"Portable data-parallel programming through ufuncs and gufuncs for single core CPU, multicore CPU and GPU
Bindings to CUDA libraries: cuRAND, cuBLAS, cuFFT
Python CUDA programming for maximum control of hardware resources"
If Apple cared, they would certainly be able to address problem (a), and they also have sufficient resources to address the larger problem (b)
DDR 4 will sort out the RAM size with the next Mac Pro. It doubles the density so it will go up to 128GB. They could have offered NVidia GPUs but they'd be much more expensive. AMD's GPUs are cheaper and it's the same reason the next-gen consoles use AMD.
So, most likely, my money will be spent on upgrading one of my local linux boxes to a K40 GPU.
Yeah that's what Apple should have done - offer a $5300 GPU as an option, $10k for a pair rather than $1k for a pair of AMD GPUs (they would likely have been cheaper like the AMD ones vs retail but not nearly as cheap):
So is not a Mac Pro It should be renamed a PC Pro. You need windows to get the full power of the GPUs. Amazing staff. The Mac Pro is one of the best Windows PCs ever.
Yeah, all Macs make very good Windows PCs. CrossFire mostly affects games or real-time 3D but a single D700 is already fast on its own. It's a bit silly to actually use both GPUs all the time because that runs the game in the test at over 100FPS. That's just wasting electricity and generating excess heat for a benefit you won't even see. Obviously having the option to force the use of both GPUs would be good but app developers can decide where it's necessary without CrossFire and they have done.
"Portable data-parallel programming through ufuncs and gufuncs for single core CPU, multicore CPU and GPU
Bindings to CUDA libraries: cuRAND, cuBLAS, cuFFT
Python CUDA programming for maximum control of hardware resources"
What's the maximum amount of RAM you've used for a process in the past? GPUs only have up to 6GB so I don't see why 64GB is such a huge limitation.
DDR 4 will sort out the RAM size with the next Mac Pro. It doubles the density so it will go up to 128GB. They could have offered NVidia GPUs but they'd be much more expensive. AMD's GPUs are cheaper and it's the same reason the next-gen consoles use AMD.
Yeah that's what Apple should have done - offer a $5300 GPU as an option, $10k for a pair rather than $1k for a pair of AMD GPUs (they would likely have been cheaper like the AMD ones vs retail but not nearly as cheap):
Yeah, all Macs make very good Windows PCs. CrossFire mostly affects games or real-time 3D but a single D700 is already fast on its own. It's a bit silly to actually use both GPUs all the time because that runs the game in the test at over 100FPS. That's just wasting electricity and generating excess heat for a benefit you won't even see. Obviously having the option to force the use of both GPUs would be good but app developers can decide where it's necessary without CrossFire and they have done.
1. I dont understand the comment about "mathematical algorithms". Of course they are mathematical algorithms, but the use case is very specific.
2. pyopencl: not only do I know about it, but a student of mine improved some version of it. Programming in it is not that much fun.
3. NumbaPro:
CUDA Support & Detection
NumbaPro GPU support currently requires NVIDIA CUDA GPUs with compute-capability 2.0 or above. Users should check their hardware with the following:
4. Memory size: as I had explained, symbolic computation is what uses up RAM. I am renting a server (from hetzner.de) for 130 euro a month with 128GB and a hexacore Xeon. Apparently the current fab density has not stopped whoever made the box.
5. AMD GPU have good specs on paper, but if you just google around you will see that the drivers are generally less stable. Presumably Apple verified that all is well for FCPX, but for the rest, we are on our own. Tesla cards are the industry standard, although, as I said, the nMP is quite attractive on a price-performance basis. However, my time is worth a lot to me. In particular, the fact that the two D700 have a (somewhat) higher performance than one K40 is of questionable relevance, given that it is a headache to get them to work together (from what I gather).
1. I dont understand the comment about "mathematical algorithms". Of course they are mathematical algorithms, but the use case is very specific.
2. pyopencl: not only do I know about it, but a student of mine improved some version of it. Programming in it is not that much fun.
3. NumbaPro:
CUDA Support & Detection
NumbaPro GPU support currently requires NVIDIA CUDA GPUs with compute-capability 2.0 or above. Users should check their hardware with the following:
4. Memory size: as I had explained, symbolic computation is what uses up RAM. I am renting a server (from hetzner.de) for 130 euro a month with 128GB and a hexacore Xeon. Apparently the current fab density has not stopped whoever made the box.
5. AMD GPU have good specs on paper, but if you just google around you will see that the drivers are generally less stable. Presumably Apple verified that all is well for FCPX, but for the rest, we are on our own. Tesla cards are the industry standard, although, as I said, the nMP is quite attractive on a price-performance basis. However, my time is worth a lot to me. In particular, the fact that the two D700 have a (somewhat) higher performance than one K40 is of questionable relevance, given that it is a headache to get them to work together (from what I gather).
Also, re D700 performance: the D700 is a bit slower than the Radeon R280X (it has ECC ram, and more of it, but it IS slower), and the nVidia titan (modulo the same limitations) is comparable to the K20. Notice that in double precision (which is what you need for scientific computation, less so for rendering), the nVidia part is utterly dominant.
