Slashdot Mirror

What? You don't have to lie about anything. Just use the pulldowns to select what you have and download the driver

Note that laptop GPUs are listed right there among the desktop GPUs. The driver for your M2100 even says:

QUADRO DESKTOP/QUADRO NOTEBOOK DRIVER RELEASE 367

LOL Complications?

Valve: OpenGL is faster than DirectX — even on Windows (20% faster)

Bringing Unreal Engine 4 to OpenGL

The only reason developers should consider DirectX at this point is if they need to run on an XBONE.

Funny you bring up Unreal Engine 4. UE4 is not currently at feature parity between OpenGL and DirectX11. My studio continues working with DX11 exclusively primarily for this reason. (Vulkan support will probably turn this around, but I'm assuming it's a year or two until Epic solidifies their support). We have no interest in supporting OS X, Linux, or Windows XP, so there's no good reason not to work with DX11 anyways.

LOL Complications?
Valve: OpenGL is faster than DirectX — even on Windows (20% faster)
Bringing Unreal Engine 4 to OpenGL
The only reason developers should consider DirectX at this point is if they need to run on an XBONE.

It is so god damned hard to get nVidia drivers let alone NEW drivers that its fucking pointless to try.

Oh, yeah, it sure is...

--Nvidia/Linux desktop user.

This is all very interesting. However, there is no indication of when the sticks will become generally available.

There also seems to be very little actual information about it. How much memory does it have? How many FLOPS? The product sheet says it uses 16 bit floats, which are generally good enough for NNs. But can it do FP32 and FP64 at all? The power consumption is ~1W, so I doubt if it can do much with that. The USB interface would be a major bottleneck, as you fed information in, and pulled results out. A GPU on a PCIe bus would be way faster at that ... and nearly all computers already have a GPU. I think I will continue to run my NNs on a Tesla K80.

I respectfully disagree. First of all, USB3 can pretty much guarrantee a 1 way stream of 0.5 GigaBytes/Sec. That's quite a lot of data to process, until communication becomes a bottleneck. And, you can stream your inputs in this processor, whereas you have to memcpy the whole memory to process stuff in a GPU.

It can afford to be incredibly power frugal, because it has very simple pipelines, focused on efficient vector processing, a centralised SRAM cache, and no power hungry external memories. Yes it has fewer "cores" than a GPU, but you'd be surprised how little complexity you need to perform some vector instructions quickly.

Which means, you can run quite complex NNs, on the go with pretty much just a windows/linux tablet and this stick, on a total budget of say, 10Watts. That's quite impressive.

-K

This is all very interesting. However, there is no indication of when the sticks will become generally available.

There also seems to be very little actual information about it. How much memory does it have? How many FLOPS? The product sheet says it uses 16 bit floats, which are generally good enough for NNs. But can it do FP32 and FP64 at all? The power consumption is ~1W, so I doubt if it can do much with that. The USB interface would be a major bottleneck, as you fed information in, and pulled results out. A GPU on a PCIe bus would be way faster at that ... and nearly all computers already have a GPU. I think I will continue to run my NNs on a Tesla K80.

There are only two sets of memory if you consider the register file to be memory instead of cache (which you apparently do). The problem is, the published specifications demonstrate that you are simply wrong.

4MB of L2 cache and ~14MB of register file space per GPU means that there is about 151 million bits associated with cache and "memory." On a chip with 15.3 billion transistors, that comfortably means that you have about 15 billion transistors for GPU logic.

There is everything to indicate the specs of the chip, and you've lost your bet. Now go away.

NVIDIA is in fact working on something like that. It's called NVLink and will, according to that page, enable "data sharing at rates 5 to 12 times faster than the traditional PCIe Gen3 interconnect". There's no commercial hardware yet that supports it, but it should be coming soon.

Is AMD the only company with Linux Vulkan drivers?

No, and learn to internet. Are you new? nVidia has Vulkan support for more cards than AMD does.

