What card for compute performance ?

[Midnitewarrior4]

So I've got a customer that running lots of vm's cad software, and other computational stuff. His other server has a couple quadros with xeon, this build is using dual 18 core xeon, and he said he wanted dual 980tis but I was thinking something like titan x's, what kind of card would be best for this case ? His budget is 22,000 and we're about the 18,000 mark with the 980tis it's in CAD currency btw

[Moondrelor]

So I've got a customer that running lots of vm's cad software, and other computational stuff. His other server has a couple quadros with xeon, this build is using dual 18 core xeon, and he said he wanted dual 980tis but I was thinking something like titan x's, what kind of card would be best for this case ? His budget is 22,000 and we're about the 18,000 mark with the 980tis it's in CAD currency btw

 

I'd say go with the dual 980 ti's simply because the cost.

The performance isn't much different.

[The Official Czex]

Why no Quadro' s or Tesla' s? Those things crush compute tasks.

[GodSpoon]

For CAD and VM work 980ti's would be just fine The Official Czex, they have better compute performance than an M6000 but they obviously aren't workstation rated

[MageTank]

Honestly, both Titan X's and 980 Ti's suck at compute. Maxwell is not a compute architecture. If you can find some used Titan Z's for cheap (I saw one for $400 on Ebay with one broken DVI port) that would do a lot better. If CAD/VM is all he is doing, they should work fine, but if he is doing heavy FP64, Kepler would be stronger (Original Titan, Titan Black, Titan Z, Any of the Kepler Compute cards, etc). 

[Midnitewarrior4]

I'd say go with the dual 980 ti's simply because the cost.

The performance isn't much different.

cost is not an issue with this one, so straight up what's better ? It's all being water cooled so cooling don't mater

[Midnitewarrior4]

Honestly, both Titan X's and 980 Ti's suck at compute. Maxwell is not a compute architecture. If you can find some used Titan Z's for cheap (I saw one for $400 on Ebay with one broken DVI port) that would do a lot better. If CAD/VM is all he is doing, they should work fine, but if he is doing heavy FP64, Kepler would be stronger (Original Titan, Titan Black, Titan Z, Any of the Kepler Compute cards, etc).

ya this guy definetly won't go for used components, so the original titan would be better then titan x for cad, compute and vm's ?

[Moondrelor]

cost is not an issue with this one, so straight up what's better ? It's all being water cooled so cooling don't mater

 

Then Titan Z's or Quadro M6000's

[i_build_nanosuits]

cost is not an issue with this one, so straight up what's better ? It's all being water cooled so cooling don't mater

Titan X

[GodSpoon]

Honestly, both Titan X's and 980 Ti's suck at compute. Maxwell is not a compute architecture. If you can find some used Titan Z's for cheap (I saw one for $400 on Ebay with one broken DVI port) that would do a lot better. If CAD/VM is all he is doing, they should work fine, but if he is doing heavy FP64, Kepler would be stronger (Original Titan, Titan Black, Titan Z, Any of the Kepler Compute cards, etc). 

 

Yeah I would only recommend a Titan/Black/Z if you need a lot of FP64 performance, as the architecture is better for it 

 

Just get Titan X's and call it a day Midnitewarrior4 since he doesnt use FP64 related work

 

I wouldn't watercool, since this is a workstation 

 

If something leaked...

 

[MageTank]

ya this guy definetly won't go for used components, so the original titan would be better then titan x for cad, compute and vm's ?

http://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/15

 

The current Maxwell cards CAN compute, and in single precision workloads, they are superior. However, they completely fall apart in double precision workloads (FP64) to the point where even the GTX 580 out performs the Titan X in such a workload. 

 

You just need to pick the right tool for the job. Ask him exactly what kind of compute he plans on doing, and if its more SP than DP, then he will be fine with Maxwell. 

[i_build_nanosuits]

http://www.anandtech.com/show/9059/the-nvidia-geforce-gtx-titan-x-review/15

 

The current Maxwell cards CAN compute, and in single precision workloads, they are superior. However, they completely fall apart in double precision workloads (FP64) to the point where even the GTX 580 out performs the Titan X in such a workload. 

 

You just need to pick the right tool for the job. Ask him exactly what kind of compute he plans on doing, and if its more SP than DP, then he will be fine with Maxwell. 

do you know what double precision mean? and actually how many softwares benefit from this type of computing?!

CAD softwares don't, and most softwares people actually use does not either.

[Moondrelor]

Not sure how much stuff you need, if just the GPU's or what but anyways... Here? Feel free to criticize.