1. I dont understand the comment about "mathematical algorithms". Of course they are mathematical algorithms, but the use case is very specific.
What use case? We don't know all the places Apple uses OpenCL. A lot of them are in image processing, which GPU compute excels at but OpenCL isn't limited to that.
Some programs are going to work better with CUDA but Apple developed OpenCL to allow using multiple hardware options. They have no reason to turn round and back proprietary NVidia options. Maybe they could have offered a dual Quadro as an option but the point is developers shouldn't use CUDA. If OpenCL isn't good enough, they need to say why and have the likes of AMD and the OpenCL spec developers improve it.
4. Memory size: as I had explained, symbolic computation is what uses up RAM. I am renting a server (from hetzner.de) for 130 euro a month with 128GB and a hexacore Xeon. Apparently the current fab density has not stopped whoever made the box.
And what's the largest amount you've seen a single process use? Also, like I said if 128GB is your preference, DDR4 will bring it next year:
AMD GPU have good specs on paper, but if you just google around you will see that the drivers are generally less stable. Presumably Apple verified that all is well for FCPX, but for the rest, we are on our own. Tesla cards are the industry standard, although, as I said, the nMP is quite attractive on a price-performance basis. However, my time is worth a lot to me. In particular, the fact that the two D700 have a (somewhat) higher performance than one K40 is of questionable relevance, given that it is a headache to get them to work together (from what I gather).
You'd manage to find faults with NVidia setups too. Here's one freezing up the computer display:
Apple wouldn't realistically ship the Mac Pro with a dual-Tesla setup. The price would be far too high. Ultimately, Apple or any manufacturer can offer a dream-machine spec but they have to hit a large enough target audience for it to be worth doing. If a 128GB RAM option and a dual Tesla applies to 1000 people worldwide, it's not worth doing because it means custom GPU boards just for them.
What use case? We don't know all the places Apple uses OpenCL. A lot of them are in image processing, which GPU compute excels at but OpenCL isn't limited to that.
Is there a fun part to running compute functions manually on a graphics card?
Some programs are going to work better with CUDA but Apple developed OpenCL to allow using multiple hardware options. They have no reason to turn round and back proprietary NVidia options. Maybe they could have offered a dual Quadro as an option but the point is developers shouldn't use CUDA. If OpenCL isn't good enough, they need to say why and have the likes of AMD and the OpenCL spec developers improve it.
And what's the largest amount you've seen a single process use? Also, like I said if 128GB is your preference, DDR4 will bring it next year:
Apple wouldn't realistically ship the Mac Pro with a dual-Tesla setup. The price would be far too high. Ultimately, Apple or any manufacturer can offer a dream-machine spec but they have to hit a large enough target audience for it to be worth doing. If a 128GB RAM option and a dual Tesla applies to 1000 people worldwide, it's not worth doing because it means custom GPU boards just for them.
I think you are changing (have changed?) the subject. My argument was that Apple doesn't care that much about the scientific computing crowd, and you are now explaining why they should not. Whatever faults Apple has, poor marketing is not among them, so I am sure they have excellent reasons. As for the specific points:
1. note that I had said that nVidia was MORE stable
2. There are a number of companies (including such stalwarts as Dell and HP, but also dozens/hundreds of smaller vendors, including the likes of Boxx, SiliconMechanics, XiComputer, etc, etc, etc) who will be happy to sell me a machine with 128GB of ram, and four Tesla Ks. This, to me, means that there IS a market, which probably numbers in (at least) hundreds of thousands of machines.
3. Re 128GB: I repeat, 128GB configurations have existed on servers/workstations for a very long time, my point about my server was that by now this configuration is quite cheap. Undoubtedly, next year it will be cheaper yet, and in five years toasters will have that much RAM, but that's neither here nor there.
4. CUDA vs OpenCL: again, I am not saying that Apple is doing anything wrong from the viewpoint of their self-interest, but their "Pro" machines are not as useful as they could be for me.
5. Cost of tesla cards: the GTX Titan has 95% of the functionality for 1/4 the price. In fact, given that the main use case for the MacPro is video editing, it is not clear that ECC is especially useful (to be honest, before writing this, I tried to find any examples where it IS useful, but failed -- they are apparently whole clusters churning happily away with non-ECC cards). A GTX titan is $750 at NewEgg, and probably can be had for half that if you are purchasing in Apple bulk.
My argument was that Apple doesn't care that much about the scientific computing crowd, and you are now explaining why they should not.
Apple doesn't go out of their way to cater to special interest groups and the small volume of those groups is enough justification for that but that doesn't mean they don't expect anyone from those groups to be able to use their products. They have a close relationship with Wolfram who use OpenCL for scientific computation:
1. note that I had said that nVidia was MORE stable
That's quite a broad generalisation. I doubt that you could say NVidia would automatically perform better in more computing scenarios, there are too many variables - drivers, GPU model, system design, code, OS versions. If you've had better experiences with NVidia, that's understandable and the option to have NVidia GPUs would be good for cases where CUDA is required but it shouldn't be necessary.