"I think it's a fair point. nvidia can do all they want to improve graphics performance, but that's nowhere near enough to warrant a whole new distro"

The 1990s called, and want your understanding of modern GPU architecture back. Yes. NVIDIA makes Graphics cards which do things like display an image on a monitor. Those cards also also a a whole lot more these days. For example, with CUDA one can leverage the GPU(s) for everything from Bitcoin Mining to Facial Recognition.

https://redeem.nvidia.com/

The feature list can be seen on sites like: http://www.geforce.com/hardwar...
http://www.geforce.com/hardwar...
http://www.geforce.com/hardwar...
https://developer.nvidia.com/w...
If Linux has that gpu support vs CPU bound, thats great :)

Link me to the powerpoint presentations that have those points then. I doubt they exist, they are your points and they are derived from your own misinterpretation of the information.

Here's a short summary from a Vulkan IMPLEMENTOR on where OpenGL is better than Vulkan, and vice-versa:
https://developer.nvidia.com/e...
Then if you read Graham Seller's own Powerpoint AND have experience in developing LARGE applications you can see massive potential pitfalls in using Vulkan that you won't get with OpenGL:
http://nextgenapis.realtimeren...
Of course, as an amateur you are not used to looking for indications of the limitations of an API rather than the marketing points - as doing this only comes with experience (and as I was a fanboi once but learned to look through the marketing spiel with a critical eye - a habit which it appears you have not yet developed).

If that is how you feel then use one of the various alternative solutions I already outlined for you.

Nope, OpenGL is more than sufficient - especially with the "Approaching Zero Driver Overhead" approaches and extensions. You don't seem to grok that the increased productivity of OpenGL for desktop workloads more than makes up for the few cases where Vulkan is better. For example, Vulkan produces zero benefit when rendering tessellated terrain on the GPU in my flight simulator, because the GPU is doing all the work and the use of the Vulkan command-buffer model gains nothing, but increases development cost. Hence, OpenGL is still a better option when your problem is rendering hundreds of millions of polygons per second on a discrete workstationGPU, and not submitting tens of thousands of sprites to a phone GPU with the phone's limited CPU.

But it isn't poorly designed, your specific criticism is that it has memory barriers - just like GLSL does - so when you say you have all this experience yet you don't understand the basics of asynchronous programming and don't even know GLSL very well certainly calls your claims into question.

Yes, OpenGL has that feature (so why use Vulkan?) but you don't HAVE to use that feature. This is the two-level API design I mentioned a few posts back, and why OpenGL gives you some control if you need it, plus higher-level options too - but you are not forced to do everything as if you were programming for a console with a shelf-life of a few years. Plus, on my MacPro with dual D700s the glMemoryBarrier call s not yet supported (grrrrrr crappy Apple spending effort on Metal instead).

No, I said post your objective criticism, the problem is you are so misinformed that you don't have one. You've gone to great effort to make baseless claims about things you clearly don't understand but then tell me you don't have time to post such things on the Khronos forum? No sorry that's just rubbish. If you really thought it was a waste of time and not that you're just going to look like an idiot then you wouldn't be spending all this time trying to convince me. Quite frankly when your only criticism of the design of the API is because it has memory barriers (and you think that is telling you the internal state of the GPU, which it isn't) then you have proven all this credential-dropping is untrue since you don't even have any experience with the basics of asynchronous programming and you think application code belongs in the driver. Every time you get backed into a corner you throw out the obviously false claim that you are expderienced, which you clearly are not.

Wrong. EVERYBODY says that Vulkan is more work. Everyone says that Vulkan drivers don't check for errors, which means it is easier to not only crash your application, but crash the whole damn machine - you wait until Vulkan is used in the Real World (tm). Everyo

Amen to that. Installing and updating the proprietary Nvidia drivers manually is a colossal pain, especially if you're running Ubuntu or similar and they fart out a new kernel revision every couple of weeks, requiring new kernel headers, and a rebuild of the driver. Not to mention that the proprietary driver still is missing some features that they solved in the Windows driver years ago. For example, you still can't enable TwinView (multiple monitors) and SLI at the same time, and that was a problem that Nvidia managed to fix in Windows Vista.

How about nVidia's Jetson TK1 board? It has a great 192-core Kepler GPU, a nice quad-core ARM CPU, on-board gigabit ethernet, all the ports you're likely to need and comes pre-installed with Ubuntu Linux.