Translates to 19,800 CAD

If you just need the GPU's, then two Titan Z's will put you 4500 CAD deeper. So 22,500 CAD

 

PCPartPicker part list / Price breakdown by merchant

 

CPU: Intel Xeon E5-2698 V3 2.3GHz 16-Core OEM/Tray Processor  ($3678.12 @ Amazon) 

CPU: Intel Xeon E5-2698 V3 2.3GHz 16-Core OEM/Tray Processor  ($3678.12 @ Amazon) 

CPU Cooler: Cooler Master GeminII M4 58.4 CFM Sleeve Bearing CPU Cooler  ($36.09 @ Amazon) 

CPU Cooler: Cooler Master GeminII M4 58.4 CFM Sleeve Bearing CPU Cooler  ($36.09 @ Amazon) 

Motherboard: Asus Z10PE-D16 WS SSI EEB Dual-CPU LGA2011-3 Motherboard  ($564.01 @ Amazon) 

Memory: Crucial 128GB (4 x 32GB) DDR4-2133 Memory  ($1267.36 @ Amazon) 

Memory: Crucial 128GB (4 x 32GB) DDR4-2133 Memory  ($1267.36 @ Amazon) 

Video Card: EVGA GeForce GTX Titan Z 12GB Superclocked Video Card (2-Way SLI)  ($1620.99 @ Amazon) 

Video Card: EVGA GeForce GTX Titan Z 12GB Superclocked Video Card (2-Way SLI)  ($1620.99 @ Amazon) 

Case: Silverstone GD07B (Black) HTPC Case  ($151.98 @ Newegg) 

Power Supply: EVGA SuperNOVA T2 1600W 80+ Titanium Certified Fully-Modular ATX Power Supply  ($437.98 @ Newegg) 

Total: $14359.09

Prices include shipping, taxes, and discounts when available

Generated by PCPartPicker 2015-12-16 09:59 EST-0500

[MageTank]

do you know what double precision mean? and actually how many softwares benefit from this type of computing?!

CAD softwares don't, and most softwares people actually use does not either.

I take it you did not read my post? 

 

Honestly, both Titan X's and 980 Ti's suck at compute. Maxwell is not a compute architecture. If you can find some used Titan Z's for cheap (I saw one for $400 on Ebay with one broken DVI port) that would do a lot better. If CAD/VM is all he is doing, they should work fine, but if he is doing heavy FP64, Kepler would be stronger (Original Titan, Titan Black, Titan Z, Any of the Kepler Compute cards, etc). 

I was very upfront with that IF, on both parts of that sentence. If CAD/VM is all he is doing, Maxwell will be fine. IF he is doing heavy FP64, Kepler would be stronger. That is a true statement that cannot be refuted, right? I also already gave him a link comparing performance between the two and told him to pick the right tool for the job. You've seen me before in the CPU subforum nanosuits, you already know i do not mislead people. Why assume i started now? 

[iHardware Shelden]

Why no Quadro' s or Tesla' s? Those things crush compute tasks.

Yea Quadro's and Tesla's would be a good option. 980ti's are not meant for tasks like that also assuming the systems are running under significant load 24/7.

[i_build_nanosuits]

I take it you did not read my post? 

 

I was very upfront with that IF, on both parts of that sentence. If CAD/VM is all he is doing, Maxwell will be fine. IF he is doing heavy FP64, Kepler would be stronger. That is a true statement that cannot be refuted, right? I also already gave him a link comparing performance between the two and told him to pick the right tool for the job. You've seen me before in the CPU subforum nanosuits, you already know i do not mislead people. Why assume i started now? 

...the reality is that only a handful of scientific applications actually require 64-bit floating point operations...some CAD programs might do as well but i suspect them to be VERY few and far between...and anyways even then what kind of Workstation Market would like to use 190 or 200 GFLOPs of peak FP64 performance? A 150$ CPU can do more than that. Reason why FP64 compute was striped from maxwell is that it's not useful to have on there, applications don't care about it and if/when they do your CPU will be around 5 to 10 times better at those calculations anyways.

[MageTank]

You're clueless...the reality is that only a handful of scientific applications actually require 64-bit floating point operations...some CAD programs might do as well but i suspect them to be VERY few and far between...and anyways even then what kind of Workstation Market would like to use 190 or 200 GFLOPs of peak FP64 performance? A 150$ CPU can do more than that. Reason why FP64 compute was striped from maxwell is that it's not useful to have on there, applications don't care about it and if/when they do your CPU will be around 5 to 10 times better at those calculations anyways.

I've seen this before... On another forum. Ah yes, There it is. http://www.techpowerup.com/forums/threads/why-include-fp64.199352/#post-3086889

 

You are calling me clueless, and yet you are googling reasons as to why FP64 is useless, and copying the response of someone else? It's a known fact by ANYONE that has ever done 3d modeling, that DP64 can be useful. Plenty of different CAD softwares that can utilize it. Why are you pretending otherwise? 

 

CPU's might be more accurate when it comes to FP64, but GPU's are worlds faster, and always will be. If a $150 CPU could do the job, then tell me, why is Intel investing so much time and effort into Knights Landing? God knows it's not going to be used for gaming. 

 

It's sad that the clueless one such as myself has to explain this to you, seeing as you seemed rather intelligent in the past. I'll simplify it. FP64 is used in CAD/3d Modeling, simulation, several different encoders, etc. Since OP mentioned CAD work, my asking of his use of FP64 was justified. With your logic, one could say "AVX2 is useless because no games use it". GPU acceleration is only getting used more and more as time progresses. Even encryption and compression use it these days. I eagerly await to see your half baked, googled response to this. 

