Titan: Excellent Performance per Dollar

[BallaTheFeared]

Geez, AMD is doing great here. I wonder why they failed so hard at penetrating this market with GCN. I wonder what the people actually buying these cards know, that forum trolls don't?

 

[Arkadrel]

I wonder what the people actually buying these cards know, that forum trolls don't?

They probably know how to code in CUDA, because its been out alot longer, and rather use CUDA than learn a new program languge.

So dispite not being "excellent performance per dollar" it still sells.

 

[Rakehellion]

For the PhD minions out there, its time to convince your advisor to buy Titan. We're experiencing a x2 speedup over GTX680 without any code modifications!!! Time to tell your advisor time is money, double the research throughput w/ Titan!!

Titan cost more than twice as much as the GTX 680, so how is that good performance per dollar?

 

[Lonyo]

Geez, AMD is doing great here. I wonder why they failed so hard at penetrating this market with GCN. I wonder what the people actually buying these cards know, that forum trolls don't?

I think it was somewhat to point out that it very much depends on the task which card is the best per dollar.
The OP didn't specify any particular tasks, which is why the advantage over the 580 vs 680 was brought up in the first place.

Some people have also mentioned the HD7970 because in some tasks that too is better in performance per $ than Titan.

Basically Titan CAN be good if you are doing the right task compared to a GTX680, or 580, but sometimes the 580 can be a good deal where Titan still struggles, or the HD7970 can be best per dollar if it's an area where that is strong.

The OP was rather lacking in his breakdown of where exactly his gains were coming from though, and it's probably that while his stuff has found significant gains, those won't be the same for everyone, as we already know, due to different relative strengths of GTX580 vs 680 vs HD7900

 

[bystander36]

Another big part of performance per dollar in this scenario is the cost of labor. The faster it works, the less labor they pay.

 

[BallaTheFeared]

I think it was somewhat to point out that it very much depends on the task which card is the best per dollar.
The OP didn't specify any particular tasks, which is why the advantage over the 580 vs 680 was brought up in the first place.

Some people have also mentioned the HD7970 because in some tasks that too is better in performance per $ than Titan.

Basically Titan CAN be good if you are doing the right task compared to a GTX680, or 580, but sometimes the 580 can be a good deal where Titan still struggles, or the HD7970 can be best per dollar if it's an area where that is strong.

The OP was rather lacking in his breakdown of where exactly his gains were coming from though, and it's probably that while his stuff has found significant gains, those won't be the same for everyone, as we already know, due to different relative strengths of GTX580 vs 680 vs HD7900

Where does Titan struggle against the 580?

7970 is a non factor, it can't do CUDA which is the major language in this field.


Titan is much better than the 580, 580 has crippled DP, and operates at the same power envelope for considerably less performance. Start looking at cost of operation, 3 times the duration is not only a time thing, but it's 300% more power consumption.

But then again you have people who have no idea what they're talking about using percentage of peak performance achieved charts as performance charts, I mean seriously...

 

[lambchops511]

Where does Titan struggle against the 580?

7970 is a non factor, it can't do CUDA which is the major language in this field.


Titan is much better than the 580, 580 has crippled DP, and operates at the same power envelope for considerably less performance. Start looking at cost of operation, 3 times the duration is not only a time thing, but it's 300% more power consumption.

But then again you have people who have no idea what they're talking about using percentage of peak performance achieved charts as performance charts, I mean seriously...

7970 is a non factor,, unless we get like literally 100 cards for free we will probably never switch to AMD. Our codebase is CUDA and it not economical to switch to OpenCL. We also heavily invested into NVIDIA's libraries (thrust, cublas, etc..) which OpenCL may or may not have similar alternatives.

DP performance actually means nothing in our line of work, we stick to SP.

Actually, we found the performance has high variance, really depends on what you are doing. If your model allows you to fit most of your number crunching in the (smaller) Kepler cache, we notice a significant delta in performance over cache spilling... the delta can be ~30% although we haven't measured too hard on this. This is of course highly dependent on your nature of work.

 

[Cloudfire777]

This forum frightens me. Its so full of Average Joe`s that is strongly pulled toward AMD that have no idea how things really work in the real world.

 

[Arachnotronic]

This forum frightens me. Its so full of Average Joe`s that is strongly pulled toward AMD that have no idea how things really work in the real world.

People just hate companies that make money. AMD offers "value" and it's an "underdog", so everyone roots for 'em. Badly run company that had a brief moment in the sun during 2003 - 2006 before "Conroe" left them bleeding in an alley. Graphics division is still pretty good, but it lacks the finesse and brand recognition that Nvidia has in key markets (Asia). That being said, I used to adore their "Athlon II" parts before Intel came in and gobbled up all of the value price points there, too.

AMD has bunch of loyalists still waiting for the next "Hammer".

 

[badb0y]

Yea screw AMD!

 

[BallaTheFeared]

CPUs fell off, GPUs are starting to fall behind due to lack of funds.

RIP.

Too soon? ()

 

[tincart]

I'm actually working on a paper on ethics systems in computer games. I'll see if I can get a grant to buy a few Titans. My paper will clearly be better if I can game with 8X MSAA.

 

[wand3r3r]

This forum frightens me. Its so full of Average Joe`s that is strongly pulled toward AMD that have no idea how things really work in the real world.

Thankfully you're here to educate. How do things work in the "real" world?

 

[gamervivek]

Thankfully you're here to educate. How do things work in the "real" world?

http://hgpu.org/?page_id=3529

 

[Cloudfire777]

Thankfully you're here to educate. How do things work in the "real" world?