I have one, and it is a great little Linux box for the most part. The x86 compatibility issue pops-up every once in a while, and 2GB RAM is a little tight, but after that it's all roses. Thanks to Raspberry Pie, ARM support for Linux is surprisingly complete, and the Tegra K1 graphics vastly outperformed the GPU on any Intel CPU that cost less than the entire JetsonTK1 board. Video cam, Google Hangouts, LibreOffice, Gimp, Inkscape, Java, Webstorm, etc. all work very nicely. My son uses it for his development computer, and it is attached to a 240GB SSD and a 32" 1920x1200 monitor. Libreoffice launches in 2s cold.

How about nVidia's Jetson TK1 board? It has a great 192-core Kepler GPU, a nice quad-core ARM CPU, on-board gigabit ethernet, all the ports you're likely to need and comes pre-installed with Ubuntu Linux.

Q: What is NVIDIA Tesla?

With the world’s first teraflop many-core processor, NVIDIA® Tesla computing solutions enable the necessary transition to energy efficient parallel computing power. With thousands of CUDA cores per processor , Tesla scales to solve the world’s most important computing challenges—quickly and accurately.

One example Tesla K40: 2880 CUDA cores. That's a LOT of cores.

Beowolf? Don't think so. If anything then maybe Spark with the map or reduce step being executed on the GPU, or better yet Tensor Flow. But as pointed out elsewhere, this chip is not even for that, especially next year when the new cards with NVlink blow away 980/Titan-X stuff of this year. No this thing is for drones, AR, or image recognition on embedded anything where power consumption and latency are the overwhelming factors. Otherwise graphics cards will outperform, or if latency is not a factor, then the whole thing can be offloaded to AWS or similar. Also, 16 bit (half-precision) floats normally bad for numerics are fine for neural networks with the bonus advantage of effectively doubling the memory bandwidth and problem size which are the current limitations.

You mean this? http://www.nvidia.com/object/f...

FreeBSD. And it is growing. Admittedly, from a VERY small share, but...

Get me an up-to-date nVidia driver, and support for vmware, and I'll switch all my systems right now. Cold day in hell, you say? That's about when I'll go BSD, then.

Well, I guess you will be reinstalling for a while... VMware since FreeBSD8 and current Nvidia drivers. http://www.nvidia.com/object/f... PC-BSD is a little easier for a Desktop then pure FreeBSD.

Also, nVidia has drivers "up-to-date" as of about three weeks ago.

http://www.nvidia.com/object/f...

So, not sure what your issues are. Can you perhaps restate them?

It's not signed code per-se, but nvidia does provide checksums from the website:

cuda_7.5.18_linux.run (md5sum: b22ef6bc073f7cf767f547a84fb0e3c2)

(see https://developer.nvidia.com/c... for more versions)

If your unfamiliar with how to use a checksum, I suggest reading http://lifehacker.com/247262/h...). Basically though, if they don't match, don't trust it.

I haven't tried running craps on my Nvidia video card yet. That should be a hell of a lot faster.

https://developer.nvidia.com/how-to-cuda-python

You probable do not know what is in your smart phone at all. http://www.nvidia.com/object/t... http://www.nvidia.com/object/t... Nor should you when it just works.

You probable do not know what is in your smart phone at all. http://www.nvidia.com/object/t... http://www.nvidia.com/object/t... Nor should you when it just works.

Yep. No console can drive my 4k TV.

http://shield.nvidia.com/andro...

Depends on what you mean by "faster." If you mean clock frequency, then perhaps. Also perhaps if you mean an individual core of a CPU vs a core of a GPU.

In this sense, it's the time to perform massively parallel instructions. GPUs are generally hundreds of times faster than CPUs for such calculations. Part of this is because a CPU can have a few cores, but a GPU generally has thousands of floating point units. The other part is that CPUs are general purpose central processors while GPUs are very specialized to optimize them for specific kinds of tasks.