[i_build_nanosuits]

snip!

https://en.wikipedia.org/wiki/Double-precision_floating-point_format

[MageTank]

https://en.wikipedia.org/wiki/Double-precision_floating-point_format

How exactly is this helping with the discussion? Also, you said FP64 was stripped from Maxwell "because it was not useful". That is simply not true. It was stripped from Maxwell because Nvidia made more room on the die for cache and other things to make it better at gaming. Take a look at the Maxwell Tesla's. Notice how NONE OF THEM are sold as compute cards anymore. They are sold as VDI/Machine Learning cards. 

 

[The Official Czex]

For CAD and VM work 980ti's would be just fine The Official Czex, they have better compute performance than an M6000 but they obviously aren't workstation rated

 

Also, workstation load times need to be counted as a factor. Quadro' s and Tesla' s are rated for much longer 100% load endurance compared to a consumer card like the 980 Ti.

[LokiFire]

GTX 580

TITAN

TITAN Black

TITAN X

[i_build_nanosuits]

i must admit, upon further reading on the subject it does looks like some CAD applications might use FP64 computing in some modules depending on what you do with them..if the precision is required
so it might be worth it to select an older card that has an Fp64 block on the die...being aware he would be making a huge sacrifice in FP32 performance though, obviously...and energy efficiency which should not be an issue i assume for such an high grade workstation machine.

Also, depending on the application the double precision computing can be done by the CPU, it will not be as fast as it can be GPU accelerated.

 

It would also be worth looking into what softwares he use exactly and wheter or not they can use OpenCL acceleration or CUDA.
If his softwares are mostly supporting OpenCL and not CUDA then an AMD card might be a better choice.
Cuda when supported is faster than OpenCL, Nvidia cards can do both OpenCL accleration (not as fast as AMD cards) and CUDA...AMD can only do OpenCl.

[MageTank]

i must admit, upon further reading on the subject it does looks like some CAD applications might use FP64 computing in some modules depending on what you do with them..if the precision is required

so it might be worth it to select an older card that has an Fp64 block on the die...being aware he would be making a huge sacrifice in FP32 performance though, obviously...and energy efficiency which should not be an issue i assume for such an high grade workstation machine.

Also, depending on the application the double precision computing can be done by the CPU, it will not be as fast as it can be GPU accelerated.

 

It would also be worth looking into what softwares he use exactly and wheter or not they can use OpenCL acceleration or CUDA.

If his softwares are mostly supporting OpenCL and not CUDA then an AMD card might be a better choice.

Cuda when supported is faster than OpenCL, Nvidia cards can do both OpenCL accleration (not as fast as AMD cards) and CUDA...AMD can only do OpenCl.

That is exactly why I asked him if he intended on using an application that supported FP64. If he did not, Maxwell would be the better choice. I firmly believe in picking the right tool for the job. I do not know what he intends on doing, which is why I gave him both options, and left the final decision up to him. If his customer uses an application that can be accelerated using FP64, then a Kepler card might be a great option. Since his customer does not want a used card however, he might not be able to find one. Maxwell would work fine still too, and there are plenty of options available.

 

CPU's are catching up in compute. AVX2 (256-bit) allows for 4 FP64 operations on each register, at the same time. It's why i mentioned AVX2 before. At the end of the day, it's up to OP choose the tool that will best serve his customer. 

[GodSpoon]

Also, workstation load times need to be counted as a factor. Quadro' s and Tesla' s are rated for much longer 100% load endurance compared to a consumer card like the 980 Ti.

 

Thats what I was referring to when I said workstation rated

[i_build_nanosuits]

That is exactly why I asked him if he intended on using an application that supported FP64. If he did not, Maxwell would be the better choice. I firmly believe in picking the right tool for the job. I do not know what he intends on doing, which is why I gave him both options, and left the final decision up to him. If his customer uses an application that can be accelerated using FP64, then a Kepler card might be a great option. Since his customer does not want a used card however, he might not be able to find one. Maxwell would work fine still too, and there are plenty of options available.

 

CPU's are catching up in compute. AVX2 (256-bit) allows for 4 FP64 operations on each register, at the same time. It's why i mentioned AVX2 before. At the end of the day, it's up to OP choose the tool that will best serve his customer. 

indeed...it's like the ''perfect'' compute card does not exist...some are great for openCL, some are great for CUDA...some can blast through FP32 instructions but will low down to a crawl in dual-precision compute (still believe FP64 is mostly irrelevant for MOST use case though...it does seem that not many ''normal'' tasks and applications will require you to crunch massive FP64 instruction sets)

All in all, OP if you don't know what he's doing, and want something that will not be particularly brilliant in any ways, but support FP64 acceptable level, FP32 better than average, OpenCL support average, Cuda performance better than average: Titan Black or TitanZ

Something that will crush FP32 and be very efficient: TitanX/980ti

Something for OpenCL and FP64/FP32: AMD R9 FuryX or a firepro of some sort i would assume (i don't know firepro's)