NVIDIA Maximus Fuels Workstation Revolution With Kepler Architecture

[Casecutter]

The Skyraider - The Warhog before there was an A10

Quadro already running, while AMD can't make Firepro commercials for enormous power consumption

And Nvidia has fanciful old planes while no product till Q4, if… The saga of Kepler continues… dragging on!

As to compute for what business actual use such cards for, I’d concur AMD's GCN performance/Watt is hard to hard to be incredulous. And yes let hope the driver team that was working monthly for the desktop updates has now been allocated to workstation and really succeeding in augmenting that software's support.

 

[Steevo]

Don't understand.


.

Farva: It doesn't matter cause I'm going to win ten million dollars.
Thorny: What are you going to do with ten million dollars, and you can't say buy the Cleveland Cavaliers.
Farva: I'd buy a ten million dollar car.
Thorny: That's a good investment but I'd still pull you over.
Farva: Bull Shit. You couldn't pull me over, and even if you did I'd activate my car's wings and I'd fly away.


Are we really arguing about a card that hasn't even seen the light of day as far as we know? Fanboyisim at its best worst.

 

[TheMailMan78]

Farva: It doesn't matter cause I'm going to win ten million dollars.
Thorny: What are you going to do with ten million dollars, and you can't say buy the Cleveland Cavaliers.
Farva: I'd buy a ten million dollar car.
Thorny: That's a good investment but I'd still pull you over.
Farva: Bull Shit. You couldn't pull me over, and even if you did I'd activate my car's wings and I'd fly away.


Are we really arguing about a card that hasn't even seen the light of day as far as we know? Fanboyisim at its best worst.

I still have no clue what you just said. But I do like the idea of a flying car.

 

[Steevo]

I still have no clue what you just said. But I do like the idea of a flying car.

I feel sorry for your soul.


Super Troopers, watch it.

 

[TheMailMan78]

I feel sorry for your soul.


Super Troopers, watch it.

I have a busy life man. I just saw Sling Blade for the first time this past week. But Ill check it out.

Anyway we should get back on topic.
Does anyone get the feeling NVIDIA and AMD charge more for "pro" parts for no good reason other then its a industrial product?

 

[Benetanegia]

Waaaat? Were in the world do you get your info from? I'm yet to see a compute benchmark with Fermi beating gcn, unless an app is heavily optimized for cuda its very unlikely, anything opencl is amd for the win.
Theoretical compute figures demonstrate the peak computational ability in a perfect world I understand that point, but I'm aware that gcn is more efficient in general compute than anything NVIDIA has, if anything is holding their professional business back is probably their customer support, but considering the fact they waited this long to release these cards even though they had the designs readily and selling in the mainstream tells me they made sure stability and drivers were optimized, but that part is my optimistic assumption.

First of all my comment was clearly not meant to say that Fermi rules or anything. Just that Tflops means next to nothing as can be seen by Fermi cards beating some GCN based cards with twice as high maximum theoretical flops. AMD HD6000 and HD5000 had very high TFlops too and were piss poor for compute.

As you can see above the Fermi card can beat every other card (AESEncrypt) and pwns HD5000 and 6000 cards, all of which have 2x or even 3x as high maximum theoretical flops. And in DirectX11 compute Kepler apparently pwns. We could try to conclude several things from this, but the only real thing is that GPU compute is still in its infancy and it heavily depends on which app has been optimized for which GPU or which architecture and most consumer level apps and benchmarks are simply not properly optimized for every card and architecture. It's a hit or miss thing.

The GCN that handily beats Fermi is not the $899 card, it's the $3999, big difference. Against the $899 the top Fermi card wins as many available benches as it loses, which demostrates that you just can't simply compare 2 architectures based on an arbitrary metric such as maximum theoretical flops. I could design a GPU with 200.000 very simple shaders and it would pwn everything on Earth in that metric, but it would mean nothing unless the ISA was complex enough to actually do something with those ALUs, unless I had a cache and memory system capable of feeding it etc, etc.

