Titanic questions

[LaurV]

Some member of this forum wrote me a PM with the title above, and the content below. I won't say the person, maybe he has his reasons to write privately, but I consider this an excellent topic for public discussion, so I say sorry to him for bringing it public.

I see the plain Titan GTX (not x, or z, or whatever) occasionally as a refurbished card on Newegg for $499 and I have an old computer it can go in. Titan is also second on the cudalucas performance page for LL, which is my interest. My concern is that this is a 2013 card which is pretty old. From what I see from Nvidia, NVlink is going to be the next big step. But that will take more time, and it will presumably cost a lot. So I'm wondering what's going to be the next big DP FP performing card, and can I afford it. The teslas are crazy expensive, and even $1000 for a card is pushing it for me. But are they going to come out with something maxwell-like that's more efficient than GK110 and that can actually do reasonable DP FP, or is it just going to be NVlink at wildly high prices. If so, I'll buy the Titan. Thoughts?

My "thoughts" are that NV did a couple of two big "mistakes" in the past. Those mistakes are called "gtx 580" and "Titan". They were not exactly mistakes, that is why the quotes. Say "planned mistakes", hehe. It was partially market survey, or strategy, partially it helped them to work the internal mysteries of the complicated DPFP architectures, to be able to come out with stronger "scientific" cards (like Tesla). What were sold as "late" titans (i.e. Black) were "failed-production-test-number-x" Teslas, in fact. They had the same chip. The best Titan (for DP) is still the original, old, plain, Titan.

First, NV will not repeat these mistakes in the future. Not in the near future, anyhow. "Normal" people are not interested in DPFP precision. I am not talking about gamers (not only!). The most of the applications I know (including all sorts of CAD programs, for mechanical-, electrical-, electronic-, furniture- (interior-) design, constructions, chemistry, biology, etc.) are happy enough and faster with SPFP. The more "scientific" DP is - you guessed - for scientists. Or for silly people like us. None of these two categories have the money for Teslas, and there is a reason why a Tesla costs $5k (we posted in the past about this). Therefore the market wants super-fast single precision cards, which can handle those 5 billions frames per second of I don't know which Crisis or Assassins, and that is what the market will get.

Second, the "consumer" cards will ALWAYS stay in the current range of prices. At least in the foreseeable future. As new roosters appear, they will get their top of the list ($1000 or so, or a bit more) pushing down the prices for older roosters. (Do you know the old rooster joke? Don't play with the old rooster! ) This is what the people in the "normal" category defined above (those with the SP interest) are willing to pay. The prices are not defined by some "production related" issues, believe me, I was working in electronics manufacturing companies for 30 years. What is on a Titan costs below $100. Including the big mountain of silicon in the middle. The price is given by what the market wants to pay.

Now, if someone will offer me new Titans for around $500, or hydro-coppers for around $700 (all in current US dollars), then I will immediately buy few. For 4 (four) of them I have already the free slots, preferably water cooled, not necessarily hydro-copper, but I like HC because they are very thin.

However, you should not do what the preacher does, without thinking first. My grandpa always said "do what the priest says, not what the priest does", but that is a different story.

There are at least 2 or 3 problems with the Titans.

First, they are old. As my OP said, this is a 2-3 year old card, which means a lot in this market, its chip is "end of life" already, and they are (almost) all second-hand now, because there were only few produced and all sold like crazy. So, there are extremely low chance that you can buy a "new" one, which was never used. And buying second hand stuff is risky, there might be a "secret" reason why the seller is selling the card, and even if there is no "secret" reason, you will still worry if there is one. For example, I was lucky with my Titan shopping but I know people who were not. You have to define your own "risk affinity", like in Forex trading.

Second, they consume a lot. This partially because of the first reason - they are old, so they are old-technology, and partially to the fact that a lot of computing power is concentrated on a very small surface, so they are not extremely power-efficient. If you compare the work they do, with the power they take, and if you have the free slots, then you may be better buying 2 or 3 "newer" cards which will cost more, and occupy more slots, but will do the same amount of work for a fraction of electricity, and if you plan to use them for few years, then you may be better buying new cards. Especially if the electricity in your area is not cheap (or it does not come for free, like in an office, etc). Because they consume a lot, they need good cooling, which adds to the price - if is air, it will add to the consumption for fans, if is water, then think about pumps, etc. However, they are good room heaters, so if you live in a cold area and use electricity to heat your house (I envy you, son of a bleach!) then you may seriously consider buying lots of Titans!

Third, they are big, and - if not water cooled - they are noisy. Related to the first and second reason, too. But I mentioned it because it may help with the calculus of the slots, what I said above related to the fact that more new cards will occupy more slots may be false, because the new cards are thinner. Titans are double slots mostly, and few are triple slots. Most of good 580s are triple slots too, i.e. you may find 3 new cards to put in 3 single-wide slots, but you will succeed at most to put 1 (one) Asus 580 DC2 card in those 3 slots, because the card is 2.8 slots wide. Of course, you can use cables, which add to the costs too (and they are not cheap!).

