NVIDIA Quadro M6000 12GB Maxwell Workstation Graphics Tested Showing Solid Gains

MojoKid writes: NVIDIA's Maxwell GPU architecture has has been well-received in the gaming world, thanks to cards like the GeForce GTX Titan X and the GeForce GTX 980. NVIDIA recently took time to bring that same Maxwell goodness over the workstation market as well and the result is the new Quadro M6000, NVIDIA's new highest-end workstation platform. Like the Titan X, the M6000 is based on the full-fat version of the Maxwell GPU, the G200. Also, like the GeForce GTX Titan X, the Quadro M6000 has 12GB of GDDR5, 3072 GPU cores, 192 texture units (TMUs), and 96 render outputs (ROPs). NVIDIA has said that the M6000 will beat out their previous gen Quadro K6000 in a significant way in pro workstation applications as well as GPGPU or rendering and encoding applications that can be GPU-accelerated. One thing that's changed with the launch of the M6000 is that AMD no longer trades shots with NVIDIA for the top pro graphics performance spot. Last time around, there were some benchmarks that still favored team red. Now, the NVIDIA Quadro M6000 puts up pretty much a clean sweep.

Comments

No use

[maestroX]

Doesn't even have a D-sub connector!

Re:No use

[Anonymous Coward]

How did you manage a score of two? Is my sarcasm detector on the blink?

It has a DVI-I output, so a DVI-I to D-Sub adapter could be used.

Re:No use

[Anonymous Coward]

Neither does any monitor made in the last 10 years that didn't have it for legacy compatibility.

Good luck getting a 4K 60p video stream (530MHz pixel clock) down an analog cable that isn't 3BNC.

Re:No use

[sexconker]

Cables aren't analog or digital. They're cables.

Re:No use

[Anonymous Coward]

True, but a BNC cable and connector is designed mainly for analog applications.

Re:No use

[maestroX]

How did you manage a score of two?

I created an account over 10-15 years ago, treated it with utmost care and love, only now and then jesting, but very cautiously and carefully.
With the occasional post, my precious karma went silently up and up, and up and up.
Until today, I felt the longing for OMG frist post, succumbed and finally blew it all on an Nvidia topic.

Perhaps on topic, what useful task would this card capabilities be insufficient for? (useful in the sense with video output)

Re:No use

[Anonymous Coward]

Thus you say the signal going through said cable is analog?

Re:No use

[Anonymous Coward]

Good answer.

Re:No use

[Anonymous Coward]

Cable = Coaxial
Connector = BNC

Re:No use

[Anonymous Coward]

Thus you say the signal going through said cable is analog?

Yep, sinewave.

Re:No use

[silas_moeckel]

BNC are only for analog you say? SDI disagree's with you, as does CXP. Hell 4x CXP6 which goes over 4 BNC terminated coax cables is pretty close bandwidth wise to display port but over a far greater distance.

Re:No use

[Blaskowicz]

Some people do use 2560x1440 over VGA (at a reduced refresh), that's a workaround that can be used on some Intel graphics.

Re:No use

[dale.furno]

tell that to Monster

Re:No use

[Anne Thwacks]

Get off my lawn. Or, you might care to investigate how Ethernet cables used to be fat and yellow and coaxial.

Re:No use

[Half-pint HAL]

tell that to Monster

..but make sure that the wires in your phone are they right way round, or the signal might not be clear enough for them to understand what you're saying.

Re:No use

[rgbatduke]

I like to make sure that the wires are very slightly tipped when I talk so that the electrons go downhill into my phone. Otherwise the wire gets clogged.

Isn't that why the phone lines run overhead? Otherwise the people on the second floor couldn't use the phone!

rgb

Re:No use

[Anonymous Coward]

BNC are only for analog you say?

Nope. I said that BNC is designed mainly for analog applications. Which is true.

Re:No use

[Anonymous Coward]

seriously, ultimately all signalling is analog, and treating the waveforms as digital (binary or otherwise) is a convenient abstraction. an abstraction that gets increasingly difficult to maintain as signalling rate goes up, or bandwidth goes down, etc.

Re:No use

[Anonymous Coward]

on the "blink"?

Re:No use

[drkim]

what useful task would this card capabilities be insufficient for? (useful in the sense with video output)

Video editing, image processing, or VFX.

Avid and Premier, Photoshop, and After Effects can use the CUDA processing power of this card.

SILVER HAMMER HAS ARRIVED

[Anonymous Coward]

Cower AMD Cower. May as well make trinkets now.

Too bad

[Anonymous Coward]

its too bad there is no double precision performance to speak of on these newer cards lately. good for games, not much else.

Re:Too bad

[sexconker]

nVidia doesn't care about double precision. The benchmark suites don't, either.

But researchers do, and that (and price/performance) makes them go AMD.