2. There are a number of companies (including such stalwarts as Dell and HP, but also dozens/hundreds of smaller vendors, including the likes of Boxx, SiliconMechanics, XiComputer, etc, etc, etc) who will be happy to sell me a machine with 128GB of ram, and four Tesla Ks. This, to me, means that there IS a market, which probably numbers in (at least) hundreds of thousands of machines.
If you spec out a Boxx with dual 12-cores, 128GB RAM, 4x NVidia K5000s, the price is $36,000. If the market for that was 100,000 units, they'd make $3.6b in revenue and you can bet Apple would cater to that group. Estimates for Boxx's revenue are about $20-30m:
4. CUDA vs OpenCL: again, I am not saying that Apple is doing anything wrong from the viewpoint of their self-interest, but their "Pro" machines are not as useful as they could be for me.
At best they'd offer dual Quadros but if they did, you'd then say it's only two GPUs and you can get four from Boxx. If they offered 128GB RAM, you'd say you could get 256GB from Boxx. If they offered everything Boxx did, you'd say you could get it cheaper from Boxx. The choice they make is simple, they don't bother catering for the rare event that someone is willing to part with $36,000 for a computer.
5. Cost of tesla cards: the GTX Titan has 95% of the functionality for 1/4 the price. In fact, given that the main use case for the MacPro is video editing, it is not clear that ECC is especially useful (to be honest, before writing this, I tried to find any examples where it IS useful, but failed -- they are apparently whole clusters churning happily away with non-ECC cards). A GTX titan is $750 at NewEgg, and probably can be had for half that if you are purchasing in Apple bulk.
The 780 Ti is faster than the Titan and cheaper, two of those would have been a good option to have. This always happens though how the likes of Boxx is mentioned with Teslas and it always ends up at using cheap gaming cards, preferably in a machine under $2k with a 6-core i7.
At the end of the day, there's nothing stopping people building whatever servers and supplementary computers they want, Apple builds their machines to be nice, reliable workstations. You can see in the following video someone's gaming rig next to the Mac Pro:
[VIDEO]
While that design will allow for whatever NVidia GPU config, there's no way they'd convince more people to buy it than the compact workstation.
Apple doesn't go out of their way to cater to special interest groups and the small volume of those groups is enough justification for that but that doesn't mean they don't expect anyone from those groups to be able to use their products. They have a close relationship with Wolfram who use OpenCL for scientific computation:
That's quite a broad generalisation. I doubt that you could say NVidia would automatically perform better in more computing scenarios, there are too many variables - drivers, GPU model, system design, code, OS versions. If you've had better experiences with NVidia, that's understandable and the option to have NVidia GPUs would be good for cases where CUDA is required but it shouldn't be necessary.
If you spec out a Boxx with dual 12-cores, 128GB RAM, 4x NVidia K5000s, the price is $36,000. If the market for that was 100,000 units, they'd make $3.6b in revenue and you can bet Apple would cater to that group. Estimates for Boxx's revenue are about $20-30m:
If they only sold those high-end machines, that would be under 1,000 units. There's a market for small vendors, not for Apple.
At best they'd offer dual Quadros but if they did, you'd then say it's only two GPUs and you can get four from Boxx. If they offered 128GB RAM, you'd say you could get 256GB from Boxx. If they offered everything Boxx did, you'd say you could get it cheaper from Boxx. The choice they make is simple, they don't bother catering for the rare event that someone is willing to part with $36,000 for a computer.
The 780 Ti is faster than the Titan and cheaper, two of those would have been a good option to have. This always happens though how the likes of Boxx is mentioned with Teslas and it always ends up at using cheap gaming cards, preferably in a machine under $2k with a 6-core i7.
At the end of the day, there's nothing stopping people building whatever servers and supplementary computers they want, Apple builds their machines to be nice, reliable workstations. You can see in the following video someone's gaming rig next to the Mac Pro:
While that design will allow for whatever NVidia GPU config, there's no way they'd convince more people to buy it than the compact workstation.
First, a technical point: the 780Ti is useless for computing applications, since it has the usual (for gaming cards) throttled double precision. On the other hand, the titan can only be gotten for $3K these days, for which money you can get an actual K20. And as for nVidia vs AMD, a D700 does 800 giga flops double precision, a Titan (and hence a K20) does 1.3 teraflops, and so a K40 is the equivalent of the TWO D700s (K40 is faster than a K20x, by around 30%).
Otherwise: I repeat, I am not presuming to know Apple's marketing better than Apple does, and you are impugning that my point is to trash talk Apple (see your comments above re Apple vs Boxx.) It is not. I am saying that the Mac Pro will not be popular in scientific computing circles, because it is not quite what is needed -- I really like the form factor of the machine, and I would like to convince myself that it is the right thing to get, but it just isn't.
