Anyone still using Alpha 12.1?

[gbrak30]

Am I the only person still running alpha 12.1???  In my opinion, the physics/collisions run MUCH better than newer versions.  The fact that I am able to select the Compute device may have something to do with it.  I have an SLI setup, so in the control panel I set up US2's profile as follows:
SLI Rendering Mode: Singe GPU
CUDA Devices: manually select GPU #2 as CUDA Device and disable GPU #1
With these settings, I can have massive collisions with no freezing, hickups, etc. 
Also, I noticed that creating fragment grids in 12.1 works better than newer versions.  I feel like a broken record here because I've been saying this over and over again.

It seems like the focus of the development team is slowly moving away from the physics engine and more towards visuals.  I guess I'm the minority that cares more about the physics than visuals.

[Physics_Hacker]

Am I the only person still running alpha 12.1???  In my opinion, the physics/collisions run MUCH better than newer versions.  The fact that I am able to select the Compute device may have something to do with it.  I have an SLI setup, so in the control panel I set up US2's profile as follows:
SLI Rendering Mode: Singe GPU
CUDA Devices: manually select GPU #2 as CUDA Device and disable GPU #1
With these settings, I can have massive collisions with no freezing, hickups, etc. 
Also, I noticed that creating fragment grids in 12.1 works better than newer versions.  I feel like a broken record here because I've been saying this over and over again.

It seems like the focus of the development team is slowly moving away from the physics engine and more towards visuals.  I guess I'm the minority that cares more about the physics than visuals.

The reason for that is the commonpeople, and thus the majority, want better visuals: any common person is going to just sit there and crash planets together, and all that matters with that really is visuals. While some of us actually care about the physics, but that's the minority. They do still work on polishing the engine, but adding new features is what's going to keep people on the train, now.

[Prometech]

I recently went back to 12.1 for similar reasons, and in 14, the total body fragmentation wasn't disabled or horribly optimized.

[gbrak30]

I guess its all about money at this point...I reckon they're going to charge full price for a final, visually stimulating product....

[Jar]

It's true that we do want to make it look better and better, but it's not true that this is taking priority over anything else... We want to continue making everything better.

I can't speak too much to the technical side of things, but I know that for many users Alpha 15 was a huge performance boost for them due to the physics rewrite... Thomas is still our dedicated physics programmer, and for a while now he has been working on some further physics optimizations, which should be ready soon. This should only make things faster.

In general, it may just be harder to notice physics changes than something more eye-catching like visual changes, especially when it comes to performance and stability tweaks which don't really draw attention. But that doesn't mean that we're not continually working to optimize it.

[gbrak30]

I noticed the "Managed" option as a Compute Device, does that mean the program balances/shifts load between CPU<>GPU? I'm currently using a GTX 980 Ti with alpha 12.1 with max attracting/non-attracting objects each set to 1024.  The performance is amazing!  What I personally like most is the amount of bodies I can run at once, all interacting with each other, just couldn't get that with newer versions.  One thing I do is crash stars into each other, with the debris from the collisions condensing from fragment to body and falling back down to the star, creating more debris, like drops of water. 

[Greenleaf]

In my opinion, the physics/collisions run MUCH better than newer versions.  The fact that I am able to select the Compute device may have something to do with it.  I have an SLI setup, so in the control panel I set up

Previously we had support for computation in managed code https://en.wikipedia.org/wiki/Managed_code as well as using opencl https://en.wikipedia.org/wiki/OpenCL . The nbody code had, however, drifted too far away from the initial design and needed a rewrite. This rewrite increased performance for most people, but it removed opencl and placed all computation in managed code. For a select few with a much more powerful gpu setup (than cpu) this could have made performance worse.


The plan was to build back different computation modes, which perform better, such as a native computation library (not using .net/mono) and gpu computing using C++ AMP. The native code has been 98.5%done for a long time and recently completed. Tests show a very high general performance increase of the NBody computations, from managed code.


Other physics related improvements (entirely new and interesting features, actually) are also being added right now and when that is done GPU computing will likely come back.


As to us not caring about physics.... bah! ;-)

[vmorgo]

Any thoughts/comments/recommendations as to when the GPU compute capabilities are coming back?

Also:  Suggestions as to which compute hardware would best benefit Universe Sandbox?