Re:Too bad

[Anonymous Coward]

No they don't. They go nvidia, wrong or right.

Re:Too bad

[HetMes]

If your algorithm is unstable at single precision floating point, it's going to be unstable at double precision as well.

I think there's only very small niche of applications for GPGPU cards where double precision is absolutely necessary.

Re:Too bad

[Anonymous Coward]

If your algorithm is unstable at single precision floating point, it's going to be unstable at double precision as well.

Do you even know what you're talking about? error propagation, does that ring a bell? I'll give you a hint, if your computation requires a large number of operations then the absolute magnitude of machine rounding errors is critical. And, lest you think this is a 'very small niche', any matrix multiplication has O(n) operation per element. Chain a few those (say, in a Markov chain type of random walk) and it's an exponential growth. Using double instead of single is like being able to do periodic (expensive) full recomputations to control stability of a fast-updating chain instead of barely having enough precision when doing the full recomputations. Fast and stable versus extremely slow and perhaps (depending on today's $DEITY's mood) barely stable.

To put it differently, by the time error accumulation in double precision leaves you with a single-precision-worth of valid digits, single precision error accumulation has long ago made the computed value completely meaningless.

Re:Too bad

[Anonymous Coward]

Yea, they do go Nvidia is the unfortunate reality but it's somewhat understood if you look at all the extra capabilities Nvidia's architecture exposes - this becomes clear once you've really soaked in on AMD and OpenCL. Also AMD flubbed up on double precision after the 7990s. No one gives a crap about double precision for the foreseeable future except all the researchers and engineers programming the darn things. Haha. Wait, why is no one laughing?

Re:Too bad

[Anonymous Coward]

This. Seriously. And this stuff is everywhere. Very few things only require some small number of accumulations or have no recursive uses to a system.

Re:Too bad

[Anonymous Coward]

You haven't done any computational research, have you?

Re:Too bad

[Anonymous Coward]

Single precision is good enough for training deep neural networks. Actually those network might even run faster (due to memory bandwidth limitations) if CUDA would support float16 (half precision floats).

Re:Too bad

[rgbatduke]

To paraphrase, you can't be too rich, too thin, or have too many bits of precision in a calculation. With single precision you have to be enormously careful not to drop digits even in comparatively modest loops; with double precision you can many digits before you run out. You can see it in almost any computations involving trig and pi -- single precision pi degrades in series much faster than double precision pi. It isn't just a matter of not using forward recursion to evaluate bessel functions, which is unstable in any precision (or for that matter, using book definitions of e.g. spherical bessel functions in terms of trig functions) or reordering series to avoid subtracting big numbers and running small to big instead of big to small -- there is simply a big difference between cumulating a random walk with a random digit at the 16th place and one at the 8th place.

A second problem is the exponent. 10^38 just isn't "big" in a modern large scale computation. It is easy to overflow or underflow a single precision computation. 10^308 is a whole lot closer to big, even expressed in decibels. One can concentrate a lot more on writing simple code, and a lot less on handling exponent problems as they emerge.

A final problem is random numbers. This is actually a rather big problem, as lots of code (all Monte Carlo, for example) relies on a stream of algorithmically random numbers that (for example) do not have a period less than the duration of the computation and that do not have significant bunching on low dimensional hyperplanes or other occult correlations. It is much more difficult to build a good random number generator on fewer bits, because the periods of the discretized iterated maps scale (badly) with reduced numbers of bits and it is more difficult to find acceptable moduli for various classes of generators from the significantly smaller discretized space. You can watch this problem emerge quite trivially by building a Mandelbrot set generator in float and rubberbanding in -- oops, you hit bottom rather quickly! Rebuild it in double and you at least have to work to rubberband in to where it all goes flat. You have to build it in a dynamically rescaleable precision to rubberband in "indefinitely" as the details you wish to resolve eventually become smaller than any given finite precision. This actually illustrates the overall problem with single precision quite nicely -- the emergent flat patches in an graphical representation of an iterated map are isomorphic to the establishment of unintended correlations in long runs of iterated maps in a random number generator and the clipping of the graphical representation of small numbers illustrates the problems with mere underflow in real computations of interest.

Personally, I dream of default quad precision and 128 bit processors. 34 decimal digits of precision means that a random walk with n unit steps (which accumulates like \sqrt{n}) require (10^30)^2 = 10^60 steps to get to where I don't still have 4 significant digits. Even a rather large cluster running a rather long time would have a hard time generating 10^60 add operations. In contrast, with only (say) 8 decimal digits a mere 10^16 operations leaves you with no digits at all, assuming you haven't overflowed already. I've run computations with a lot more than this number of operations. I also like the idea of having overflow around 10^5000. It takes quite a while adding numbers at the overflow of double precision to hit overflow, and one basically could add overflow scale single precision floats forever and never reach it. That gives me comfort. It would also make writing a Mandelbrot set explorer tool where one would be likely to give up before rubber banding all the way to the "bottom" -- there are a whole lot of halvings of scale in there to play with that still leave you with much more resolution than needed on the screen.

rgb

Re:Too bad

[Anonymous Coward]

Nvidia has clearly started to segment their product line on a deeper level to have even larger margins on Tesla.