This conversation has consisted of me explaining why this is true, and you explaining to me why Apple should not care.
as for nVidia vs AMD, a D700 does 800 giga flops double precision, a Titan (and hence a K20) does 1.3 teraflops, and so a K40 is the equivalent of the TWO D700s (K40 is faster than a K20x, by around 30%).
you are impugning that my point is to trash talk Apple (see your comments above re Apple vs Boxx.) It is not. I am saying that the Mac Pro will not be popular in scientific computing circles, because it is not quite what is needed -- I really like the form factor of the machine, and I would like to convince myself that it is the right thing to get, but it just isn't.
This conversation has consisted of me explaining why this is true, and you explaining to me why Apple should not care.
You're not explaining why this is true universally. You assume that your specific software applies to the entire scientific community and make generalised statements about the Mac Pro's unsuitability for it and that Apple therefore doesn't care at all about that market segment. That's trash talk. The Mac Pro may be less optimal for that portion of the scientific community but just because they opted for AMD GPUs doesn't eliminate the Mac Pro from it entirely.
Worst case, there is the possibility of running an NVidia GPU externally over Thunderbolt:
so two D700s with 870 gigaflops each is 1.74 teraflops double precision. The two D700s cost $1000, K40 retail is $5300.
You're not explaining why this is true universally. You assume that your specific software applies to the entire scientific community and make generalised statements about the Mac Pro's unsuitability for it and that Apple therefore doesn't care at all about that market segment. That's trash talk. The Mac Pro may be less optimal for that portion of the scientific community but just because they opted for AMD GPUs doesn't eliminate the Mac Pro from it entirely.
Worst case, there is the possibility of running an NVidia GPU externally over Thunderbolt:
"I've run a GeForce Titan through thunderbolt 1 on a Mac, and saw only a 5% performance drop compared to a custom bult PC with a 3770k."
For computation, if the data set is in the GPU memory, it's not going to be limited by the connection.
That's interesting, about running an external GPU. Otherwise, I know it is shocking, but I do actually talk to (a lot of) other people, and their needs are not so different from mine. And you are right that MOST computations do not require 128GB, but if you have one that does, do you suggest one go out and buy another machine? As for trash talking, you are (for reasons that elude me, since you are not Tim Cook, as far as I know) being defensive. I have no issue with Apple largely punting on the scientific computing segment. They were never big in it, they don't really understand it, and they are wise to concentrate more on what they do know. In any case, most hardware vendors prefer to build a lot of jellybean machines -- apparently DOE had had to ask IBM really really nicely to build them a supercomputer, since it is not really worth a vendor's time. Sad, but people find workarounds.
I know it is shocking, but I do actually talk to (a lot of) other people, and their needs are not so different from mine.
It's not shocking that you talk to a lot of people but even if it's 1000 people, it's still a fraction of a community that you describe as being in the hundreds of thousands and I'm sure not every single one of the people you talk to entirely dismisses AMD or OpenCL. What happens if NVidia goes bankrupt? All that code would have to be rewritten.
And you are right that MOST computations do not require 128GB, but if you have one that does, do you suggest one go out and buy another machine?
It will still work, it'll just go into swap. Plus, in the context of GPU computing using 6GB of memory, 64GB is a lot of RAM to work with and again DDR4 will boost it back up to 128GB.
I have no issue with Apple largely punting on the scientific computing segment. They were never big in it, they don't really understand it, and they are wise to concentrate more on what they do know.
You say they don't understand it but they developed a compute language that works across multiple hardware architectures so what don't they understand? Have they designed OpenCL incorrectly?
It's not shocking that you talk to a lot of people but even if it's 1000 people, it's still a fraction of a community that you describe as being in the hundreds of thousands and I'm sure not every single one of the people you talk to entirely dismisses AMD or OpenCL. What happens if NVidia goes bankrupt? All that code would have to be rewritten.
It will still work, it'll just go into swap. Plus, in the context of GPU computing using 6GB of memory, 64GB is a lot of RAM to work with and again DDR4 will boost it back up to 128GB.
You say they don't understand it but they developed a compute language that works across multiple hardware architectures so what don't they understand? Have they designed OpenCL incorrectly?
1. All the code will have to be rewritten anyway, since the technology changes quickly.
2. "Go into swap" means "sayonara" [with SSDs the hit will not be quite as massive, but still I am guessing a couple of orders of magnitude in speed -- I would love to be proven wrong.]
3. They don't understand what the community needs. The fact that OpenCL is still not that popular is an indicator of this -- no one wants to be beholden to a single vendor, but CUDA comes with VERY nice libraries and tools. My understanding is that OpenCL does not. That said, it is certainly true that you can get Matlab or Mathematica on your Mac, and it will work quite well, but that just shunts the dependence to a different proprietary language.