Think of it like a CPU is 4 guys with Swiss Army Knives while a GPU is a team of 1,600 guys each with a battery powered, professional screwdriver. Guess which one's faster at screwing 1,600 wood screws into 400 posts for a building. Now guess which is faster at cutting a traced outline on a single piece of paper.

http://www.nvidia.com/object/w...

You see, if you (the one who posted the question) were a numerical mathematician or a computational physicist and looking for adequate performance in a research setting at rock-bottom cost, I'd say:have a look at GPU's (see e.g. here http://www.nvidia.com/object/c... ) and e.g. the Navier Stokes solver from Stanford U. (see here: http://mc.stanford.edu/cgi-bin... ). For such applications, hardware based on GPU's tends to give a much better price/performance ratio than rigs based on CPU's. But they're s lot harder to use as well. So find a suitable solver and see what hardware makes it shine.

But you probably aren't, or you'd have known that already (or looked it up in the literature or figured it out yourself).

Err ... if you are (as far as I can make out) just a computing guy who doesn't know Navier from Stokes but wants to put a "FEA/CFD rig" together, I think you're simply not the right person to do that. Yes, you're probably capable a few PCB's together that can run a generic Matlab-based solver, and will then find that it gives you an abysmal price-performance ratio on your particular workload.

And why? Well, the problem with "supercomputers" is: they're much more powerful than a general purpose computer only on very *specific^ problems. Change the problem and watch the performance change as well.

So you've got to tune your hardware to your problem in order to get realy good price-performance. And your "problem" is your solver. You need to choose that as well. And for that you need yo understand a bit about what a solver does relative to the problem you really want to solve. And it sounds as if you haven't a clue. Sorry.

The alternative is to let other people (consultants, vendors) do the thinking, and buy a custom solution. That will work, and will give you reasonable (but not great) price-performance ratios at a reasonable price level. But make darned sure that your hardware-software combinations is a good fit.

My suggestion: talk to your team member who knows what the formulas look like, what solver you're going to be using, whether low-accuracy is acceptable (e.g. if the objective is to obtain a graphical solution) or whether high accuracy is a must (e.g. for engineering purposes).

If low accuracy is acceptable but you want the best speed, think GPU's. If not ... determine what the particularities of your problem instances are, what your solver grid will look like, and what type of computing resources you'd need for that and how much.

Simply saying "a CFD problem" isn't nearly specific enough. And getting to a hardware configuration that has truly good price/performance levels is something for a specialist (or a team effort).

This is Nvidia. Don't buy into this based on any promises of what is to come, no matter how reasonable they seem.

All that was promised is right here, right now: https://developer.nvidia.com/d...

From TFA there was no pic of the UI, nor any mention of tech specs aside from a lot of nebulous details. From nVidia's website ...

* https://developer.nvidia.com/d...

DIGITS DevBox includes:

* Four TITAN X GPUs with 12GB of memory per GPU
* 64GB DDR4
* Asus X99-E WS workstation class motherboard with 4-way PCI-E Gen3 x16 support
* Core i7-5930K 6 Core 3.5GHz desktop processor
* Three 3TB SATA 6Gb 3.5â Enterprise Hard Drive in RAID5
* 512GB PCI-E M.2 SSD cache for RAID
* 250GB SATA 6Gb Internal SSD
* 1600W Power Supply Unit
* Ubuntu 14.04
* NVIDIA-qualified driver
* NVIDIA® CUDA® Toolkit 7.0
* NVIDIA® DIGITSâ SW
* Caffe, Theano, Torch, BIDMach

.. holy crap is that a lot of GPU horsepower "just" for AI. Oh look, they are running Ubuntu :-)

They are really trying to get people on board about how much better / faster their GPU solutions are ...

* http://www.nvidia.com/object/m...

The problem is that there are lot of "niche" use cases. If your problem domain maps to the GPU then yeah, mjaor speedup. If not, well, then you're SOL running on "slow" CPUs.

From TFA there was no pic of the UI, nor any mention of tech specs aside from a lot of nebulous details. From nVidia's website ...

* https://developer.nvidia.com/d...