Re:Too bad

[Anonymous Coward]

With the advent of full-scale planetary systems in games, double precision would be a very handy thing to have even for gaming. Such systems are being implemented in single precision today, but having good double precision support would make the job that much easier on the engineers.

Re:Too bad

[mrchaotica]

With single precision you have to be enormously careful not to drop digits even in comparatively modest loops; with double precision you can many digits before you run out.

Ha! I see what you did there. Unfortunately, your verb wasn't double-precision.

Re: Too bad

[Anonymous Coward]

AMD GPU accelerators have a miniscule percentage of the HPC and data analysis market share. Researchers use NVIDIA, Intel Xeon Phi, or unaccelerated CPUs. If they need double precision, they just don't use Maxwell. But Intel and AMD haven't come out with anything in the space newer than Kepler, either.

Re: Too bad

[Anonymous Coward]

Completely wrong. The reason the previous generation high end Quadro cards had double precision was because they used the same GPU that was designed for use with Tesla cards.

Re: Too bad

[Anonymous Coward]

the last time I sat around a table discussing this kind of thing, it was all about *quad* precision (128 bit) floats anyway. at the time, this is why SPARC had a massive lead, because it actually cared about floats that big.

Re:Too bad

[Anonymous Coward]

A second problem is the exponent. 10^38 just isn't "big" in a modern large scale computation.

This is rarely a problem in large scale computational work, as most problems are already scaled by relevant parameters of a problem. It pretty much all comes back to the first problem of how much precision you have near a value of 1.

Can we trust specs this time

[Anonymous Coward]

How much of the 12GB GDDR5 is actually usable?

Re:Can we trust specs this time

[Anonymous Coward]

11.476 of it.

12GB for?

[Whiteox]

This is old news. Reviews have pointed out that the 12GB ram just can't be used fully as no game can use any more than about 6GB even using multiple 4k monitors. Their conclusion reminds one of high end audiophile scams. Here's a very expensive card that doesn't do too much more than the base Titan model.
So... The AMD R9 295X2 as a single card still beats the Titan. This has a more usable 8GB ram. With a recent price drop, this is a serious card and hitting the sweet spot atm.

Back to the workstation: So maybe if you are mining bitcoins or have specific software to take advantage of the ram, then maybe it would be worth it. But you can buy 4 AMD 295X2 cards for the price of that Titan.

Re:12GB for?

[Anonymous Coward]

You're talking about a completely different card.

Re:12GB for?

[Anonymous Coward]

Reviews have pointed out that the 12GB ram just can't be used fully as no game

This is a workstation card, 12GB is barely enough for some users.

Re:12GB for?

[Anonymous Coward]

Please describe any scenario that would realistically require 12GB video memory.

Re:12GB for?

[Anonymous Coward]

Please describe any scenario that would realistically require 12GB video memory.

No problem! The list for VFX and engineering apps is considerable BTW.

I see this none argument all the time. Just because you have no requirement, that's far from protesting that there's no requirement in industry. In fact there obviously is, otherwise NVidia would not be making a workstation card with 12GB RAM.

Re:12GB for?

[Anonymous Coward]

I like BBW porn. 12Gb is 300kg or 661lb.

Re:12GB for?

[Smidge204]

High end 3D CAD packages can benefit greatly from accelerated graphics processing, and having gobs of memory on the video card can help store all that data.

Alternatively; imagine you work for Pixar.
=Smidge=

Re:12GB for?

[Anonymous Coward]

Reviews have pointed out that the 12GB ram just can't be used fully as no game can use any more than about 6GB even using multiple 4k monitors.

Either you can't read the summary or don't know what a workstation is. In any case, your opinion is so woefully, hilariously under-informed as to be worthless.

Re:12GB for?

[Anonymous Coward]

I do work in Maya, Having a better card will easily cause my render job to go faster, having more RAM will mean that my texture maps will be kept in GPU memory instead of being saved to disk then reloading it on demand.

Yes, only 6 gb may be in active use, but you have 6 gb to actively use as a ram drive to store all that extra data that might not be needed at a given moment.

Re:12GB for?

[Anonymous Coward]

You're talking about playing games on a Quadro card? WTF are you smoking.

They don't make this card for games.

Re: 12GB for?