Comments
Whether they run the code on image data or something else, they are running mathematical algorithms:
http://developer.amd.com/community/application-showcase/scientific/
http://www.wolfram.com/mathematica/new-in-8/cuda-and-opencl-support/
http://developer.amd.com/community/application-showcase/
http://mathema.tician.de/software/pyopencl/
http://enja.org/2011/02/22/adventures-in-pyopencl-part-1-getting-started-with-python/
The NumbaPro site says it uses any GPU, it just has additional functionality for CUDA, hardly a show-stopper:
http://docs.continuum.io/numbapro/
"Portable data-parallel programming through ufuncs and gufuncs for single core CPU, multicore CPU and GPU
Bindings to CUDA libraries: cuRAND, cuBLAS, cuFFT
Python CUDA programming for maximum control of hardware resources"
What's the maximum amount of RAM you've used for a process in the past? GPUs only have up to 6GB so I don't see why 64GB is such a huge limitation.
DDR 4 will sort out the RAM size with the next Mac Pro. It doubles the density so it will go up to 128GB. They could have offered NVidia GPUs but they'd be much more expensive. AMD's GPUs are cheaper and it's the same reason the next-gen consoles use AMD.
Yeah that's what Apple should have done - offer a $5300 GPU as an option, $10k for a pair rather than $1k for a pair of AMD GPUs (they would likely have been cheaper like the AMD ones vs retail but not nearly as cheap):
http://www.newegg.com/Product/Product.aspx?Item=N82E16814132015
The dual D700 should be faster than a single K40 and a fraction of the price:
http://www.geeks3d.com/20131118/nvidia-tesla-k40-announced-tesla-k20-amd-firepro-s10000-best-performance-watt-solution/
Yeah, all Macs make very good Windows PCs. CrossFire mostly affects games or real-time 3D but a single D700 is already fast on its own. It's a bit silly to actually use both GPUs all the time because that runs the game in the test at over 100FPS. That's just wasting electricity and generating excess heat for a benefit you won't even see. Obviously having the option to force the use of both GPUs would be good but app developers can decide where it's necessary without CrossFire and they have done.
Whether they run the code on image data or something else, they are running mathematical algorithms:
http://developer.amd.com/community/application-showcase/scientific/
http://www.wolfram.com/mathematica/new-in-8/cuda-and-opencl-support/
http://developer.amd.com/community/application-showcase/
http://mathema.tician.de/software/pyopencl/
http://enja.org/2011/02/22/adventures-in-pyopencl-part-1-getting-started-with-python/
The NumbaPro site says it uses any GPU, it just has additional functionality for CUDA, hardly a show-stopper:
http://docs.continuum.io/numbapro/
"Portable data-parallel programming through ufuncs and gufuncs for single core CPU, multicore CPU and GPU
Bindings to CUDA libraries: cuRAND, cuBLAS, cuFFT
Python CUDA programming for maximum control of hardware resources"
What's the maximum amount of RAM you've used for a process in the past? GPUs only have up to 6GB so I don't see why 64GB is such a huge limitation.
DDR 4 will sort out the RAM size with the next Mac Pro. It doubles the density so it will go up to 128GB. They could have offered NVidia GPUs but they'd be much more expensive. AMD's GPUs are cheaper and it's the same reason the next-gen consoles use AMD.
Yeah that's what Apple should have done - offer a $5300 GPU as an option, $10k for a pair rather than $1k for a pair of AMD GPUs (they would likely have been cheaper like the AMD ones vs retail but not nearly as cheap):
http://www.newegg.com/Product/Product.aspx?Item=N82E16814132015
The dual D700 should be faster than a single K40 and a fraction of the price:
http://www.geeks3d.com/20131118/nvidia-tesla-k40-announced-tesla-k20-amd-firepro-s10000-best-performance-watt-solution/
Yeah, all Macs make very good Windows PCs. CrossFire mostly affects games or real-time 3D but a single D700 is already fast on its own. It's a bit silly to actually use both GPUs all the time because that runs the game in the test at over 100FPS. That's just wasting electricity and generating excess heat for a benefit you won't even see. Obviously having the option to force the use of both GPUs would be good but app developers can decide where it's necessary without CrossFire and they have done.
1. I dont understand the comment about "mathematical algorithms". Of course they are mathematical algorithms, but the use case is very specific.
2. pyopencl: not only do I know about it, but a student of mine improved some version of it. Programming in it is not that much fun.
3. NumbaPro:
CUDA Support & Detection
NumbaPro GPU support currently requires NVIDIA CUDA GPUs with compute-capability 2.0 or above. Users should check their hardware with the following:
4. Memory size: as I had explained, symbolic computation is what uses up RAM. I am renting a server (from hetzner.de) for 130 euro a month with 128GB and a hexacore Xeon. Apparently the current fab density has not stopped whoever made the box.
5. AMD GPU have good specs on paper, but if you just google around you will see that the drivers are generally less stable. Presumably Apple verified that all is well for FCPX, but for the rest, we are on our own. Tesla cards are the industry standard, although, as I said, the nMP is quite attractive on a price-performance basis. However, my time is worth a lot to me. In particular, the fact that the two D700 have a (somewhat) higher performance than one K40 is of questionable relevance, given that it is a headache to get them to work together (from what I gather).
1. I dont understand the comment about "mathematical algorithms". Of course they are mathematical algorithms, but the use case is very specific.
2. pyopencl: not only do I know about it, but a student of mine improved some version of it. Programming in it is not that much fun.