DIGITS DevBox includes:

* Four TITAN X GPUs with 12GB of memory per GPU
* 64GB DDR4
* Asus X99-E WS workstation class motherboard with 4-way PCI-E Gen3 x16 support
* Core i7-5930K 6 Core 3.5GHz desktop processor
* Three 3TB SATA 6Gb 3.5â Enterprise Hard Drive in RAID5
* 512GB PCI-E M.2 SSD cache for RAID
* 250GB SATA 6Gb Internal SSD
* 1600W Power Supply Unit
* Ubuntu 14.04
* NVIDIA-qualified driver
* NVIDIA® CUDA® Toolkit 7.0
* NVIDIA® DIGITSâ SW
* Caffe, Theano, Torch, BIDMach

.. holy crap is that a lot of GPU horsepower "just" for AI. Oh look, they are running Ubuntu :-)

They are really trying to get people on board about how much better / faster their GPU solutions are ...

* http://www.nvidia.com/object/m...

The problem is that there are lot of "niche" use cases. If your problem domain maps to the GPU then yeah, mjaor speedup. If not, well, then you're SOL running on "slow" CPUs.

Not so.
There are more devices out there than you think:
http://www.nvidia.com/object/t...

My favorite thing in this discussion would be Nvidia's rendering of a moon landing, and their diatribe:

http://www.nvidia.com/coolstuf...

My car now has Nvidia chips that recognize speed limit signs and displays them inside the speedometer (along with a reminder when I exceed the speed limit). For the future, Nvidia has announced the NVIDIA’s DRIVE PX self-driving car computer which has a lot of advanced image processing.
http://blogs.nvidia.com/blog/2...

The 2015 GPU Tech Conference was stuffed full of this tech.
http://www.gputechconf.com/

but the chances are it's not actually holographic.

You can be certain about it. There is no real-time holographic display as of today. For LFDs (Light Field Display), NVidia had a prototype a few years back and it is reasonable to think that Magic Leap is pursuing something similar. Yet, I don't think the technology is mature enough to be able to generate dense light fields needed for high quality scene rendering.

On OS-X it is all Apple's way, all the time. You gets the drivers you gets from Apple and live with it.

This is actually less true now - Nvidia is publishing their own driver packages for OS X because they are tired of Apple shipping ancient versions whenever they get around to including them in a point release.

They are labeled for Quadro, but they work just fine with GeForce. I'm running a Geforce GTX 780 Ti in my Mac Pro completely unmodified - all I don't get is the uEFI boot screens. Once the kext loads, everything is perfect.

You can get a fully automated drone for hundreds of dollars, not millions. Not big enough to put a person in, but still fully automated. Google and Amazon are looking to deploy huge fleets of fully automated drone delivery aircraft costing a few thousand each.

We already know that autonomous vehicle technology isn't prohibitively expensive. Nvidia already has their system on the market for carmakers to integrate. It adds thousands, not millions. And depending on how things pan out, you might recoup all of that initial outlay in insurance savings pretty quickly.

You say that as if this isn't already a solved problem. There are loads of autonomous vehicles already successfully navigating the public roads in general traffic. This technology is so far along that companies like Nvidia have off-the-shelf autonomous car kits on the market, ready for carmakers to integrate into their vehicles.

Your rant sounds like the guy in 1906 saying that travelling faster than 35 mph was impossible as the Stanley Steamer screams past at 120 mph in the background.

You're describing 'TurboCache' (a marketing name if ever there was one).

It wasn't a secret, it was only on very low end cards, and ATI was already doing the same with 'HyperMemory'. Intel, for their part, was exclusively using system RAM at the time (and largely still is).

So what graphics *have* you been buying for the last decade?

Not really, the nvidia driver uses autotools to install.

instructions from http://www.nvidia.com/object/l...
$ tar xvzf NVIDIA_kernel.tar.gz
$ tar xvzf NVIDIA_GLX.tar.gz
$ cd NVIDIA_kernel
$ make install
$ cd ../NVIDIA_GLX
$ make install

It's the GUI tools the distro came up to manage the install which do the anoying are you sure, and click click.

