Maxwell or Pascal for math/physics simulations

iceblitzed

Distinguished
Hello, i've heard pascal is going to be released feb 2016 (if I'm correct)? I've been wanting to build a computer mainly for math/physics simulation that take advantage of cuda cores in nvidia gpus. Is it worth it waiting for pascal since it (might) be released in six months or will the performance boost not be worth it and I shouldn't play the wait and see game.
 


iceblitzed,

NVIDIA seems to follow a pattern of teasing: "Pascal is going to have three times more double precision per Watt" but not the actual relative performance, not the release, date, nor the price.

Seeing that in the Quadro world- which I assume is your interest- the $5,000 M6000 is the only Maxwell GPU, and the M4000 and M5000 are known to be arriving sort of sometime, but details are speculative, I would agree with geofelt and obtain the performance you need now. If even the Maxwell GPU's are sketchy, Pascals as promising they may be, but promises don't crunch numbers now.

If I had a reasonably healthy budget and wanted to run something like MATLAB, I think I'd put a pair of used Quadro K5200's in SLI ( about $2,500) and have 2X GK180GL, 256-bit, 16GB, 4608 CUDA cores and a total double precision and TFLOPS I'd expect could exceed a Quadro K6000 (384-bit, 12GB, 2880 CUDA) and not a million miles off an M6000 (12GB 3072 CUDA) for half the price. As the K5200 are already reasonably well depreciated, as geofelt alludes, that level of GPU is likely to make a good contribution to a Pascal when they arrive and if they represent a significant improvement. If the trend represented by Kepler and to the M6000 is representative, I think Pascal may be quite remarkable, but I'm guessing that they will be remarkable in Summer 2016 since the M4000 and M5000 are still fetoid. > Notice how many if's are involved at this point whereas we can see the very high performance of the K5200 and K6000.

The M6000 is quite amazing by the way. At my local wind tunnel, I saw a presentation / demonstration for a complex device that at first appeared to be a fantastic sort of advertising rendering that could spin around in real time remarkably easily but then be sliced in sections to reveal that every nut and bolt annotated and fully realized and with the systems could run gas flow and thermal simulations.

May I ask what programs you're using, kinds of problems, and your current system?

Cheers,
BambiBoom

1. HP z420 (2015) > Xeon E5-1660 v2 six-core @ 3.7 / 4.0GHz > 32GB DDR3 ECC 1866 RAM > Quadro K4200 (4GB) > Intel 730 480GB (9SSDSC2BP480G4R5) > Western Digital Black WD1003FZEX 1TB> M-Audio 192 sound card > 600W PSU> Logitech z2300 > Linksys AE3000 USB WiFi > 2X Dell Ultrasharp U2715H (2560 X 1440) > Windows 7 Professional 64 >
[ Passmark Rating = 5064 > CPU= 13989 / 2D= 819 / 3D= 4596 / Mem= 2772 / Disk= 4555] [Cinebench R15 > CPU = 1014 OpenGL= 126.59 FPS] 7.8.15

Pending upgrade: HP /LSI 9212-4i PCIe SAS /SATA HBA RAID controller, 2X Seagate Constellation ES.3 1TB (RAID 1)

With Quadro K2200:
[ Passmark Rating = 4918 > CPU= 13941 / 2D= 823 / 3D= 3463 / Mem= 2668 / Disk= 4764

2. Dell Precision T5500 (2011) > Xeon X5680 six -core @ 3.33 / 3.6GHz, 24GB DDR3 ECC 1333 > Quadro K2200 (4GB ) > Samsung 840 250GB / WD RE4 Enterprise 1TB > M-Audio 192 sound card > Linksys WMP600N PCI WiFi > 875W PSU > Windows 7 Professional 64> HP 2711x (1920 X 1080)
[ Passmark system rating = 3490 / CPU = 9178 / 2D= 685 / 3D= 3566 / Mem= 1865 / Disk= 2122] [Cinebench 15 > CPU = 772 OpenGL= 99.72 FPS] 7.8.15

with Quadro K4200:
[ Passmark system rating = 3585 / CPU = 9346 / 2D= 683 / 3D= 4708 / Mem= 1850 / Disk= 2202]

With Quadro 4000
[ Passmark system rating = 3339 / CPU = 9347 / 2D= 684 / 3D= 2030 / Mem= 1871 / Disk= 2234]
 

iceblitzed

Distinguished
Good points, I was definitely wrong about the expected release date.
What programs I am using: LAMMPS Molecular Dynamics Simulator (lammps.sandia.gov)
Kinds of problems: Granular Flow down an inclined plane simulations
Current system? Machine with i5-3340, no gpu, I am not the owner of this machine have limited access to it
Other Notes: I was looking for a brand new system that costs about $1000 buying the parts and constructing it at home. I don't have the money for a quadro unfortunately. I am planning on using this system for regular use such as working with documents, browsing, etc. as well.
 


iceblitzed,

When looking for good cost /performance solutions, it's essential to know the details of use and a budget. If I were making a system for LAMMPS, NAMB, or MATLAB, I'd start with a budget of $7,000 or so. This is because those applications are some of the very few that in the hands of the clever can use all the cores /threads available. Still, those uses also require high double precision, and benefit from CUDA coprocessing- all terms that translate as expensive.