I'm looking at one of the following 4 possibilities:
1.  Obtaining a used NVidia Tesla board for dedicated GPU-base computing with CUDA or possibly OpenCL (if supported)
2.  Obtaining a fairly high-end ATI graphics board
3.  Obtaining a faster system perhaps a 32-core AMD system (I keep hearing AMD is much better at the types of computations used by this software than Intel is)
4.  Obtaining a Xeon PHI with either 50 or 61 cores.

Options 1, 2, and 4 can be accomplished for less than $500 through eBay.  Option 3 is much more expensive and may result in my death at the hands of my wife.

NOTE:  I happen to be using Mint 17.1 "Rebeccah", which is based on the 14.04 LTS version of Ubuntu Linux. 

I know this sounds "personal", but I think enough people are asking about which GPU to use, single vs. double-precision computation, array computation and other performance-related questions that this is worth asking in a public forum.

[Greenleaf]

It is flattering that you consider buying hardware specifically to work better with Universe Sandbox.


The first iteration of adding back GPU computing should not be far away, but I can not make any promises. I can, however, say that it is something I have been looking forward to reimplementing. This is Initially windows only, though.


The GPU computing will come in the form of C++ AMP which means that a Tesla board should be supported (except K20, it seems) http://blogs.msdn.com/b/nativeconcurrency/archive/2013/09/27/nvidia-tesla-s-support-for-c-amp.aspx 


A high end graphics card could help a lot, except that many modern cards have rather bad support for double calculations https://en.wikipedia.org/wiki/Double-precision_floating-point_format and we are currently using doubles for all our calculations. Therefore a good GPU for Universe Sandbox should have good double performance. This is another issue which will have to be investigated, since floating point is really a bad data type for our purpose for multiple reasons, especially because the resolution is not the same everywhere in space, so we may change the datatype and this can change the double requirement as well.


A high speed many core processor system will generally linearly scale performance. Double the computing power and you could the calculation performance. This is a sure way to get better performance.


I do not think a Xenon PHI will work out of the box with software not written for it (?), so that will probably not give an improvement. I have not really any insight into this device, though.


So in summary, I would say that a Tesla card will help, but would be limited in its purposes for most users while a powerful cpu would certainly help and would probably give better value for the money in other applications as well.

[tepidbread]

An old r9 280x or 290x should work great. They have 1024 and 704 Gflops for FP64 (double floating point) respectively. They should be relatively cheap too. To put this into perspective a GTX 980 only has 156 Gflops FP64. Both AMD and Nvidia have gimped their newer cards' FP64 performance in favor of efficiency and graphics performance. The theoretical FP64 performance of a haswell 8 core cpu is close to 500 Gflops. (I think that the intel haswell 8 core chips are like $600.) I don't know how AMD's server chips preform. I would expect AMD's server chips to yield less performance than the intel chips because their cores are in a module configuration where pairs of cores share floating point units and lack dedicated ones. Last I checked Amd's server market was getting crushed by Intel's server and enthusiast series of processors.

I think your best bet is to go for an older gpu Fermi or GCN 1.0/1.1.

I hope it helps...

[vmorgo]

Thank you so much for your reply!  I actually already have a machine dedicated to UniverseSandbox 2.  It is a Dell PC with 16 GB RAM and a 3.4 GHz Intel I-5 CPU.  Upgrading it from the original Intel graphics 4000 (shared memory model) to the nVidia GeForce 730 has made a world of difference for about $60.  I'm figuring on a similar boost in performance if I can add a second NVIDIA or an ATI card (such as the R9-280x or R9-290x).  I was thinking of a Tesla board if I can fit it into this computer's case, otherwise, maybe a GeForce 560 Ti.

I'd consider the Radion card, but don't think it can coexist with the GeForce cards...

However, I am running this on Mint 17.1, which is basically Ubuntu 14.04.  If you could give me some idea as to when you would be moving forward with the Linux implementation, it would be helpful, but I understand if you don't know yet....  I envy you the ability to program for this in the first place!  Perhaps, I could even assist in some way....  Please let me know....

Again, thanks!

[vmorgo]

Many thanks for your response! 

Just to make sure I understand:  The older R9-280x/290x would be preferred even over a tesla card such as the nVidia M2090 (Anything other than the K20)? 