Comparing 2 different GPU architectures' compute capabilities based on TFLOPs is much worse than comparing CPUs such as Sandy Bridge and Bulldozer, because CPUs are much more efficient at reaching their maximum flops than GPUs and thus difference from optimitations are inherently smaller. And yet comaring 2 CPU architectures based on their maximum theoretical numbers is stupid and useless, comparing GPUs is simply the next level of bad.

 

[Xzibit]

So in these charts, we can assume that your comparing.

GK104s FP32 1/24 to Tahiti FP64 1/4, SP & DP ?

I would assume so since these GPGPUs are aimed at professionals but i'd hate to find out your using these gaming charts/file encryption that dont take the benefits of a professional card in to account to exacerbate your point.

 

[Benetanegia]

So in these charts, we can assume that your comparing.

GK104s FP32 1/24 to Tahiti FP64 1/4, SP & DP ?

I would assume so since these GPGPUs are aimed at professionals but i'd hate to find out your using these gaming charts/file encryption that dont take the benefits of a professional card in to account to exacerbate your point.

If you learn to read the OP. GK104 is used in Quadro and FP64 compute ability has little to no use here.

In fact, the cards mentioned/compared by both of you are visualization cards and not compute cards. Their value is weighted in their ability to render triangles and pixels, both tasks at which GK104 excels and in the case of triangles/primitives by a big big real big margin, so yes let's take the actual benefits into account shall we? Or we can continue making a false claim based on arbitrary numbers that mean absolute nothing in the real world and pretend the value of a card must be weighted based on that false principle.

GK110 the actual compute chip will have at least 1/2 FP64 rate (plus a shitton of Tflops), but 2/3 and 1/1 has been mentioned too. But the fact is that the position of the professional compute world has never been clearer than when they asked/requested for a GK104 based card, because you know, 64 bit is not something EVERYBODY needs or wants, professional or not. Having one chip with 64 bit is required if you want to compete, yes, and if you have only one suitable high-end chip, that chip will require FP64 performance, of course, but if you have 2 winners on the top, not all of your chips need to be. In the end reality wins and professionals have been chosing their prefered solution and THAT is the best indication of which platform is ultimately the better one.

 

[HumanSmoke]

And Nvidia has fanciful old planes while no product till Q4, if… The saga of Kepler continues… dragging on!

Yup. Nvidia must be gutted.

The momentum AMD had been able to muster in the related market for professional graphics hardware has not only petered out, it's now turned backward. AMD's FirePro brand did not gain on Nvidia's Quadro in Q4'11, rather it went in reverse, coming in at 18.4%. Worse, in the first quarter of 2012, AMD's taken a bigger step backward, with FirePro now falling to just 15% of units.

[JPR workstation report]

As to compute for what business actual use such cards for, I’d concur AMD's GCN performance/Watt is hard to hard to be incredulous

Yup, and a top fuel dragster can cover a quarter mile in less than four seconds. They aren't first choice for picking up the kids from school or commuting to work. I suspect AMD could whup Nvidia for 2-3 generations of pro graphics on sheer performance (tho' I suspect they wont as AMD are still angling for the APU as their go-to product line) but it still wouldn't make a dent in the overall market share - too many off the shelf solutions that work, too many high powered/high dollar partnerships. AMD is playing catch-up from a long way behind- and the actual hardware is the least of the issues.

And yes let hope the driver team that was working monthly for the desktop updates has now been allocated to workstation and really succeeding in augmenting that software's support.

Ah, the $64,000 question. This is a company that habitually release desktop cards with unoptimized drivers (imagine if Tahiti had access to Cat 12.7 at launch), touted VCE as a major bullet point for the HD 7000 then took 7+ months to actually get it working. Add in the fact that AMD have never made a secret that they aren't a software company- whereas their competition has demonstratively shown that they are, and I suspect AMD are going to need to call on a lot more than their driver team. AMD when all is said and done, is still reliant upon third parties to weigh in for software integration.

Now, with Nvidia owning 85% of the pro graphics market (and likely more of HPC)- and a record of GPGPU prioritization, do you think it likely that they will relinquish their position? and that they don't have a solution in place to succeed existing Fermi/Tesla ?

 

[Xzibit]

If you learn to read the OP. GK104 is used in Quadro and FP64 compute ability has little to no use here.