Still, it's possible to build a reasonably competent system for $1,000. Here's a workstation idea from a few months ago:

BambiBoom Cadagrapharific WalletJoyScream TurboSignature 9000 $$©& £℞™_5.22.15

CPU: Xeon Intel E3-1231 v3 4-core @3.4 / 3.8GHz 8MB L3 Cache 80W Server Processor BX80646E31231V3 >$242

CPU Cooler: Cooler Master Hyper 212 EVO - CPU Cooler with 120 mm PWM Fan > $40

Motherboard:ASUS SABERTOOTH Z97 MARK2 LGA 1150 Intel Z97 HDMI SATA 6Gb/s USB 3.0 ATX Intel Motherboard > $170

Memory: Crucial Ballistix Sport 8GB 240-Pin DDR3 SDRAM DDR3 1600 (PC3 12800) Desktop Memory Model BLS8G3D1609DS1S00> $72

GPU: EVGA NVIDIA GeForce GTX 750 Ti 2GB GDDR5 DVI/HDMI/DisplayPort PCI-Express Video Card > $138.

PSU: CORSAIR CXM series CX600M 600W ATX12V v2.3 SLI CrossFire 80 PLUS BRONZE Certified Modular Active PFC Power Supply > $70.

Western Digital WD Blue WD10EZEX 1TB 7200 RPM 64MB Cache SATA 6.0Gb/s 3.5" Internal Hard Drive Bare Drive - OEM > $60

Case: LIAN LI PC-9NB Black Aluminum ATX Mid Tower Computer Case > $100

Operating System: Microsoft Windows 7 Professional SP1 64-bit - OEM > $138

______________________________________________________________________

TOTAL = $1000.

Probably quite good performance, and adding a $100 SSD would give it some fizz.

The problem is that an LGA1150 system is that while Xeon E3's are great, they're limited to 4-cores and a system with potential for many more cores is more future-looking.

But, the best cost performance for this use is to upgrade a used, dual CPU capable workstation. I've done this three times, the most recent being a Dell Precision T5500.

Purchased for $171:

Dell Precision T5500 (2011) (Original): Xeon E5620 quad core @ 2.4 / 2.6 GHz > 6GB DDR3 ECC Reg 1333 > Quadro FX 580 (512MB) > Dell PERC 6/i SAS /SATA controller > Seagate Cheetah 15K 146GB > Windows 7 Professional 64-bit
[ Passmark system rating = 1479 / CPU = 4067 / 2D= 520 / 3D= 311 / Mem= 1473 / Disk= 1208]

I had the GPU, SSD, and HD around, and bought the CPU and RAM used:

Xeon X5680 > $200
RAM > $120 (24GB, DDR3-1333 ECC registered)
GPU > $150 (NVIDIA Quadro 4000 2GB)
SSD: 250GB > $70 (Samsung 840)
HD::> 70 New ( Western Digital RE4 1TB)
RAID controller > $60 (PERC H310)
OS: > $0 ( Provided by Dell)

System______________ : $190 (incl. shipping)
Other parts cost / value__ $670

____________TOTAL = $860

The performance of the the revised system is very good- on Passmark this systems is the highest performing T5500 with a single CPU of 297 tested:

Dell Precision T5500 (2011)(Revised) > Xeon X5680 six -core @ 3.33 / 3.6GHz, 24GB DDR3 ECC 1333 > Quadro 4000 (2GB ) > Samsung 840 250GB / WD RE4 Enterprise 1TB > M-Audio 192 sound card > Linksys WMP600N PCI WiFi > 875W PSU > Windows 7 Professional 64> HP 2711x (1920 X 1080)
[ Passmark system rating = 3339 / CPU = 9347 / 2D= 684 / 3D= 2030 / Mem= 1871 / Disk= 2234]

The PERC H310 controller is not yet installed, but will significantly improve the disk system as it's 6GB/s instead of the original 3GB/s. I did recently changed the GPU to a Quadro K2200 left over from upgrading my main system:

[ Passmark system rating = 3490 / CPU = 9178 / 2D= 685 / 3D= 3566 / Mem= 1865 / Disk= 2122] [Cinebench 15 > CPU = 772 OpenGL= 99.72 FPS] 7.8.15

And the idea is that in the future I can add a second 6-core CPU for 12 cores / 24 threads for about $350 total, as well as a faster GPU- and/or used Telsa GPU co-processor.

These used Precisions as they have serve quality components and are designed for continuous full performance running have been completely reliable over years of constant use. This approach also saves researching, ordering, assembling, and trouble shooting from each component- you just plug in the CPU, RAM, GPU, and drives. It might be possible to find a system that you could put into use right away- I could have done that wit hthe T5500 and then upgrade as you find the right parts- it's never out of use for more than an hour or so. Also, as these systems are fully depreciated- the Xeon X5680 I bought for $200 cost $1,600 new, they will hold their value well. I'm certain I could sell the T5500 for my investment after a year's use and in effect have the use for $0- only the time and effort.

So, there are a couple of alternatives.

Cheers,

BambiBoom

HP z420 (2015) > Xeon E5-1660 v2 six-core @ 3.7 / 4.0GHz > 32GB DDR3 ECC 1866 RAM > Quadro K4200 (4GB) > Intel 730 480GB (9SSDSC2BP480G4R5) > Western Digital Black WD1003FZEX 1TB> M-Audio 192 sound card > 600W PSU> Logitech z2300 > Linksys AE3000 USB WiFi > 2X Dell Ultrasharp U2715H (2560 X 1440) > Windows 7 Professional 64 >
[ Passmark Rating = 5064 > CPU= 13989 / 2D= 819 / 3D= 4596 / Mem= 2772 / Disk= 4555] [Cinebench R15 > CPU = 1014 OpenGL= 126.59 FPS] 7.8.15

Pending upgrade: HP /LSI 9212-4i PCIe SAS /SATA HBA RAID controller, 2X Seagate Constellation ES.3 1TB (RAID 1)