3. NumbaPro:
CUDA Support & Detection
NumbaPro GPU support currently requires NVIDIA CUDA GPUs with compute-capability 2.0 or above. Users should check their hardware with the following:
4. Memory size: as I had explained, symbolic computation is what uses up RAM. I am renting a server (from hetzner.de) for 130 euro a month with 128GB and a hexacore Xeon. Apparently the current fab density has not stopped whoever made the box.
5. AMD GPU have good specs on paper, but if you just google around you will see that the drivers are generally less stable. Presumably Apple verified that all is well for FCPX, but for the rest, we are on our own. Tesla cards are the industry standard, although, as I said, the nMP is quite attractive on a price-performance basis. However, my time is worth a lot to me. In particular, the fact that the two D700 have a (somewhat) higher performance than one K40 is of questionable relevance, given that it is a headache to get them to work together (from what I gather).
Also, re D700 performance: the D700 is a bit slower than the Radeon R280X (it has ECC ram, and more of it, but it IS slower), and the nVidia titan (modulo the same limitations) is comparable to the K20. Notice that in double precision (which is what you need for scientific computation, less so for rendering), the nVidia part is utterly dominant.
What use case? We don't know all the places Apple uses OpenCL. A lot of them are in image processing, which GPU compute excels at but OpenCL isn't limited to that.
Is there a fun part to running compute functions manually on a graphics card?
Some programs are going to work better with CUDA but Apple developed OpenCL to allow using multiple hardware options. They have no reason to turn round and back proprietary NVidia options. Maybe they could have offered a dual Quadro as an option but the point is developers shouldn't use CUDA. If OpenCL isn't good enough, they need to say why and have the likes of AMD and the OpenCL spec developers improve it.
And what's the largest amount you've seen a single process use? Also, like I said if 128GB is your preference, DDR4 will bring it next year:
http://wccftech.com/intel-xeon-processors-details-unveiled-haswell-ep-feature-14-cores-35-mb-cache-2014-broadwell-ep-18-cores-45-mb-cache-2015/
http://techreport.com/news/25298/samsung-mass-producing-32gb-ddr4-modules-for-servers
You'd manage to find faults with NVidia setups too. Here's one freezing up the computer display:
http://stackoverflow.com/questions/21555763/computer-freezes-after-running-numbapro-cuda-code
Apple wouldn't realistically ship the Mac Pro with a dual-Tesla setup. The price would be far too high. Ultimately, Apple or any manufacturer can offer a dream-machine spec but they have to hit a large enough target audience for it to be worth doing. If a 128GB RAM option and a dual Tesla applies to 1000 people worldwide, it's not worth doing because it means custom GPU boards just for them.
What use case? We don't know all the places Apple uses OpenCL. A lot of them are in image processing, which GPU compute excels at but OpenCL isn't limited to that.
Is there a fun part to running compute functions manually on a graphics card?
Some programs are going to work better with CUDA but Apple developed OpenCL to allow using multiple hardware options. They have no reason to turn round and back proprietary NVidia options. Maybe they could have offered a dual Quadro as an option but the point is developers shouldn't use CUDA. If OpenCL isn't good enough, they need to say why and have the likes of AMD and the OpenCL spec developers improve it.
And what's the largest amount you've seen a single process use? Also, like I said if 128GB is your preference, DDR4 will bring it next year:
http://wccftech.com/intel-xeon-processors-details-unveiled-haswell-ep-feature-14-cores-35-mb-cache-2014-broadwell-ep-18-cores-45-mb-cache-2015/
http://techreport.com/news/25298/samsung-mass-producing-32gb-ddr4-modules-for-servers
You'd manage to find faults with NVidia setups too. Here's one freezing up the computer display:
http://stackoverflow.com/questions/21555763/computer-freezes-after-running-numbapro-cuda-code
Apple wouldn't realistically ship the Mac Pro with a dual-Tesla setup. The price would be far too high. Ultimately, Apple or any manufacturer can offer a dream-machine spec but they have to hit a large enough target audience for it to be worth doing. If a 128GB RAM option and a dual Tesla applies to 1000 people worldwide, it's not worth doing because it means custom GPU boards just for them.
I think you are changing (have changed?) the subject. My argument was that Apple doesn't care that much about the scientific computing crowd, and you are now explaining why they should not. Whatever faults Apple has, poor marketing is not among them, so I am sure they have excellent reasons. As for the specific points:
1. note that I had said that nVidia was MORE stable
2. There are a number of companies (including such stalwarts as Dell and HP, but also dozens/hundreds of smaller vendors, including the likes of Boxx, SiliconMechanics, XiComputer, etc, etc, etc) who will be happy to sell me a machine with 128GB of ram, and four Tesla Ks. This, to me, means that there IS a market, which probably numbers in (at least) hundreds of thousands of machines.
3. Re 128GB: I repeat, 128GB configurations have existed on servers/workstations for a very long time, my point about my server was that by now this configuration is quite cheap. Undoubtedly, next year it will be cheaper yet, and in five years toasters will have that much RAM, but that's neither here nor there.
4. CUDA vs OpenCL: again, I am not saying that Apple is doing anything wrong from the viewpoint of their self-interest, but their "Pro" machines are not as useful as they could be for me.
5. Cost of tesla cards: the GTX Titan has 95% of the functionality for 1/4 the price. In fact, given that the main use case for the MacPro is video editing, it is not clear that ECC is especially useful (to be honest, before writing this, I tried to find any examples where it IS useful, but failed -- they are apparently whole clusters churning happily away with non-ECC cards). A GTX titan is $750 at NewEgg, and probably can be had for half that if you are purchasing in Apple bulk.
Apple doesn't go out of their way to cater to special interest groups and the small volume of those groups is enough justification for that but that doesn't mean they don't expect anyone from those groups to be able to use their products. They have a close relationship with Wolfram who use OpenCL for scientific computation:
http://blog.stephenwolfram.com/2011/10/steve-jobs-a-few-memories/
http://www.wolfram.com/mathematica/new-in-8/cuda-and-opencl-support/
That's quite a broad generalisation. I doubt that you could say NVidia would automatically perform better in more computing scenarios, there are too many variables - drivers, GPU model, system design, code, OS versions. If you've had better experiences with NVidia, that's understandable and the option to have NVidia GPUs would be good for cases where CUDA is required but it shouldn't be necessary.
If you spec out a Boxx with dual 12-cores, 128GB RAM, 4x NVidia K5000s, the price is $36,000. If the market for that was 100,000 units, they'd make $3.6b in revenue and you can bet Apple would cater to that group. Estimates for Boxx's revenue are about $20-30m:
http://www.insideview.com/directory/boxx-technologies-inc
If they only sold those high-end machines, that would be under 1,000 units. There's a market for small vendors, not for Apple.
At best they'd offer dual Quadros but if they did, you'd then say it's only two GPUs and you can get four from Boxx. If they offered 128GB RAM, you'd say you could get 256GB from Boxx. If they offered everything Boxx did, you'd say you could get it cheaper from Boxx. The choice they make is simple, they don't bother catering for the rare event that someone is willing to part with $36,000 for a computer.
The 780 Ti is faster than the Titan and cheaper, two of those would have been a good option to have. This always happens though how the likes of Boxx is mentioned with Teslas and it always ends up at using cheap gaming cards, preferably in a machine under $2k with a 6-core i7.
At the end of the day, there's nothing stopping people building whatever servers and supplementary computers they want, Apple builds their machines to be nice, reliable workstations. You can see in the following video someone's gaming rig next to the Mac Pro:
[VIDEO]
While that design will allow for whatever NVidia GPU config, there's no way they'd convince more people to buy it than the compact workstation.
Apple doesn't go out of their way to cater to special interest groups and the small volume of those groups is enough justification for that but that doesn't mean they don't expect anyone from those groups to be able to use their products. They have a close relationship with Wolfram who use OpenCL for scientific computation:
http://blog.stephenwolfram.com/2011/10/steve-jobs-a-few-memories/
http://www.wolfram.com/mathematica/new-in-8/cuda-and-opencl-support/
That's quite a broad generalisation. I doubt that you could say NVidia would automatically perform better in more computing scenarios, there are too many variables - drivers, GPU model, system design, code, OS versions. If you've had better experiences with NVidia, that's understandable and the option to have NVidia GPUs would be good for cases where CUDA is required but it shouldn't be necessary.
If you spec out a Boxx with dual 12-cores, 128GB RAM, 4x NVidia K5000s, the price is $36,000. If the market for that was 100,000 units, they'd make $3.6b in revenue and you can bet Apple would cater to that group. Estimates for Boxx's revenue are about $20-30m:
http://www.insideview.com/directory/boxx-technologies-inc
If they only sold those high-end machines, that would be under 1,000 units. There's a market for small vendors, not for Apple.
At best they'd offer dual Quadros but if they did, you'd then say it's only two GPUs and you can get four from Boxx. If they offered 128GB RAM, you'd say you could get 256GB from Boxx. If they offered everything Boxx did, you'd say you could get it cheaper from Boxx. The choice they make is simple, they don't bother catering for the rare event that someone is willing to part with $36,000 for a computer.
The 780 Ti is faster than the Titan and cheaper, two of those would have been a good option to have. This always happens though how the likes of Boxx is mentioned with Teslas and it always ends up at using cheap gaming cards, preferably in a machine under $2k with a 6-core i7.
At the end of the day, there's nothing stopping people building whatever servers and supplementary computers they want, Apple builds their machines to be nice, reliable workstations. You can see in the following video someone's gaming rig next to the Mac Pro:
While that design will allow for whatever NVidia GPU config, there's no way they'd convince more people to buy it than the compact workstation.
First, a technical point: the 780Ti is useless for computing applications, since it has the usual (for gaming cards) throttled double precision. On the other hand, the titan can only be gotten for $3K these days, for which money you can get an actual K20. And as for nVidia vs AMD, a D700 does 800 giga flops double precision, a Titan (and hence a K20) does 1.3 teraflops, and so a K40 is the equivalent of the TWO D700s (K40 is faster than a K20x, by around 30%).
Otherwise: I repeat, I am not presuming to know Apple's marketing better than Apple does, and you are impugning that my point is to trash talk Apple (see your comments above re Apple vs Boxx.) It is not. I am saying that the Mac Pro will not be popular in scientific computing circles, because it is not quite what is needed -- I really like the form factor of the machine, and I would like to convince myself that it is the right thing to get, but it just isn't.
This conversation has consisted of me explaining why this is true, and you explaining to me why Apple should not care.
K40 is 1.43 teraflops, K20 does 1.17:
http://www.nvidia.com/object/tesla-workstations.html
so two D700s with 870 gigaflops each is 1.74 teraflops double precision. The two D700s cost $1000, K40 retail is $5300.
You're not explaining why this is true universally. You assume that your specific software applies to the entire scientific community and make generalised statements about the Mac Pro's unsuitability for it and that Apple therefore doesn't care at all about that market segment. That's trash talk. The Mac Pro may be less optimal for that portion of the scientific community but just because they opted for AMD GPUs doesn't eliminate the Mac Pro from it entirely.
Worst case, there is the possibility of running an NVidia GPU externally over Thunderbolt:
http://forums.macrumors.com/showthread.php?t=1594407
"I've run a GeForce Titan through thunderbolt 1 on a Mac, and saw only a 5% performance drop compared to a custom bult PC with a 3770k."
For computation, if the data set is in the GPU memory, it's not going to be limited by the connection.
K40 is 1.43 teraflops, K20 does 1.17:
http://www.nvidia.com/object/tesla-workstations.html
so two D700s with 870 gigaflops each is 1.74 teraflops double precision. The two D700s cost $1000, K40 retail is $5300.
You're not explaining why this is true universally. You assume that your specific software applies to the entire scientific community and make generalised statements about the Mac Pro's unsuitability for it and that Apple therefore doesn't care at all about that market segment. That's trash talk. The Mac Pro may be less optimal for that portion of the scientific community but just because they opted for AMD GPUs doesn't eliminate the Mac Pro from it entirely.
Worst case, there is the possibility of running an NVidia GPU externally over Thunderbolt:
http://forums.macrumors.com/showthread.php?t=1594407
"I've run a GeForce Titan through thunderbolt 1 on a Mac, and saw only a 5% performance drop compared to a custom bult PC with a 3770k."
For computation, if the data set is in the GPU memory, it's not going to be limited by the connection.
That's interesting, about running an external GPU. Otherwise, I know it is shocking, but I do actually talk to (a lot of) other people, and their needs are not so different from mine. And you are right that MOST computations do not require 128GB, but if you have one that does, do you suggest one go out and buy another machine? As for trash talking, you are (for reasons that elude me, since you are not Tim Cook, as far as I know) being defensive. I have no issue with Apple largely punting on the scientific computing segment. They were never big in it, they don't really understand it, and they are wise to concentrate more on what they do know. In any case, most hardware vendors prefer to build a lot of jellybean machines -- apparently DOE had had to ask IBM really really nicely to build them a supercomputer, since it is not really worth a vendor's time. Sad, but people find workarounds.
It's not shocking that you talk to a lot of people but even if it's 1000 people, it's still a fraction of a community that you describe as being in the hundreds of thousands and I'm sure not every single one of the people you talk to entirely dismisses AMD or OpenCL. What happens if NVidia goes bankrupt? All that code would have to be rewritten.
It will still work, it'll just go into swap. Plus, in the context of GPU computing using 6GB of memory, 64GB is a lot of RAM to work with and again DDR4 will boost it back up to 128GB.
You say they don't understand it but they developed a compute language that works across multiple hardware architectures so what don't they understand? Have they designed OpenCL incorrectly?
It's not shocking that you talk to a lot of people but even if it's 1000 people, it's still a fraction of a community that you describe as being in the hundreds of thousands and I'm sure not every single one of the people you talk to entirely dismisses AMD or OpenCL. What happens if NVidia goes bankrupt? All that code would have to be rewritten.
It will still work, it'll just go into swap. Plus, in the context of GPU computing using 6GB of memory, 64GB is a lot of RAM to work with and again DDR4 will boost it back up to 128GB.
You say they don't understand it but they developed a compute language that works across multiple hardware architectures so what don't they understand? Have they designed OpenCL incorrectly?
1. All the code will have to be rewritten anyway, since the technology changes quickly.
2. "Go into swap" means "sayonara" [with SSDs the hit will not be quite as massive, but still I am guessing a couple of orders of magnitude in speed -- I would love to be proven wrong.]
3. They don't understand what the community needs. The fact that OpenCL is still not that popular is an indicator of this -- no one wants to be beholden to a single vendor, but CUDA comes with VERY nice libraries and tools. My understanding is that OpenCL does not. That said, it is certainly true that you can get Matlab or Mathematica on your Mac, and it will work quite well, but that just shunts the dependence to a different proprietary language.