In fact, the cards mentioned/compared by both of you are visualization cards and not compute cards. Their value is weighted in their ability to render triangles and pixels, both tasks at which GK104 excels and in the case of triangles/primitives by a big big real big margin, so yes let's take the actual benefits into account shall we? Or we can continue making a false claim based on arbitrary numbers that mean absolute nothing in the real world and pretend the value of a card must be weighted based on that false principle.

GK110 the actual compute chip will have at least 1/2 FP64 rate (plus a shitton of Tflops), but 2/3 and 1/1 has been mentioned too. But the fact is that the position of the professional compute world has never been clearer than when they asked/requested for a GK104 based card, because you know, 64 bit is not something EVERYBODY needs or wants, professional or not. Having one chip with 64 bit is required if you want to compete, yes, and if you have only one suitable high-end chip, that chip will require FP64 performance, of course, but if you have 2 winners on the top, not all of your chips need to be. In the end reality wins and professionals have been chosing their prefered solution and THAT is the best indication of which platform is ultimately the better one.

I'm sure you will correct me if i'm wrong

AMD FirePro (W & S series)
*9000 - K20 (MSRP $3,999 vs $3,199)
*8000 - K10 (MSRP $1,599 vs $2,249) <- This is the comparison in question performance/price and if so the W8000 has 4x throughput then the K5000
*7000 - K10 & NVS (MSRP $899 vs n/a)
*5000 - NVS (MSRP $599 vs n/a)
*x00 - The hundred denotation will be Entry level (MSRP Sub $599 vs n/a)

So when i did my comparison i was being generous. I dont doubt the K20 will be better then *9000 if Nvidia does come through with what is said it would do but since they decided to use the GK104 in quadro its expectations can go either way.

For your assertion of FP64. Just cause you use FP64 doesnt mean your locked into it. Companies may need the efficiency of a FP64 DP for just One step of their development process where the margin of error needs to be of that standard and then do the rest of the work on FP32 SP. It be silly for them to have them seperate just for that purpose.

Can you provide resources to your big big real big margin & plus a shitton of Tflops referance. It sounds so scientific.

 

[Recus]

Top Firepro W9000 bites the dust by previous gen Quadro 6000.

http://hothardware.com/Reviews/AMDs-New-FirePro-W8000-W9000-Challenge-Nvidias-Quadro/?page=1

 

[Frick]

Top Firepro W9000 bites the dust by previous gen Quadro 6000. http://i56.tinypic.com/2e1sp3m.gif

http://hothardware.com/Reviews/AMDs-New-FirePro-W8000-W9000-Challenge-Nvidias-Quadro/?page=1

Ouchouchouch. That is quite bad. Lets hope it's drivers.

 

[Benetanegia]

I'm sure you will correct me if i'm wrong

AMD FirePro (W & S series)
*9000 - K20 (MSRP $3,999 vs $3,199)
*8000 - K10 (MSRP $1,599 vs $2,249) <- This is the comparison in question performance/price and if so the W8000 has 4x throughput then the K5000
*7000 - K10 & NVS (MSRP $899 vs n/a)
*5000 - NVS (MSRP $599 vs n/a)
*x00 - The hundred denotation will be Entry level (MSRP Sub $599 vs n/a)

So when i did my comparison i was being generous. I dont doubt the K20 will be better then *9000 if Nvidia does come through with what is said it would do but since they decided to use the GK104 in quadro its expectations can go either way.

For your assertion of FP64. Just cause you use FP64 doesnt mean your locked into it. Companies may need the efficiency of a FP64 DP for just One step of their development process where the margin of error needs to be of that standard and then do the rest of the work on FP32 SP. It be silly for them to have them seperate just for that purpose.

Can you provide resources to your big big real big margin & plus a shitton of Tflops referance. It sounds so scientific.

Bla bla bla you keep insisting on comparing them based on the arbitrary value of maximum theoretical Tflops or as you named it throughput. I don't want to spend any more time on you. The link by Recus is more than enough. Now continue on insisting that Tflops == performance all you want, all day long if you want.

