Sign in with
Sign up | Sign in

OpenCL And CUDA Are Go: GeForce GTX Titan, Tested In Pro Apps

OpenCL And CUDA Are Go: GeForce GTX Titan, Tested In Pro Apps
By

We initially had trouble getting the GeForce GTX Titan to work with OpenCL and CUDA. Finally, though, there are drivers available that fix all of that. Now we can figure out if the Titan makes a good workstation-oriented alternative to Nvidia's Quadros.

We covered Nvidia's still-new GeForce GTX Titan in Nvidia GeForce GTX Titan 6 GB: GK110 On A Gaming Card and Benchmarking GeForce GTX Titan 6 GB: Fast, Quiet, Consistent. As a gaming product, we know it to be the fastest single-GPU board you can buy. But how does the vaunted Titan fare in professional applications? It wasn't possible to run a number of tests for the launch because Nvidia's drivers weren't working in most of the non-gaming titles we tried.

Nevertheless, you'd think that, given its GK110 GPU, first introduced on a couple of Nvidia's Tesla accelerator boards, the GeForce GTX Titan would be a shoo-in for a market that doesn't flinch at $1,000 graphics cards. So, we're taking it, along with a number of other desktop-oriented graphics cards (like the Radeon HD 7970, Radeon HD 6970, GeForce GTX 680, and GeForce GTX 580) to see how the last two generations of flagship gaming products handle workstation-class software.

We're using a Titan card that Gigabyte sent over. It's based on Nvidia's reference design, though Gigabyte does throw in some extras to set its offering apart. There's a large mouse pad, a deck of playing cards, some cables, and obligatory adapters. 

The previous-gen processor in our test bed was swapped out in favor of an overclocked Core i7-3770K to help minimize platform bottlenecks. Getting to the point where we didn't see application performance change based on processor performance took a clock rate of 4.6 GHz, which just goes to show that older software is still CPU-limited. Optimizations for threading, CUDA, and OpenCL are playing a larger role in rendering tasks, but some workloads still aren't being parallelized.

Benchmark System
CPU
Intel Core i7-3770K (Ivy Bridge), 22 nm, 4C/8T, 8 MB Shared L3 Cache, Hyper-Threading Enabled, Overclocked to 4.6 GHz
RAM
32 GB Corsair Dominator Platinum @ 2,066 MT/s
Motherboard
Gigabyte G1.Sniper 3, Intel Z77 Express
SSD
2 x Corsair Neutron 480 GB
OS  
Windows 7 Ultimate x64 (Fully Patched)
Drivers
GeForce 314.22 WHQL
Catalyst 13.3 Beta 3


We already know what happens when the Tesla's GK110 GPU is tossed into a gaming environment. So, what happens when we put that same hardware to work in a professional sense?

Today's story also serves as a preview for a big workstation graphics card round-up we have coming up with all of the new Kepler-based Quadro cards. We're going to use the same benchmarks (and a lot more) to compare two generations of Nvidia and AMD offerings. Right now, we're still sorting out some driver issues that show why it's so important for these companies to seek out certifications for their premium products. You'll see us add the results from these gaming cards to that piece, too.

Display 86 Comments.
This thread is closed for comments
Top Comments
  • 18 Hide
    k1114 , April 16, 2013 9:23 PM
    Why are there not workstation cards in the graphs?
  • 10 Hide
    DoDidDont , April 17, 2013 3:57 AM
    I use 3ds max mainly for ultra high detail hard surface 3D modelling, rigging and production rendering.

    These results prove that the Titan is more than twice as fast as the rest of the cards tested here when it come to rendering in CUDA based app’s like iray, blender and most likely V-ray too. Even if the Titan turned out to be on par with the older gen GTX cards, which it didn’t, the 6GB of onboard memory for me is an absolute must. My current 3GB GTX 580’s are almost maxed out on Vram because of the high level of detail I model at, and that’s before texturing, so the alternative is buying an older generation card like the quadro 6000 or Tesla c2075 at over twice the price of the Titan, or spending more than three times the price of the Titan on the newer Tesla K20/X.

    * Good view port performance.
    * Great gaming performance.
    * More than twice as fast as the older GTX cards in CUDA based production rendering.
    * 6Gb of onboard memory for huge data sets, at less than half the price of an older 6GB Quadro/Tesla cards.

    This card is a win win win for the apps I use, so the “its price is just too high for the performance it offers in professional applications” remark, is completely wrong.

    Would you rather spend £7600 on 4x older 6gb Tesla cards in your render node, or spend £3300 on 4x Titan’s and get over twice the performance, do the math.