Wondering when the Linux version will be--want to make sure I have the right card ready when you're ready...would it be better to have one or multiple GPU's?  I am assuming multiple of the same type (or at least the same manufacturer) is preferred.

Again, many thanks! 

(I'll work on seeing if my wife will let me get the multi-core CPU....  )

Thank you!

A./

[tepidbread]

I failed to notice that you were running linux and are running a Dell computer.

Linux is known for its bad Amd drivers. I cant seem to find anything on how the nvidia Tesla cards stack up against other graphics cards. Considering that Tesla cards are for the server market. They are really likely to have a lot lower performance per watt ratio. The r9 280x or the more expensive r9 290x will be much faster than your current graphics card for graphics rendering even with bad driver support. They will also crush all the  double floating point tasks that you can throw at it. I have heard that the driver situation with Amd on linux is improving since AMD separated its CPU and GPU branches.

http://www.geeks3d.com/20140305/amd-radeon-and-nvidia-geforce-fp32-fp64-gflops-table-computing/

This link is where I am trying to find the relative compute performance for each card.

My second concern is your power envelope and room in your case. Do you have a power supply that can handle an 250 watt graphics card enough room to fit one into your case? You may need a new power supply for either a tesla gpu or a r9 290x or r9 390x.

Sorry, I just can't find any benchmarks on Tesla cards in order to compare them to the really good amd cards. I tried to consider things that are within your price envelope.

This was really no problem for me. I love this kind of stuff. I am in college to become an electrical engineer after all.

Edit: Regarding some of your other questions.
 
It is generally better to go for one powerful card over several weaker cards. For example I have 2 GTX 660s in my system and the second one would not see any utilization. Greenleaf has said that he planned on adding multiple gpu computing but it seems to be a ways off.

It is possible for Amd and Nvida cards to coexist. People used to do it all the time. In the case of universe sandbox one could do the computing and the other could do the rendering. However, Their would be no point to doing this because a more powerful gpu (excluding the tesla models) have more than ample resources left over for rendering that are not being utilized by the compute process.

Also upgrading to a multicore cpu seems like that would involve upgrading your entire system or something. If that is the case I would just suggest building your own computer. Sorry for the wall of text.

[vmorgo]

Well, hello, TepidBread--and everyone else!

Just rebuilt the system.  New server... $285 off eBay.  Dedicated HP Z600 12 GB Workstation with dual hyperthreaded Xeon 5645's (6 cores each) running at 2.4GHz.  May not sound like all that much, until I mention that, with hyperthreading, we're showing 24 cores on this beast!  Still using the old nVidia 730, though (I think it has only 96 graphics cores, but it might be the one with 384.).  Moved the 2 TB drive and the NVIDIA out of the Dell into this machine this AM.   Tesla M2090 6GB (I picked that up for $85 off eBay) is on the way.  650 watt P/S should handle that okay....

I DID remember that Linux ATI support is not that great, so I went with the nVidia (which works with Solaris, and I may eventually switch over to that anyway).  Anywho:  UniverseSandbox 2 doesn't seem to be able to get the attention of more than three to six of my 24 cores at any one time.  It will bog down at times and doesn't seem to want to take advantage of all those CPU cores.

Now, I did do one possibly bad thing:  I just moved the disk drive to the new machine WITHOUT either reinstalling Linux or reinstalling the UniverseSandbox2/Steam app.  The old machine had an i5 with 4 cores.  The new one has two (slighly older) Xeon 6-cores CPUs.  The old machine ran at 3.4 GHz with no hyperthreading.  The new one has hyperthreading, but because of the higher core count, it runs at only 2.4 GHz.

So, perhaps all my trouble is because I didn't "reinstall linux"/recompile the kernel.

On the other hand, BOINC/Seti At Home runs just fine on this rig and it DOES see and run on all 24 cores.

In fact, I can run both Seti@home and UniverseSandbox 2 at the same time on this rig with only a slight hit in performance (the GUI slows and gets just a bit choppy, but it's not bad, and since BOINC runs the CPU's at full throttle, it actually seems to speed UBox2 up a bit.)

I have disabled TurboBoost2 and ExtendedTurboBoost2, based on recommendations I found elsewhere (neither seems to do much of anything other than make the CPUs run hot).

Any suggestions as to how I might take better advantage of this new system would be appreciated!

Anyone who wants to suggest experiments/research or submit models for high-speed processing are also welcome to drop me a line...