As to my comment saying "big big real big" margin in triangles/primitives I assumed I was speaking to informed people and it was not necessary to cite known numbers. Since you know the Tflops of each cards so well I think it was safe to assume you knew other specs too. But I should have realized you don't or you probably wouldn't insist on that single one metric so much.

Sigh... anyway, Gk104 and GF110 do 4 triangles per clock; Cayman, Pitcairn and Tahiti do 2 per clock. The result is this:

Theoretical:

On practice. One implementation, for other examples look at Recus' link.

I'd say more than 2x-3x the throughput can be labeled as real big.

The mistery that Hothardware didn't seem to find a response to in their "These results make little sense" section is not such a mistery once you know what you're talking about and where to look for a bottleneck.

 

[Xzibit]

Oh snap, Benchmarks that dont take Fp32 or Fp64 SP, DP into account

http://www.tomshardware.com/reviews/geforce-gtx-690-benchmark,3193-12.html

What good is that throughput if its not accurate or up to the specified standards ?

By all means buy a graphic card right ?


Show me some more benchamrks please

I atleast have to give credit to Recus for providing a subtanent link. Something you never do. Might want to bookmark that one Benetanegia. You also might not want to lead with the whole Driver Optimization thing since it will disprove the advantage of wheather test where done CUDA vs OpenGL and such. I'm sure you know since you spend a good amount on talking about application, driver optimization when Nvidia is on top right ? Oh wait, my mistake...

 

[Urlyin]

That because video games and tech demonstrators always represent reality? Let me guess, you send a weekly email to Unigine to berate them about the the impossibility of floating islands.
Just for the record- the aircraft seems modelled on the Granville Brothers Gee Bee Model Z which first flew in 1931
http://www.passion-aviation.qc.ca/images/musflight08/geebee.jpg

The underwing missiles appear to be Raytheon AIM-9 Sidewinder (1956-present), and the accompanying biplanes (Hawker Hardy's?) have no offensive armament apparent. No doubt the pilots are all sporting non-realistic prosthetic appendages and phasers not set to "stun"

And just for the record, here's an A-1 Skyraider (as mentioned by TheMailman) with less than half it's stores pylons occupied. Similarly/heavier equipped Skyraiders recorded air-to-air dogfight kills against NVAF MiG-17 fighters during the Vietnam War.
And just to get us back on topic

Means little or nothing.
Case in point; My partner works at the my districts hospital- in the medical imaging department. She tells me that AMD did not even compete for the image computation contract. Not because they have don't have a graphics compute solution, but because the contract stipulates 24/7 support (on site as well as instant replacement of parts, 24/7 phone and guaranteed replacement stock for 3-5 years) and a coherant and all encompassing software enviroment...so along with the Tesla/Quadro heavy lifting required for MRI, CT, PET and SPECT (her area of expertise), Nvidia also gets the contract for hundreds of laptop GPU's that technicians/doctors/maintenance use - fixed workstation being largely phased out. Factor in inertia - how often big business upgrades its systems, retraining, and a reluctance to move away from technology/support which already delivers and you should be able to see that theoretical compute numbers mean sweet fuck all.

AMD are basically reinforcing that view. Look at AMD's WS card press release- I count SIX distinct references to "our competitor/competition" - worried much? AMD referenced one design win (Supermicro)- clearly, if AMD's solution (as opposed to GPU) was so much superior, surely they would be counting HP, Dell, Lenovo et al as already on board ?

Even more off topic is that AMD still makes cards for PACS system but is under the BARCO brand....

 

[Steevo]

This thread is still going? Its like an elevator that has reached the bottom floor, but it doesn't stop......



Will all the people who own one of these cards for actual use please line up?

 

[Casecutter]

Yup. Nvidia must be gutted.

Yup, and a top fuel dragster can cover a quarter mile ...
AMD is playing catch-up from a long way behind- and the actual hardware is the least of the issues...

Ah, the $64,000 question...

Never do I want to not to see competition, or one company continually be able to oppress the another...

And yes, for AMD to gain any market share it doesn't need to deliver the utmost performance it just needs to offer a stable software/driver optimization for a $/perf. The one thing they can't pass over is sustaining and vetting business workstation clients software against their product. That’s what the purchaser is expecting for such extortionate prices, they want no hick-ups and white glove support. AMD has to provide that or get out!

 

[Benetanegia]

Oh snap, Benchmarks that dont take Fp32 or Fp64 SP, DP into account

http://www.tomshardware.com/reviews/geforce-gtx-690-benchmark,3193-12.html

What good is that throughput if its not accurate or up to the specified standards ?

By all means buy a graphic card right ?


Show me some more benchamrks please

I atleast have to give credit to Recus for providing a subtanent link. Something you never do. Might want to bookmark that one Benetanegia. You also might not want to lead with the whole Driver Optimization thing since it will disprove the advantage of wheather test where done CUDA vs OpenGL and such. I'm sure you know since you spend a good amount on talking about application, driver optimization when Nvidia is on top right ? Oh wait, my mistake...

Aaand you keep on insisting on compute benchmarks for cards that are NOT for compute. Congratulatios your posts are the most useless ever made!!

It's funny how every benchmark I've shown is not valid according to you, but you post a Sandra bench (omg ) and Luxmark which is known to be non-optimized for Nvidia cards. And this is THE point that I've been pointing out all along, so very funny you try to dismiss it. Some compute benches are won by Nvidia loosely and some are won by AMD loosely. This does not indicate any advantage on the chip/architecture it only proves that optimization is much required and it's not there yet, which again only proves that GPU compute is a new technology.

I've posted 6+ benches that cover ALL the characteristics necessary for a Quadro card and you have shown NOTHING substantial. So congratulations.

So this card @ $2,249 MSRP, USD is suppose to compete against AMD FirePro W7000 @ $899 MSRP, USD.

Talk about a price premium...

Nvidia Quadro K5000
2.1 TFLOPs @ SP

AMD FirePro W7000
2.5 TFLOPs @ SP

The above statement is what we are discussing about. And my answer couldn't be easier to understand. It can compete because it's not the maximum theoretical Tflops the basis by which these type cards are bought. They are bought because they are very good cards for visualization (and you can see this in ALL of my posts as well as the one be Recus). If you can't understand this and you keep on insisting on useless arithmetic FP benchmarks, I'll have to think that you are deficient or something. I don't want to reach that point.

 

[HumanSmoke]

Never do I want to not to see competition, or one company continually be able to oppress the another...

I think that any right minded person would agree with that sentiment. Please don't see my post as "anti-AMD" or lump me in with the nonsensical "want to see AMD dead" trolls- in an ideal world we would have AMD, VIA and ARM challenging Intel, Nvidia and Intel challenging AMD, anyone challenging Microsoft. Unfortunately, what I want, and the reality of the situations rarely converge.

Something, something, shooting the messenger

And yes, for AMD to gain any market share it doesn't need to deliver the utmost performance it just needs to offer a stable software/driver optimization for a $/perf. The one thing they can't pass over is sustaining and vetting business workstation clients software against their product. That’s what the purchaser is expecting for such extortionate prices, they want no hick-ups and white glove support. AMD has to provide that or get out!

Again, a truism. Nvidia's drive has been to market an ecosystem (CUDA, GPGPU, SDK's, Design software, pro drivers) almost entirely predicated upon the high dollar professional market, which usually lumbers the desktop variants with an abundance of overkill in unneeded features and a wasteful perf/watt, perf/$, perf/mm^2, and an overcomplicated solution (for desktop). AMD seem to be on the learning curve that Nvidia started navigating some time back. I've no doubt that AMD given the resources can develop the necessary hardware to challenge anything Nvidia can come up with- the problem is marketing, ongoing support, and most importantly, long range vision. As you say, AMD needs to start delivering- and I hope they do...but what I see presently is a cash-strapped AMD trying to cover a lot of bases with very few resources.

 

[SIGSEGV]

Luxmark which is known to be non-optimized for Nvidia cards.

i want to ask you something, where do you know luxmark is known to be non-optimized for Nvidia cards? it just because nvidia lose on the luxmark bench then you have right to say like that or what? please give me a scientific explanations. thanks.

*afaik luxmark bench has debug infos

 

[Casecutter]

what I see presently is a cash-strapped AMD trying to cover a lot of bases with very few resources.