    I understand the advantages of Quadro/Tesla cards, optimised drivers, higher yield chips, better stability, durability, but using the GTX 480 vs. Quadro 6000 as an example up to 30% extra view port performance, but over 700% in cost, from a business point of view the math just doesn’t add up.

    I have owned Quadro cards in the past, and always ended up being disappointed by the very slight view port performance increase over the desktop equivalent, and feeling I have just wasted a lot of cash for nothing. One of my mechanical models I am working on has over 30 million polygons so far, and the GTX 580 throws it around the view ports with ease.

    For gaming yes this card is over priced, and you are better off getting a cheaper sli/crossfire configuration, but for some professionals that need fast render times and working on large data sets, this card will be a much cheaper and faster option than spending a lot more cash on Quadro/Tesla cards.

    I already ordered two Titans this morning. I will order another two when my other kidney sells on Ebay.



Other Comments
  • 18 Hide
    k1114 , April 16, 2013 9:23 PM
    Why are there not workstation cards in the graphs?
  • 1 Hide
    bit_user , April 16, 2013 9:31 PM
    Thanks for all the juicy new benchmarks!

    BTW, I'm hoping the OpenCL benchmarks all make it to the GPU Charts. I'd like to know how the HD 7870 stacks up, at least. Being a new owner of one, I'm pleased at the showing made by the other Radeons. I had expected Titan to better on OpenCL, based on all the hype.
  • -8 Hide
    bit_user , April 16, 2013 9:35 PM
    k1114Why are there not workstation cards in the graphs?

    Because it would be pointless. They use the same GPUs, but clocked lower and with ECC memory.

    The whole point of Titan was to make a consumer card based on the Tesla GPU. I don't think AMD has a separate GPU for their workstation or "SKY" cards.
  • 3 Hide
    crakocaine , April 16, 2013 9:46 PM
    Theres something weird with your ratGPU gpu rendering openCL test results. you say lower is better in seconds but yet the numbers are arranged to make it look like more seconds is better.
  • 3 Hide
    mayankleoboy1 , April 16, 2013 9:49 PM
    Too much data here for a proper conclusion. Here is what i conclude :

    In Pro applications :

    1. 7970 is generally quite bad.
    2. Titan has mixed performance.
    3. Drivers make or break a card.

    In more consumer friendly 'general' apps :

    1. 7970 dominates. Completely.
    2. 680 is piss poor (as expected)
    3. 580 may or may not compete.
    4. Titan is not worth having.

    AMD needs to tie up moar with Pro app developers. Thats the market which is ever expanding, and will bring huge revenue.

    Would have been interesting to see how the FirePro version of 7970 performs compared to the HD7970.
  • 6 Hide
    Yuka , April 16, 2013 9:51 PM
    bit_userBecause it would be pointless. They use the same GPUs, but clocked lower and with ECC memory.The whole point of Titan was to make a consumer card based on the Tesla GPU. I don't think AMD has a separate GPU for their workstation or "SKY" cards.


    It isn't pointless, since it helps put into perspective where this non-pro video card stands in the professional world. It's like making a lot of gaming benchmarks out of professional cards with no non-pro cards. You need perspective.

    Other than that, is was an interesting read.

    Cheers!
  • 2 Hide
    slomo4sho , April 16, 2013 10:04 PM
    I would honestly have liked to see a GTX 690 and 7990(or 7970 x-fire) in the mix to see how titan performs at relatively equal price points.
  • 2 Hide
    tiret , April 16, 2013 10:27 PM
    I find the rendered scenes more interesting than the graphs.
  • 0 Hide
    SuperGamerBoy , April 17, 2013 12:12 AM
    I would likd to see the Ares II in action :D 
  • 2 Hide
    Cryio , April 17, 2013 2:16 AM
    Couldn't they just, you know, mod some professional-oriented GPU drivers to work on the Titan and then show us some results?

    I've been running Quadro drivers on my 9600GT & 560 Ti for years now.
  • 2 Hide
    softplacetoland , April 17, 2013 3:46 AM
    260511Show me one benchmark where AMD actually does well? the amount of fanboyism in your comment is unsettling, go back to your cave, Troll.


    You aren't quite settled with what troll should be, are you? :trollface
  • 10 Hide
    DoDidDont , April 17, 2013 3:57 AM
    I use 3ds max mainly for ultra high detail hard surface 3D modelling, rigging and production rendering.

    These results prove that the Titan is more than twice as fast as the rest of the cards tested here when it come to rendering in CUDA based app’s like iray, blender and most likely V-ray too. Even if the Titan turned out to be on par with the older gen GTX cards, which it didn’t, the 6GB of onboard memory for me is an absolute must. My current 3GB GTX 580’s are almost maxed out on Vram because of the high level of detail I model at, and that’s before texturing, so the alternative is buying an older generation card like the quadro 6000 or Tesla c2075 at over twice the price of the Titan, or spending more than three times the price of the Titan on the newer Tesla K20/X.