You can start by looking at the people saying the 7970 is a better options since it offer better performance/price than Titan and then look at OPs last reply.
And read the link above from gamervivek.

I`m not counting that you are able to do a simple 2+2, but I`m not gonna waste my time explaining it to you

 

[96Firebird]

Another big part of performance per dollar in this scenario is the cost of labor. The faster it works, the less labor they pay.

Yes, people need to understand this. But most of the people here only play games with their video cards, so when the discussion moves towards the professional side of things, they are stuck in their ways.

 

[Rvenger]

People just hate companies that make money. AMD offers "value" and it's an "underdog", so everyone roots for 'em. Badly run company that had a brief moment in the sun during 2003 - 2006 before "Conroe" left them bleeding in an alley. Graphics division is still pretty good, but it lacks the finesse and brand recognition that Nvidia has in key markets (Asia). That being said, I used to adore their "Athlon II" parts before Intel came in and gobbled up all of the value price points there, too.

AMD has bunch of loyalists still waiting for the next "Hammer".

K6-2, Thunderbird, Athlon XP (2001) were all great CPUs as well. So AMD did have competitive CPUs for much longer than you think.

 

[Arkadrel]

I was wrong nvm.

 

[NTMBK]

Geez, AMD is doing great here. I wonder why they failed so hard at penetrating this market with GCN. I wonder what the people actually buying these cards know, that forum trolls don't?

-Existing codebase- CUDA has been pushed hard in these markets for longer, making getting into it an uphill struggle for AMD. Porting code from CUDA to OpenCL takes time and effort, which costs money and wastes precious time which could be used to do research.

-OpenCL is implemented subtly differently by every vendor, making portability a nightmare.

-Ease of working with the code- OpenCL is frankly horrible, compared to CUDA. The amount of boilerplate garbage you have to deal with just to set up a context. The horrendous C-style API.* Academics want something that they can just pick up and run with, with the minimum of hassle, and CUDA just flat out wins there. There has been some progress with the release of things like the CLU, and new language extensions like C++AMP, but there is still a loooooong way to go.

I love OpenCL as a concept, and wish CUDA would go away. Being tied to a single hardware vendor is just... blech. I want a portable GPGPU language which will run anywhere, and be nice to work with. But OpenCL just isn't there yet.

*As an example of how horrid it is, let's take a very simple example. You want to get a list of all the OpenCL cards available in the API. So what do you do? First, you create an integer variable. Then you pass in a pointer to this variable (cast to a void*, because who needs type safety?) to a function, along with an enum telling the function what you want it to do, and certain other parameters left set to NULL as you're not using them. The function then sets this integer to the size of the list of cards which it is about to give you. Then you need to allocate an array that is this length. This array is then passed into the same function again, but as one of the other parameters (which was NULL the first time you called it), and finally it gives you your list of OpenCL devices. Utterly, utterly bloody horrible. (I admit that I probably got one or two minor details wrong here, because it's such a complex pain in the arse that I've struggled to memorise how to work with it... and have no desire to go trawling through those docs again.)

 

[wand3r3r]

You can start by looking at the people saying the 7970 is a better options since it offer better performance/price than Titan and then look at OPs last reply.
And read the link above from gamervivek.

I`m not counting that you are able to do a simple 2+2, but I`m not gonna waste my time explaining it to you

Average joes aren't doing CUDA/GPU programming, again, how does the real world affect the average joe such that titan is good performance for the $ or whatever you are trying to promote?

The OP finds some benefit, sure, but you're claiming the average joe is ignorant and uneducated, please enlighten us.

 

[Anarchist420]

Compute performance vs gaming performance is a false dichotomy. AMD's forward+ runs much better with compute than without.

 

[lambchops511]

Average joes aren't doing CUDA/GPU programming, again, how does the real world affect the average joe such that titan is good performance for the $ or whatever you are trying to promote?

The OP finds some benefit, sure, but you're claiming the average joe is ignorant and uneducated, please enlighten us.

If you read my OP

For the PhD minions out there

This isn't for the "average joe", unless you are suggesting every "average joe" here is a PhD that requires intensive computing for their research.

Titan is a high end product (and priced such), and definitely not for the "average joe". However, if your work / research requires high performance computing such as GPUs, Titan is a very very very excellent product and has an extremely high return on investment.

Consider the option between $0 capital cost for not purchasing any hardware, versus $5 to purchase a simple calculator. If your work permits it, even though the $5 calculator costs infinitely more than not purchasing a calculator, it still has an excellent return on investment since it saves you time.

Similar analysis can be done for Titan, sure you are spending more dollars for initial upfront capital costs, but you should do your cost analysis over a period of time. How much is your time worth? Hence, if your time is worth squat, it is very very very advantageous to purchase Titans. Our approved supply channels are actually all out of Titans and generally have a 2 week wait period to get them, they are selling very very well as far as I know (K20s are available, but they do cost quite a bit more, esp when our group does not need the ECC or enterprise features).


Some benchmarks here as well:
http://forums.anandtech.com/showthread.php?t=2312284

 

[Cloudfire777]

May I ask what research you are doing OP?

 

[notty22]

Here is a lot of young people doing gpu computer research and receiving grants.
NVIDIA Awards $275,000 to 11 GPU Computing Geniuses

 

[thilanliyan]

Here is a lot of young people doing gpu computer research and receiving grants.
NVIDIA Awards $275,000 to 11 GPU Computing Geniuses

That's a front for shilling I tell ya!!
Quick...trace their IPs...they may be on these forums ALREADY!!! We all have to be vigilant! Anyone recommending Titan is suspect right now! :biggrin: