Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

All We Know About Maximizing CPU Related Performance

Topper said:
I wonder how that thing compares to the 4GB GTX 680.
It's got more memory and more configuration cores. Can I assume it will get better benchmarks? Any guesses on the bandwidth?
Click to expand...

Regarding benchmarks, in the article referenced, the author - Hassan Mujtaba - states: "The only thing that would be changed [from the current Titan] is the GK110 with reduced core count around 2304-2496 and a 5 GB GDDR5 (384-bit) memory. Hopefully, we would be looking at a lower price range around $659-$759." [Emphasis and bracketed material added.] If that is true it would mean that the Titan LE would get much better benchmarks than the 4GB Galaxy GTX 680s that I own, even though mine are the highly overclocked version.

Regarding bandwidth, here's what Nvidia says, in relevant part, about the Titan: "NVIDIA® Kepler™ architecture gives you 2,688 NVIDIA CUDA® cores and 4.5 teraflops of gaming horsepower. Plus, GTX TITAN delivers a 384-bit memory interface running at a blazing fast 6 Gbps for an amazing 288 GB/s of memory bandwidth ... ." So as far as the memory bandwidth is concerned, I would't expect it to change so long as the memory clock (6.0 Gbps) remains the same. The memory interface (GDDR5) and interface width on both cards would be the same (384-bit GDDR5). Only the cores would be reduced by from 8% - 15%, depending on the final core count and the amount of memory would be reduced by 1 GB. That wouldn't be much of a hit to creative applications that use CUDA. But this would indicate that contrary to my earlier belief that OEMs, with Nvidia's blessing and more importantly Nvidia's price support, would soon drop the price of the Titan to the $700 - $800 range, that isn't going happen any time soon and, instead, a slightly weaker Titan (the Titan LE) will just be inserted to fill the gap. However, the more that I think about the prowess of the Titan, particularly when compared to the $4.5K+ Tesla K20X, I guess that I should not be at all surprised, given what you can do with a Titan to enhance its performance. The Titan has actually flipped positions with K20X for a whole lot less. So the $1K+ price tag isn't that bad in the end.
 
Tutor said:
The Titan has actually flipped positions with K20X for a whole lot less. So the $1K+ price tag isn't that bad in the end.
Click to expand...

That really is amazing. A thousand dollars is a bargain when you look at it in that way.
The Titan LE sure sounds good though.
The following image "MAY" be of a Titan LE. What do you think?
View attachment 408891
 
Don't think that I'm jaded. I'm just constantly adjusting to inflated prices.

Topper said:
That really is amazing. A thousand dollars is a bargain when you look at it in that way.
The Titan LE sure sounds good though.
The following image "MAY" be of a Titan LE. What do you think?
View attachment 408891
Click to expand...

That's what it may look like, but Photoshop could be at work.

I'm not so jaded that I view, in the absolute, the current Titan's prices as a bargain, but I do consider them, in the relative, a bargain. I'll put it in better perspective this way. My son who is beginning his work life while in school is paid the federal minimum wage for an on campus job. The federal government currently mandates a nationwide minimum wage level of $7.25 per hour. My son and I had a discussion wherein I revealed to him that long ago in the 1960's that I too had a minimum wage job. The federal government then mandated a nationwide minimum wage level of $1.81 per hour. My son asked how could I purchase anything making so little money per hour. Then we proceeded to have a discussion about economic relativity. I pointed out to him that while I was living in Southern California at that time that I used a motorcycle for transportation and that my motor cycle had two tanks: one gallon active and one gallon reserve, and that I could, at most, purchase 30 cents of gas at a time. Then I told him something that completely startled him, namely, that I was making more money then than he is making now. Then he looked at me in disbelief and asked, "how?" I next told him about the then existing prices and that because of the gas wars that I could, e.g., purchase a gallon of gas for 15 cents and that now the price of goods, in general, are so high that there isn't even a cent symbol clearly exposed on any keyboard (including, but not limited to, the Mac Pro, the iPhone or the iPad - I was just trying to show him that his old dad knows something about the iToys his generation uses to mark "coolness." Yes I know "coolness" is out). Then I showed him some math: $1.81 / .15 per gallon of gas = 12.07. Yes, I could buy 12 gallons of gas if there weren't then any deductions for things like taxes, etc. and that now this is the math that he sees: $7.25 / $3.60 per gallon of gas = 2.01. For his one hour of labor he now get paid enough, assuming that there were no deductions, 2 gallons of gas. When I was making less than a third of what he makes for an hour of labor, I could purchase 6x the amount of gas that he can now purchase. And those relative price differences apply not only to gas, but to everything else. Just to make his mouth water, I told him about the 25 cent full-size, fully-decked, all beef - back yard style hamburger, fries and coke combo.

Do the Teslas cost to much? DR*/
Do the Xeons cost to much? DR
Do the Titans cost to much? DR
Do the _____ cost to much? DR

*/ DR = Damn right. I'm just trying to conserve water by not overusing/abusing damns.
 
Last edited:
Tutor said:
I'm not so jaded that I view, in the absolute, the current Titan's prices as a bargain, but I do consider them, in the relative, a bargain.
Click to expand...

I am in agreement.
I made the following graph for myself.
I am interested in the Titan, Titan LE, and GTX 680 (4GB). Truth is I already have the 4GB 680.
The disclaimer on this chart is we have no idea if the specifications for the Titan LE are as I have presented.
The chart is for my benefit, but I decided to share it anyway. Did I make any typos?

View attachment 408929
 
Topper said:
I am in agreement.
I made the following graph for myself.
I am interested in the Titan, Titan LE, and GTX 680 (4GB). Truth is I already have the 4GB 680.
The disclaimer on this chart is we have no idea if the specifications for the Titan LE are as I have presented.
The chart is for my benefit, but I decided to share it anyway. Did I make any typos?

View attachment 408929
Click to expand...

Thanks for the chart. You guesses are as good as mine and Hassan's. Only time will tell. See post no. 9 here: https://forums.macrumors.com/threads/1572951/ . It'll give you some idea of the power of the Titan when tweaked.
 
Last edited:
Topper said:
How can your 3 Titans get beaten by two 670s or two 680s or one 580 or one 680 or one 670 or one 660 TI or one 460 or two 550s?
Even a guy with 4 Titans was low on the list.
I'm guessing the above were extremely overclocked with liquid cooling and maybe connected with SLI?
Or even more likely, I don't understand the test.
http://www.luxrender.net/luxmark/top/top20/LuxBall HDR/GPU
Click to expand...

I think I've found the answer on insanelymac:

mitch_de said:
Luxmark main dev , is non commercial, develops in an AMD system (Win). He started dev Luxmark (and Luxrender) with AMD 5870, now has AMD 7970. He will not & cant buy some expensive Gefore to optimize the OpenCL code also for Nvidia, as he already did (with OpenCL profiling) for AMD.
So this info shows, that beside OpenCL driver speeds & GPU hw speed also the software, the OpenCL source code does matter. At least up to 20% performance can be optimized by source code opti for some gpu series (AMD or Nvidia, which are in detail different in gpu computing usage / opti!)
Click to expand...
http://www.insanelymac.com/forum/to...ench-07-march-2013/page__st__300#entry1899238
 
Last edited:
Want To Give Your GPU A Personality MakerOver?

Have you every wished that you owned a Tesla or Qradro GPU or other very high end Nvidia card, but you're sure that you don't? Well, think again. If you own an Nvidia gaming card in the GTX/GT category, then you do most likely already own a Tesla, Grid or Qradro card on steroids. See, to save money Nvidia just disables the high end card's personality in the corresponding GTX/GT line by assigning the gaming card a different PCI device ID (imagine someone monkeying with birth certificates), using various resisters on you GPU's PCB. That way when you install Nvidia's drivers, the installer will know which personality to assign to your card - low end (but faster) GTX/GT or high end (but slower) Tesla, Quadro, or Grid. Gnif and his crew can help you give your GTX/GT a personality makeover (or if you don't like my analogy here, think of it as something like correcting identity theft). What kinds of makeovers are there? Take a look at, just some, of the the personality makeovers that Gnif and his crew are helping GTX/GT owners to perform:

A GTX Titan becoming a Tesla K20X;
A GTX 690 becoming a Tesla K10 or Grid K2 or Quadro K5000;
A GTX 680 becoming a Tesla K10 or Grid K2 or Quadro K5000;
A GTX 670 becoming a Tesla K10 or Grid K2 or Quadro K5000;
A GTX 650 becoming a NVIDIA GRID K1 or Quadro K600 or Quadro K2000;
A GT 640 becoming a NVIDIA GRID K1 or Quadro K600 or Quadro K2000;
A GTS 450 becoming a NVIDIA GRID K1 or Quadro K600 or Quadro K2000 (pre 2011 model); and
A GTX 580 becoming a Tesla M2090.

Will it change the speed of your GPU? No. It will run at its pre-makeover speed. Your Nvidia GTX/GT GPU, if anything, is already running faster than its high end counterpart - so it may even be fitting for you to try to underclock your GPU a little because what this makeover really does is to allow your GPU to express all of the locked in expressions that, pre-Gnif, were never driven by driver to the surface. So for you Nvidia GTX/GT owners, its time for you to study up on the other personalities your GPUs have been hiding, i.e., what was the high end Nvidia card of the day (and just as importantly what was its features) released at about the same time as was your GPU (or as a part of the same GPU chip family) because if you want to bring those high end family traits to the surface, all you have to do is:

(1) Arm yourself with knowledge (you can get that here from:
(a) [ http://www.nvidia.com/object/tesla-supercomputing-solutions.html ],
(b) [ http://www.nvidia.com/object/quadro.html ],
(c) [ http://www.nvidia.com/object/cloud-gaming.html ],
(d) [ http://www.nvidia.com/object/visual-computing-appliance.html ],
(e) [ http://www.nvidia.com/object/enterprise-virtualization.html ], and
(f) anywhere else your own resourcefulness takes you, like wikiing it for another example.);

(2) Get out your ohm meter, flux, solder and solder iron (or take a quick course or find the solder maven in your family or other circle of contacts) and

(3) Head on over to Gnif's Place: http://www.eevblog.com/forum/projects/hacking-nvidia-cards-into-their-professional-counterparts/.

If done correctly, you'll save yourself thousands of dollars and have a really fast version of the highest end, most comparable GPU card that Nvidia ever imagined that it was hiding from you.
 
Last edited:
Tutor said:
If done correctly, you'll save yourself thousands of dollars and have a really fast version of the highest end, most comparable GPU card that Nvidia ever imagined that it was hiding from you.
Click to expand...

Thank you!
That's the one thing I can do... solder.

p.s. That looks like one of our members.
 
How many Titan can I install in my MP and what are some alternatives to access more?

If you install an FSP booster or similar PSU in your Mac Pro, you'll be able to install, at most, two Titans inside that case (See posts #'s 1 - 106 here: [ https://forums.macrumors.com/threads/1565650/ ] ). If you want/need to access to more than 2 Titans, you need to:

(1) Purchase an external PCI-e chassis, like the NetStor NA250A that'll cost you about $2200 at the low end [ See posts #'s 106 to 122 here: https://forums.macrumors.com/threads/1565650/ ];

(2) Build a Windows system for about $3,000 that can hold up to four Titans [ mine is an EVGA SR-2 Westmere based system, but its 2 years old or an ever better value now is the SuperMicro GPU SuperWorkstation barebones system - see posts #s 116 - 123 here: https://forums.macrumors.com/threads/1565650/ and post #606 here, above: https://forums.macrumors.com/threads/1333421/ ]; or if you have two computer systems and have given your Titans the personality makeover describe in my immediately preceding post, then

(3) Use a combination of either 1 doubled (= 2 systems) or 2 doubled (= 2 systems) or 1 and 2 (= 2 systems) with the technique described at post # 123 here: [ https://forums.macrumors.com/threads/1565650/ ].

So to sum it up (kind of):
(1) two Titans can be put in a Mac Pro if you have adequate power from another PSU [= 2];
(2) one Mac Pro with two Titans, adequately powered, can be combined with a four or (up to eight if you've got the cash to spring for one of those 16 single, thus) eight slot dual width PCI-e chassis [=10];
(3) two Mac Pros each with two Titans with Tesla personality restored, adequately powered from another PSU + infiniband [=4];
(4) two Mac Pros each with one Titan with Tesla personality restored, adequately powered, + infiniband + 8 dual slot chassis on each [=18];
(5) two Mac Pros with one Mac Pro with two Titans with Tesla personality restored, adequately powered from another PSU, and the second Mac Pro with one Titan, adequately powered, + one 8 dual slot chassis connected to that Mac Pro, all max no. of Titans with Tesla personality restored, + infiniband [=11];
(6) one Mac Pro with two Titans with Tesla personality restored, adequately powered from another PSU + infiniband with one SuperMicro GPU SuperWorkstation barebones system completed, with four Titans with Tesla personality restored [=6];
(7) one Mac Pro with one Titan with Tesla personality restored, adequately powered, + infiniband with one SuperMicro GPU SuperWorkstation barebones system completed with 3 Titans, + 8 dual slot chassis on each full of Titans with Tesla personality restored [=20]; etc.
 
Information synthesis increases and quickens as we learn more.

Topper said:
Stupid of me to have mentioned it.
I just meant that I think I recognize "GNIF" as being a member here.
Click to expand...

Never is it stupid when your neurons fire, for stupid is just the opposite. Never stop arming yourself with knowledge. When and if an autopsy is ever done on your brain you want it to be the most wrinkled, pathway filled brain that the medical examiner has ever seen. Those depressions forming the wrinkles get filled with pathways from learning; more pathways provide the basis for future connections to be made. It's like the theory of folding space to decrease time travel.

I too recall prior reference to this site [ https://forums.macrumors.com/threads/1561013/ ], but at that time I don't recall any specific mention of why what Gnif's doing is so important, especially since Gnif openly states that it won't make your GTX/GT any faster {however, the GTX is already faster than it's Tesla counterpart to satisfy the gamers}. I think that the question of why he's doing it was raised, but it was not answered fully. Actually, Gnif says that he started doing it because Nvidia pissed him off: "No, just anger at NVidia for not clearly advertising that surround does not work under linux." But, be that as it may, why not take advantage of his serendipitous discovery. Christopher Columbus didn't set sail to "discover" what would much later become the home of WallStreet because he had any idea that it existed, but he is widely credited with "discovering" America [tho' my Choctaws et al. friends would differ]. My cerebral neurons related to the many practical applications of Gnif's work didn't fire until yesterday when I was researching an alternative method to the expensive, external PCI-e chassis to use multiple Titans on one project simultaneously without having to parcel or breakup the project to run them separately on different machines. The least expensive external chassis for adding > 2 Titans costs from $2200 [PCI-e 2.0] - $2448 [PCI-e 3.0] for four extra, external PCI-e slots. Thus, two of them would cost from $4400 - $4896. Being Mr. Cheapo as a few friends call me, but feeling the need for speed, I couldn't stomach paying $2448 for one headless chassis, not to mention paying pretty close to $5,000 for two of them. I'd rather have two new Sandy Bridge computers for about $6,000, with each having five PCI-e slots [with one slot per system being the connection point to the other system, leaving a net of eight slots for Titan cards]. From finding David Hunt's post where he details how he got two computers connected through an Infiniband Network that he set up at his home, I eventually found documents on the web site of the company that manufactured the cards that David used, that revealed that the company had a solution (costing about one-have the price of the headless chassis) for connecting two computers together as one for purposes of CUDA rendering, but that it works only if you're using Tesla cards. In Titans, certain Tesla card features have been disabled (i.e., the Titan drivers don't activate them). RDMA for GPU Direct is a Tesla feature that enables a direct path of communication between the Tesla GPU and another or peer device on the PCI-E buss of your computer, without CPU intervention. Device drivers can enable this functionality with a wide range of hardware devices. For instance, the Tesla card can be allowed to communicate directly with your Mercury Accelsior card or another Tesla card without getting your Xeon or i7 involved. Titans do not support the RDMA for GPU Direct feature. So that's when the neurons holding memories about Gnif's work began to fire => Why not give the Titans a personality makeover by changing their identities to that of Teslas? Then I could take advantage of the RDMA for GPU Direct features of the cards that I'd prefer to uses, namely the less expensive, faster ones called "Titans." New mental connection made <->; Pathway in wrinkle forms V̶; Circle completed O.
 
Last edited:
Gigabyte is making a huge entry into the server and workstation markets.

Not sure if this one [ http://b2b.gigabyte.com/products/product-page.aspx?pid=4498#sp ] has enough PCI-e power cables to power eight Titans, but if it does, I'd surely want a chassis like this one. It's not headless, i.e., you can load it with two Sandy Bridge CPUs, eight hot-swappable hard drives and 512 Gigs of ram, all in addition to the 8 GPUs. The GPU/coprocessor wars are really heating up and attracting lots of mercenaries. With a single Titan being able to muster over 2 Tflops of double precision floating point peak performance and over 4.5 Tflops of single precision floating point peak performance when tweaked, we'd be talking about a chassis, that from it's GPUs alone, could muster over 16 Tflops of double precision floating point peak performance and 36 Tflops of single precision floating point peak performance. So, of course, with only four Titans in one chassis, whether server or desk side, you'd only be able to muster a little over 8 Tflops of double precision floating point peak performance and 18 Tflops of single precision floating point peak performance.
 

Attachments

  • Gigabyte_GS-R22PHL.jpg
    Gigabyte_GS-R22PHL.jpg
    145.8 KB · Views: 141
Last edited:
Tutor said:
power eight Titans
Click to expand...

OMG! I couldn't even afford the electricity.
With one Titan, I'd done did thought I'd died and gone to heaven.
If I may ask, what would you do with 8 Titans besides proving M-theory?
Although I do have a friend who is doing DNA research and he could use something like that.
 
Tutor said:
Since the vast majority of my systems, including the Macs, PCs, Ataris, and Commodores, run faster than factory, I've been thinking of ways to convert the heat which is output from them to electricity to help power them. Any thoughts?
Click to expand...

Sure, spend some time on youtube. :) There's a bunch of home-made high efficiency thermoelectric generator designs using Seebeck's plates. I guess if you have a few air-cooled components running at around 60˚C or a little more then you could probably recover a few watts per machine.


I was of course thinking on a larger scale than using the computer as a differential source and there are probably other waste-heat generators I'm not thinking of - although running it off your hot girlfriend seems interesting. :D Keep in mind that a single solar panel the size of your front door and a few batteries is enough to run 3 or 4 fully loaded MacPro systems 6 to 8 hours a day - Or one or two stuffed MacPro systems 24/7.
 
Last edited:
Less can be more

Tesselator said:
Sure, spend some time on youtube. :) There's a bunch of home-made high efficiency thermoelectric generator designs using Seebeck's plates ... .
Click to expand...

Thanks a bunch - it gave me some ideas.

Here's something that I've been pondering. Having gotten my highest geekbench 2 score and my highest Cinebench 11.5 score on my Hackintosh using what I've dubbed as underclocking, I should have not have been at all surprised by the following: Most 8 GPU socketed servers advertised to run 8 Tesla K20 or K20X aren't validated to run 8 GTX Titans. Why? It has a lot to do with Topper's last comment,"OMG! I couldn't even afford the electricity." Since GTX Titans (and other GTX members) run much faster than their Tesla counterparts, the high density, high numbered PCI-e slotted servers "can't afford the electricity." That is, the Titans have a higher TDP than do the Telsas and the Titans can be overclocked to widen the gap even more. So I started, at Wikipedia, by comparing some of their respective specs and this is what I saw:

Memory Bandwidth
TeslaK20X / Titan
250 / 288
250 divided by 288 =~ 0.87

Single Precision (SP)
TOTAL (MUL=ADD+SF)
3950 / 4500
3950 divided by 4500 =~ 0.87

Memory Clock
5200 / 6008
5200 divided by 6008 = ~ 0.87


TDP Watts
235 / 250
235 divided by 250 = 0.94, but
250 * 0.87 = 217.5

Tesla K20X - 235 watts x 8 = 1880 watts, w/o counting the power needed by other system components.
GTX Titan (not overclocked) - 250 watts x 8 = 2000 watts, w/o counting the power needed by other system components.

2000 watts - 1880 watts = 120 watts

8 Titans x .87 = 6.96

So it seems that the Tesla K20X is running about 87% of the speed of a Titan and that if you you want to put as many Titans in an 8 GPU PCI-e socketed server built for Tesla K20Xs, then you may have to underclock those Titans by, at least, .87 or resort to other hacks. Or you may have to just drop your expected no. of Titan-filled PCI-e slots by a Titan or two (or three if you over clock them). At such high densities and given the current PSUs available and safe to use in a server environment, every extra watt counts.
 
Last edited:
Yeah, as you load up a system the power requirements can become an issue. And at 2,000W there's a lot of secondary systems that will be affected. Like air conditioning and cash low. :)
 
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back