So everyone has to do the homework by him/herself. One advice however, I can give you, and I would insist you follow it: don't buy Teslas. They are too expensive for what they can do, and you may have the surprise they won't work in your rig because most of them are designed for servers, have special protocols and are passive cooled (i.e. will need external fans to blow air onto them). If you have so much money to buy few Teslas and want to have a Tesla-branded thing around your house, then better buy a roadster.

I can't say much about the NV link, I am the guy who need to see first, then talk about. But I pray to all my gods the same way I was praying when keplers and maxwels came, let them come and be faster and have a lower price so anybody afford to buy them, so the prices of Titan go down so I can buy more Titans...

[Mark Rose]

People do sell them because them for efficiency reasons. But it turns out efficiency isn't everything. I can buy 3 GTX 580 for less than $250, versus 2 GTX 980 Ti for $1600. Sure, I'd spend about an extra $30 a month in electricity, but it would take 4 years to come out ahead with the 980 Ti's. In four years, the 980 Ti's will be as slow as 580's today and not worth much.

The original Titan is still probably the best deal for LL. I would certainly consider buying used for $500, if LL were my task. Ebay has one for $297.

The only thing I could see changing the equation soon is if Pascal has better FP64. Apparently it will have the ability to split the FP32 units into 2x FP16, so maybe it will work the other way, too. Also, Pascal is getting a massive process shrink, so there should be space to add more FP64 back in. Pascal is still a year out though.

[LaurV]

That is an auction, and there are ~5 days left. You won't get that for less than $450-500, to which I (from Thailand) have to add ~$150 shipping and 10% (of the declared value of the parcel) custom (import) taxes.

(this is to justify why I say 150 for shipping! which of course, is robbery!)

As a side talking, I had a discussion with Mike (Xyzzy) in the past, where I was asking him to help me buy some hardware, due to the fact that he is there, in the US. He agreed, but said he can't understand how I can get a better deal doing such roundabouts. The current offer is one more example of a "bad" deal I could make, if I try to buy it directly, at a betting price close to 500-600, it will cost me close to 800-900 to get it on my hands, and this is an air-cooled. Not counting the risk that, after paying, the seller may say that he, in fact, can not ship to Thailand (for whatever reason, it happened to me in the past - grr it took me a while to find the link, someone renamed the thread ). He will return me the money promptly, so I can't accuse him of foul play, but I will lose a lot due to exchange rates (to US$ and back) and paypal taxes - of course, my loses are profit (free money) for paypal. I know for sure that they employ people ("sellers") to do this trick (no, it is not conspiracy theory!), from which the "sellers" get a commission. And ebay/paypal may say that they are now separated, i.e. they are "not the same company anymore, but two different companies", but they still have the same people, same owners, doing same business.

Edit: Disclaimer: to be clear, I don't try to buy this card, I only used it as an example, because it was linked by Mark

[Mark Rose]

Yeah, I have the same issue with the outrageous shipping costs. To Canada it's $100. So I basically stopped buying things on Ebay US.

[xilman]

Originally Posted by Mark Rose

Yeah, I have the same issue with the outrageous shipping costs. To Canada it's $100. So I basically stopped buying things on Ebay US.

It works the other way too. I bought a 1870-ish family bible on the off chance a friend in Florida would like it. Cost me £105. The cheapest shipping I could find was over £80. In the end I rode 240 miles up to Yorkshire where she was staying on vacation ~10 months after the initial purchase and handed it over in person. Good excuse to go for a nice long bike ride, meet up with old friends, visit a very pretty part of the country and drop in on my folks roughly half-way back home.

[Mark Rose]

Sounds like a very nice ride! How many days did it take?

[xilman]

Originally Posted by Mark Rose

Sounds like a very nice ride! How many days did it take?

Nine hours spread over two days.

Having an 1100cc 90HP engine makes it really rather straightforward.

[Mark Rose]

Originally Posted by xilman

Nine hours spread over two days.

Having an 1100cc 90HP engine makes it really rather straightforward.

Ahh, not one human power ;)

[ATH]

@LaurV You wrote in another thread that you can "unlock" the original Titan. How much extra DP/SP does that give? Is it possible with all brands of the original Titans?

[LaurV]

I saw this post, but I didn't have time to "prepare" a detailed reply, or tutorial. I replied to your PM today, during my lunch break, with mainly the same text like here, and later decided to put it on public too, someone may benefit from it. There are few corrections and additions compared to the reply to your PM.

Generally, there are two types of "locking" that Nvidia applies for the cards, one is power lock, and one is voltage lock. To avoid a multitude of RMAs from crazy overclockers (they advertised the card as "overclockers' dream" remember?) they won't allow you to raise the voltages artificially, and to increase the power that the card can digest.

You can generally overclock the card with different applications, like EVGA's Precision X, or MSI's Afterburner, or else (some discussed on the forum already), but when you overclock the card "too much", the card becomes unstable if you don't also increase the voltages. But the voltages are "locked" in the sense that you can not operate them "by hand", they are automatically computed from the power requirement, so actually what really happens (and you can easily see this in GPU-Z, for example) is that the voltages are reduced if you increase the clocks (to keep the power "in check").

This makes the card run unstable at higher rates of overclock. One way to "unlock" the card is to overwrite its bios with some other bios which allow the voltages to be set to higher values. Obviously, unless you want to dis-assembly by yourself, this method can not be applied to all cards (brands), you must _have_ a bios file, downloaded from somewhere on the net, from a source you trust (there are many scrap files there, which may kill your card forever - rumor has it that some of them are "put on the market" by the card manufacturers themselves, you know, this is conspiracy theory, but would you trust "upgrading" your bios after you paid a lot of money for the card?).

You can read the bios you have in a file, and compare with the new bios, and see where the differences are, they should be not many, only few bytes change, assuming a voltage table is overwritten, etc. There are always risks.

Either you do this "unlock" or not, there is another step you can do (and you must do, in case you update your bios, because higher clocks and higher voltages will require higher power), this is changing the power resistors. This can be applied to all the cards, and it will "cheat" the card about how much power consumes itself. All cards have shunt resistors of 5 milli-ohm at the 12V inputs, and they measure the power by measuring the voltage drop over these resistors. You can see them, they are the BIGGEST resistors on the board, close to the 12V power supply. If you apply a solder bridge (tin) over the resistor, you will decrease the resistance to about 3 milliOhm, therefore almost doubling the maximum power limit that the card can take in. Sometime, doing this is enough, depending on the brand, because when you overclock the card, it will measure the power consumption and establish the voltage function of this (i.e. a higher value than without "bridge").

If you do this, you have to go higher and higher with the clock, slowly, step by step, and watch the power the card takes, and you have to be able to cool the card. I mean COOL. Most of the time, air cooling is not enough, you need liquid (water) cooling. Watch the temperatures, increase the clock, watch the temperatures, increase the clock, watch the temperatures, increase the clock, watch the temperatures, increase the clock, watch the temperatures... You got the idea. Otherwise you can easy make bar-b-q of your memories or the Titan chip.

Note that Titan is really a power hungry card, if you can efficiently cool it down, you may get to really high clocks, and get about 50% or more output, for about double energy consumption (and heat produced). Which is less efficient, but well, if you want more output...

I could get about 20% more output from the old card from Xyzzy, and that was air cooled. Well, that is fried chicken now, as you know, and that may be a reason for everything

I didn't start experimenting with the new card yet. But I will, after it finishes Manfred's DC (which BTW, it passed 210M iterations with success, ETA about 20-25 days). I want to "pay my debts" first, before starting to cook fried chickens on it. Otherwise it will be a pity to have to cede the ~3000GHzDays of credit to someone else to finish the test

You can get a lot more info if you google for "unlocking gtx titan", for example here they get about 14% more, from simple overclocking and overvoltaging (a normal Titan gets about X5000 mark). Also, here they talk about power resistors. These are just first two links google provides. The guy in the second article even uses a "reversible" method to increase the power, to which I will vote against, unless you really take precautions! What can happen is that the patch of dried nail paint liquid metal gets off, due to vibrations, and makes a short inside of the board (which, at testing time, is inside of the box and under the cooler, not easy accessible). But well, it is a method worth to try, assuming you are careful.

There is a third kind of "unlocking" to which I was referring in the past on this forum, that is "unlocking a 780 to behave like a Titan". That is for 780 where you can not go into Nvidia Ctrl Panel and set it to "DP". They sell the 780s like game card, most probably after failing some DP production tests, and for that cards you can not turn DP on. Some guys found a way to do it, by moving some jumper resistors from one position to the other on the PCB. Of course that is on your own risk, you lose the warranty and the card may produce strange results when used for DP. But that is another story (which you also can check on the web, they even make 780 to believe about itself it is a Tesla, hehe, here is a starting point, but that will for sure return crap results, and there is no ECC memory, and no wide mem interface, etc, no idea how that card will behave after moding it). Doing the mod will almost double your 780's LL performance, but will almost halve its TF performance (same as when you switch a Titan from SP to DP mode in Nvidia CP).

[Madpoo]

Originally Posted by LaurV

... All cards have shunt resistors of 5 milli-ohm at the 12V inputs, and they measure the power by measuring the voltage drop over these resistors. You can see them, they are the BIGGEST resistors on the board, close to the 12V power supply. If you apply a solder bridge (tin) over the resistor, you will decrease the resistance to about 3 milliOhm, therefore almost doubling the maximum power limit that the card can take in....

Interesting. I don't have a GPU but I read this anyway...goes back to my formal education in electronics.

I wonder if you couldn't just wire in a POT in parallel, have a nifty little knob to adjust the resistance across the shunt instead of shorting it entirely? Give you a little more fine (albeit manual) control over things.

Anyway... who knew you could do so much GPU hacking? Maybe one of these days I'll find a decent used one on ebay or something and give it a whirl in my home machine.