If that whole process takes 3 seconds (which would be amazing) then your computer only performed 1 "operation per second". But computers don't perform "operations" they have to perform millions of sub-actions to accomplish your goal.It would be like saying that "Rendering a game's frame is only a single task so it would be a very serial task without any potential for multithreading." when in reality "rendering a frame" is a massively parallel task of rasterizing millions of triangles (or intersecting rays) and sampling textures, computing lighting values and performing table look ups.

Take interpreting voice. By applying multiple models simultaneously you can get better results. Seems pretty obvious.
http://devblogs.nvidia.com/par...

For the flyer maybe it'll generate 1,000 flyers simultaneously and then compare them to award winning graphic design projects to see which of the 1,000 ideas it had matches historical good ideas.

192 cuda cores, doesn't use much power, so little heat to dissapate.

https://developer.nvidia.com/j...

Or, maybe re-write your application to use an fpga?

If you are using the gpu for display, then just do the compute elsewhere, and use a low power box, with a simple frame buffer, that you can seal up.

How inexpensive?

* $200 Nexus 7
* $299 nVidia Shield Tablet http://shield.nvidia.com/gamin...
* $350+ Nexus 10

Some tech specs comparisons ...

* http://gadgets.ndtv.com/nvidia...
* http://versus.com/en/nvidia-sh...

He's talking about the controller that comes with the SHIELD Tablet, not the SHIELD Portable. Still, I don't see what's wrong with it requiring an Nvidia GPU, since it is an Nvidia product.

Previously, Nvidia said that it would license it's Kepler GPU cores to third parties. Semiaccurate maintains that this licensing program was in fact bogus and was conceived purely to justify future patent trolling activities. Semiaccurate also claims that
Nvidia tried to "shakedown" Apple with the same patents and Apple subsequently gave the contract for the Mac Pro GPU to AMD as punishment.

Sorry to be pedantic, but there is actually a real difference with regard to remote desktop, which operates using the RDP protocol and must operate under a non-accelerated graphics driver, vs. Nvidia's GRID offerings. Those GPUs have access to the same H.264 encoding capability as all Kepler-series chips as well as drivers which allow them to be shared seamlessly by multiple virtual machine instances, meaning the performance is MUCH better than most previous solutions. You can check out a demo here. Now of course a significant portion of the gamer market would still find the input lag acceptable even if we could stream uncompressed 4k (though that is being worked on), but companies that do CAD and/or 3D modeling may find that the TCO of a couple of GRID servers is lower than going through a full hardware refresh cycle every few years while offering very comparable performance, particularly if the servers are on your LAN.

There is currently no published dev tools for OpenCL for the tegrak1. If you want to do development for a *specific* device that has a k1 (such as the shield tablet) you need to use the tegra android development pack, and guess what, it does NOT support CL yet. The OP is reporting on something that is not even out yet.

https://developer.nvidia.com/t...

Since google has not put their weight behind OpenCL any CL development you do for android will be very device specific. So the fragmentation issue that Apple bashes Android for is very real if you want to use CL...

Google should do something (and soon) because Apple's metal API supports GPU compute workloads. Sure you can do GPGPU using OpenGL ES but that is more limited in scope than a full CL implementation...

The Nexus 10 I purchased on launch day had a working OpenCL implementation. I ran some kernels on it and it was definitely GPU accelerated. A software update actually removed the CL driver later on as Google backtracked on CL support and began promoting their Renderscript instead.

http://www.anandtech.com/show/...

Also, the nVidia jetson-tk1 that I purchased does NOT have a working OpenCL implementation.

Look at the comments from the nvidia employee: http://devblogs.nvidia.com/par...

This article is just free advertising for nVidia, and its false information too!

They make great hardware, why do they have to be so damn dishonest all the time?

While there are some fancy light field displays that might be able to adjust for vision defects in software, those are still years after. However the Oculus Rift has swappable lenses, so it shouldn't be to hard to design some lenses that correct whatever vision defect you might have. The consumer version will probably have some adjustable optics to correct for vision issues, at least thats how the first wave of consumer VR headsets back in 1995 worked.

Not sure if I get what you are asking. But I'll try:

AMD: http://developer.amd.com/resou...
NVidia: https://developer.nvidia.com/n...
https://developer.nvidia.com/o...