[Anonymous Coward]

12 GB is for compute, not for "scamming". Radeon 295x2 not only miss out on CUDA, they are water-cooled 500 watt cards with two GPUs and less RAM. All disadvantageous. If one wanted two GPUs at 500 watts one could use two Titan Xs and get CUDA support, much better performance, more RAM, and a simpler cooling system.

Re:12GB for?

[rl117]

3D volume rendering and large scale 2D image rendering. Modern biological and medical imaging systems can create gargantuan datasets, and visualising them is computationally expensive. The largest single images I have are ~5GiB from tiled confocal z stacks (large 3D volume e.g. 4096x4096x64 with 4x 16-bit channels is 8GiB). Current lightsheet microscopes can acquire over 1TB *per sample* in just a few minutes, with the final processed/renderable volume still being many tens of GBs. MRI scans can also be very large.

Put it this way: a 512x512x512 volume at 8 bits per pixel is 128 MiB and that's pretty pedestrian
1024x1024x1024 at 32 bits per pixel/single precision float with 4 channels is 16.4GiB
Current high-end CCDs are now 2048x2048; if I acquire with 16 bits per pixel, 128 planes and 4 channels, that's 32.8 GiB
Slide scanners can also acquire very large 2D images, well over 100000x100000 with 3 8bpp channels; that's 28.6 GiB uncompressed, and I've seen far larger.

I hope this makes the point that 12GB video memory isn't just desirable, it's still way to small for many current imaging applications without pre-processing the data to filter/downsample/reduce the sample precision. I would really appreciate a GPU with this much VRAM, and still want more.

Regards,
Roger

Re: 12GB for?

[Whiteox]

But use 2 slots, suffer from frame jitter due to the SLI and less bang per buck.
The cooling system is a tiny bit of a hassle, especially if you also have cpu water cooling requiring a specialized case. I'd like to see Fiji, the next GPU coming out soon. Nvidia does have the drop on AMD though. No question about that.

Re: 12GB for?

[Anonymous Coward]

What compute situations worry about frame jitter, and how many compute rigs have water cooling blocks for the cpu? Off-topic: The 295x2 is Crossfire which has similar disadvantages to SLI, anyway.

Not THAT surprising

[Anonymous Coward]

Seeing that the AMD offer is over a year old and the next one should arrive in a couple months.

Re: Not THAT surprising

[Anonymous Coward]

Well, when I was at AMD years ago, the graphics GM basically said that since workstation graphics was so low volume, we (AMD) weren't going to do much of anything to specifically target it. If we got wins by doing nothing, well great, otherwise, oh well. That mindset still persists, me thinks...

Shouldn't Slashdot put up a

[wiredog]

"Sponsored Content" banner at the top of this post?

Re:Shouldn't Slashdot put up a

[flopsquad]

"Sponsored Content" banner at the top of this post?

Right next to the "Guaranteed to Run Crysis" stamp!

Oh boo hoo ATI

[Anonymous Coward]

Looks like AMD dragged you down! Any idiot should have seen anything they touch turns to shit. Suckas.

Re: Oh boo hoo ATI

[Anonymous Coward]

As someone who was at ATI when AMD bought us, I can only say that you've no idea how much that announcement sucked. ATI was a great place to work, on competitive products. AMD was a terrible place to work, and management consistently screwed with products and schedules so that they ended up less and less competitive as time went on.

Don't go thinking Lisa Su is going to save it, she was GM during some very serious screw ups.

GOOD

[Anonymous Coward]

The article continued in the blog "pioneers of informatics" Newsflash for Graphics Card
http://info-goers.blogspot.com/2015/03/elsa-gtx-960.html

Other way 'round

[DrunkenTerror]

For a customer service phone drone, asking someone to turn their ethernet cables around the other way is actually a brilliant tactic. Of course we all know the cables are bidirectional, but most lusers don't have a clue. By asking them them to flip the ends around you're really asking them to reseat the connections at either end, but it's seemingly strange and arcane enough that they'll actually do it, instead of just making some fake noises into the phone and saying "There, I've done it." You might be surprised how many bad connections get fixed by "flipping the ends around."

M6000 Name? ... and a note from nvidia

[Anonymous Coward]

Is it called the M6000 because it costs $6000? Does it stand for MONEY 6000!

From nvidia:

Money. In our pockets. No you don't get 1/2 double performance... You don't even get 1/10th double performance! Suck our balls, consumers and engineers! Yes it is exactly the same as our Titan X, just not quite as crippled. All for just $6000!

Tired of Gforce / Quadro Marketing Crap

[BrendaEM]

The triangles shouldn't care if they are being rendered for a game or cad.
With Nvidia, you have a choice of a workstation card that cannot cool itself, or a a gaming card that has been intentionally crippled for CAD.

Lame.

[rthille]

Less space (12GB vs 16GB) than AMD
No DisplayPort 1.3
No Wireless
Lame.