. Then, I'll learn it, fix it, and forget about it! Seems pretty good to me.
"So long and thanks for all the fish."
Hold on guys no benchmarks yet nvidia is still paying out kickbacks for good results lol.
WTF are you babbling about? Speech and video don't even require a computer, just the gate control.
On Intel, the Caffe benchmark runs optimized for CPU. Caffe framework also has GPGPU backend for GPUs
The benchmark was image classification with Caffe AlexNet https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet
Caffe deep learning framework is optimized for CPU on Intel and has GPGPU backend for NVIDIA architectures
The benchmark was image classification with Caffe AlexNet https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet
Caffe deep learning framework uses CPU backend on Intel and has GPGPU backend for NVIDIA architecture. The results are comparing overall throughput and power (number of images classified per second with AlexNet model)
Keep the RPi for the software, it's a microprocessor, and it has enough oomph for all your software. Use an arduino to control the gate, it's a microcontroller, and should have enough oomph to drive tour gates motor. Use the right tool for the job. No real sense trying to remove a screw with a hammer.