    * Good view port performance.
    * Great gaming performance.
    * More than twice as fast as the older GTX cards in CUDA based production rendering.
    * 6Gb of onboard memory for huge data sets, at less than half the price of an older 6GB Quadro/Tesla cards.

    This card is a win win win for the apps I use, so the “its price is just too high for the performance it offers in professional applications” remark, is completely wrong.

    Would you rather spend £7600 on 4x older 6gb Tesla cards in your render node, or spend £3300 on 4x Titan’s and get over twice the performance, do the math.

    I understand the advantages of Quadro/Tesla cards, optimised drivers, higher yield chips, better stability, durability, but using the GTX 480 vs. Quadro 6000 as an example up to 30% extra view port performance, but over 700% in cost, from a business point of view the math just doesn’t add up.

    I have owned Quadro cards in the past, and always ended up being disappointed by the very slight view port performance increase over the desktop equivalent, and feeling I have just wasted a lot of cash for nothing. One of my mechanical models I am working on has over 30 million polygons so far, and the GTX 580 throws it around the view ports with ease.

    For gaming yes this card is over priced, and you are better off getting a cheaper sli/crossfire configuration, but for some professionals that need fast render times and working on large data sets, this card will be a much cheaper and faster option than spending a lot more cash on Quadro/Tesla cards.

    I already ordered two Titans this morning. I will order another two when my other kidney sells on Ebay.



  • 1 Hide
    Stephen Bell , April 17, 2013 4:03 AM
    I think there is one significant point that is overlooked in the conclusion of this review:- 6 GB of RAM!

    You benchmarked iRay but didn't mention how with that 6Gb the Titan can do scenes that most other cards cant touch and would simply reject. Then there is the issue of viewing massive scenes. to show what I am talking about, go down the page a little and click on the landscape/ocean scene, which kills a normal video card but would probably fit inside a Titan.

    The Titan is faster than its pro origins, a fraction of the price and has enough RAM on it to do serious work. I think it is a "no-brainer" as a workstation card – just buy it.
  • 0 Hide
    Stephen Bell , April 17, 2013 4:06 AM
    Sorry, link in my comment above did not come out. The article and image is at: http://www.archiform3d.com/3d-workstation/video-card.php
  • 3 Hide
    ojas , April 17, 2013 4:17 AM
    Aha! Been waiting for tan article like this. Thanks!

    I must ask, though: Was the Titan using all the compute resources available to it? There was that setting in Nvidia's Control Panel, that let you switch between more double precision float performance or more gaming performance. Was it enabled?

    Quote:
    Tapping in to the full-speed FP64 CUDA cores requires opening the driver control panel, clicking the Manage 3D Settings link, scrolling down to the CUDA – Double precision line item, and selecting your GeForce GTX Titan card. This effectively disables GPU Boost, so you’d only want to toggle it on if you specifically needed to spin up the FP64 cores.


    http://www.tomshardware.com/reviews/geforce-gtx-titan-gk110-review,3438-3.html
  • 0 Hide
    rolli59 , April 17, 2013 4:29 AM
    Looking forward to the test against the workstation cards.
  • -1 Hide
    ubercake , April 17, 2013 4:43 AM
    k1114Why are there not workstation cards in the graphs?


    Thanks for the article. It is extremely thorough with the exception K1114 raises in the first comment.

    Why not include some Quadros and FirePros? This would give a real frame of reference with regard to workstation performance.
  • 0 Hide
    cravin , April 17, 2013 5:44 AM
    Yeah I don't see why people cant just mod the workstation (quadro ie) drivers to work with 680s and titans and stuff.
  • 0 Hide
    tsmithf , April 17, 2013 6:19 AM
    Were all the 3DS Max viewport benchmarks incapable of running with the Titan?

    Because the CUDA tests are rendering speed not viewport performance, and while it's interesting seeing the CUDA performance that really does need to be compared to pro cards for those interested in such things.

    However most 3DS Max studios doing architectural visualisation use Vray using the CPU on a renderfarm, they might use RT locally, but CUDA is still, distinctly secondary to using a graphics card for viewport acceleration.

    It would also be nice to indicate any issues gaming cards have in applications like 3DS Max, benchmarks are one thing, but trying to pick a non existant vertex because the drivers haven't rendered the scene correctly quickly makes you realise why the Quadro cards exist, in my experience the money saved isn't worth the frustration when you make your living from it.
Display more comments