Yep, and that might be why here a while back AMD slowed the drive release from monthly, to as need. I think Rory and those realize there’s a good return in the workstation business, and let hope that’s where a lot of those resources where refocused to… rather than unemployment lines.

 

[Benetanegia]

i want to ask you something, where do you know luxmark is known to be non-optimized for Nvidia cards? it just because nvidia lose on the luxmark bench then you have right to say like that or what? please give me a scientific explanations. thanks.

*afaik luxmark bench has debug infos

I was too busy responding to stupid posts of you know who and didn't see your post, sorry.

Well it's a known fact for 3d artists and people related to rendering world etc. The most notable source was a Nvidia engineer himself, in Nvidia forums who simply admitted that they don't optimize for Luxmark, because it's not a priority. Keep in mind it's only one of the dozens of renderers, although it's in the spotlight now because AMD uses it on their internal benches and that made some sites start using it too (probably AMD suggests its inclusion in their reviewing guides).

So other than that what can I do? What about some benches from Anandtech which show the matter at hand:

HD6970 review:

GTX560 Ti review:

HD7970 review:

GTX680 review:

As you can see there's always progression on AMD cards. i.e HD6970 8370 -> 10510 -> 11900 -> 13500

This is normal behavior as both the application and AMD drivers are optimized.

On Nvidia on the other hand performance is erratic at best. i.e GTX570 6104 -> 15440 -> 10500 -> 9800

A massive improvement in one particular driver and then down and then even lower again. This does not happen when something is optimized. Well actually, you can see a hint of real Nvidia potential in the GTX560 review. That is how much optimization or lack thereof can affect results in a relatively new and unoptimized benchmark on a relatively new and unoptimized API.

 

[Xzibit]

A massive improvement in one particular driver and then down and then even lower again. This does not happen when something is optimized. Well actually, you can see a hint of real Nvidia potential in the GTX560 review. That is how much optimization or lack thereof can affect results in a relatively new and unoptimized benchmark on a relatively new and unoptimized API.

You might want to check the benchmark discription first aswell, HDR and then HDR+Vol before citing the GTX 560 potential

Maybe you should post CUDA benchmarks to see if there own API shows a discrepancy in "optimization" and gauge that ?

There is several out there. Here is one
A few sites and blog & forums have done comparative test aswell for CUDA vs OpenCL and the % difference is minimal. Single digit % differance. Search away... It kind of throws a wrench in that whole "optimization" theory tho.

You also might want to finish the quote from the Nvidia engineer.. It kind of paints the entire picture of the GK104.

 

[Benetanegia]

You might want to check the benchmark discription first aswell, HDR and then HDR+Vol before citing the GTX 560 potential

Maybe you should post CUDA benchmarks to see if there own API shows a discrepancy in "optimization" and gauge that ?

There is several out there. Here is one
A few sites and blog & forums have done comparative test aswell for CUDA vs OpenCL and the % difference is minimal. Single digit % differance. Search away... It kind of throws a wrench in that whole "optimization" theory tho.

You also might want to finish the quote from the Nvidia engineer.. It kind of paints the entire picture of the GK104.

Wow, you are clueless. Do you even know what OpenCL or CUDA really are? They are APIs, programming frameworks. OpenCL is not different in that regard to OpenGL, DirectX and the likes. They are not applications, they are the tools to create programs. You still have to create your programs and optimize every single one of them.

What you did is like posting a Battlefield 3 benchmark and pretend it's relevant to Crysis 2 optimization level and performance, because both have been developed using DirectX environment. Next time, you might as well compare any 2 applications because they have been created with C++ or because they used the Windows SDK.

But it's even worse, because you coul have at least posted some application benchmarks, but no, you post a link to some matrix*matrix multiplication, and fourier transformations, etc and you pretend it demostrates anything regarding LuxRender optimization. You're so clueless. It's like posting a video of some little boy doing simple arithmetic additions and pretend he must be a genius capable of doing Lorentz transformations.

 

[Xzibit]

You are funny how you somehow insert some off the wall implications that you have in your head.

Its like I dont have to say anything and you insert what ever you want into the conversation. Its halarious.

Do you still have your imaginary friend with you ?