GTX 780 TI

[kracker]

To the best of my knowledge, the 780 and the 780 Ti is not the same die.
Also, sure you could chip a 6950 or a 690 to a higher/business card, but I'll bet you that the success "stable" rate is not 100%.

The professional cards are more expensive sure, one part is from higher bin, but the majority is... because it is a professional card. I mean think of it this way:

Gamer: Oh noez, a $3000 card? I can't afford that...
Business: Oh noez, it's $3000? Oh well, I'll have to deal with that, I don't have much choice.

You see nVidia/AMD's grand scheme of things?

[Manpowre]

The actual GPU might be the same production, it is when testing the GPU it comes out to be a lower clocked or parts of the GPU is not as good.

As I see it, the PCB board might be the same in some cases, like titan and K20, but the difference are many:
* K20 has registered memory chips
* Titan and 780ti has higher clock speed, but lower DP than K20
* K20 is clocked lower due to registered memory, and secondly to 100% assure that the "pixels" or compute nodes inside GPU are correct.
* K5000/K6000 also have registered memory to reduce memory errors.
* Who cares if a pixel here and there is wrong in battlefield or any other game, blinking red for 2-3 frames ?
* The guys in TV business, which use the K5000/K6000 DO care if a red pixel is shown for 2-3 frames. it is a big deal (I work in this business, so I know).
* When you clock the GPU higher, parts of the GPU does not work as intended. It could mean that DP part of the GPU doesnt work as intended with full 1/3 enabled. Instead, Nvidia cripples it to ensure it becomes a stable GTX board competing in 3D rendering for games.

Nvidia had to produce 19000 K20's for the TITAN cluster setup first. out of that, they probably got enough to produce the Titan, thats probably WHY the titan was born so early. Made it GTX and put on unregistered memory. Remember that titan came in February 2013. 780TI comes in november 2013. 10 months later.

Nvidia with TMSC perfected the chip production to enable all cuda cores for 780ti. Also there is no new TITAN cluster machine to take the top production. And it is competition they have to deal with as the 290x is out.

As I see it, the professional cards are there to ensure the stability, same GPU, better production, but that red dot for 2-3 frames isnt there for K5000/K6000. For Tesla cards, they are clocked slower to ensure compute performance are as promised with registered memory. Less risk of FFT errors for instance.

I really love the titan cards, they did something right there.. atleast for me who loves computing.

[LaurV]

Originally Posted by kracker

To the best of my knowledge, the 780 and the 780 Ti is not the same die.

This I can confirm, from a very trustful source.
OTOH, Titan and K20(x) is the same chip.

Talking about DP performance, 1/24 is quite ridiculous. I can't understand where did they come with the number, in fact. A SP float has 23 bits of mantissa, and a DP float has 53 bits. But the most of the applications which can't live with SP floats, will feel convenient with 46 bits of mantissa too. Including most of our LL tests. The point is that when one uses SP floats to do FFT multiplications, there is not enough space in SP floats to do the carry propagation, etc. One may use two SP floats, having together 46 bits of mantissa, and 16 bits of exponent, to keep the polynomial coefficients of Karatsuba multiplication, to get 1/3 DP* performance, in software (please remark the star, this is still 128 times "less accurate" than "true" DP, as there are 7 bits less in the mantisa, but still enough for the most of the needs, and you get it 8 times faster!). Also, one can use 3 SP floats, for a mantissa of 69 bits and 24 bits of exponent, and use software Toom-Cook-like multiplication, to get 1/5 (with some overhead for addition and constant multiplication) or 1/7 (with almost no overhead) DP performance, or just do 1/9 with blind "school grade" multiplication, there is more precision than a "real" DP, and you get it 3-5 times faster (compared to 1/24). Considering that the cards could do all the Karatsuba/Toom-Cook multiplications and overhead in the same time, in parallel, and they even possess 24-bit hardware multipliers, well... It seems a bit stupid for me... Why should they release cards with 1/24 DP performance?

If one has a card which can multiply SP floats very fast, and multiply 9 of them in the same time, then he can use 3 of them to fully cover a DP float, and implement 1/9 DP performance in software. Or I am dreaming?

But this is another discussion...

[xilman]

Originally Posted by TheMawn

Neither of you understood my meaning, by the looks of things.

Perhaps so.

Originally Posted by TheMawn

I'm giving them crap for this. And this. $3500 vs $730.

It's a free market. You will pay what you think it is worth. Some people think it's worth paying $3500 whereas you do not. So be it.

[LaurV]

The difference between the two is the reliability, i.e. the ecc memory, which does not lose bits on the way, and few thousands (no joke) hours of intensive testing and few hundreds climate chamber cycles (for each chip/card) to ensure that the cards perform according with the specs. Imagine Retina is his evil lair doing ballistic rockets with k20. Do you think he will use 780's instead, because they are cheap, risking to miss his over-the-ocean targets because a pixel fail?

Someone needs to pay for the current consumption, menpower, etc, during those tests. Good comfort is expensive, anywhere.

[retina]

Originally Posted by LaurV

Imagine Retina is his evil lair doing ballistic rockets with k20. Do you think he will use 780's instead, because they are cheap, risking to miss his over-the-ocean targets because a pixel fail?

Absolutely. ECC is the only way to go if you need reliability and accuracy. But down clocking (or at least not overclocking) helps a lot towards achieving that on non-ECCed cheaper systems. But you'll never solve the cosmic ray, or PSU glitch, problem just by underclocking; those will still trip you up.

As for the ICBMs: Have you been spying on me? I thought only the NSA/GCHQ was doing that sort of thing. Anyhow, it is not true and I deny it emphatically. Not doing that. Nope. Not me. Must be my next-door-neighbour in the lair next to mine. Or Mini-me. It's probably him. Yeah, that's it, the little bugger.

[Manpowre]

Yepp, so it is not just about rebranding existing cards. The 780TI is simply made for consumer market, non-registered memory, higher clock speeds which cripples reliability but perfected chip reliability.

Though, Im surprised Nvidia didnt drop the pricetag on the Titan. I cant even imagine they sell any of these anymore. 780ti is better for gaming. only a few of us buing titans for cuda and computing.

For video editing, 290x is even better. Folding and mersenne is mabye the only usage left for Titan cards.

[TheMawn]

Alright. I stand corrected.

I still fail to see the justification in two thousand dollars more for a GPU but if someone else thinks it's worth it, so be it. To be fair, I didn't know about the ECC memory on the professional cards.

I still won't be surprised when someone finds out the 780 Ti can be successfully hacked into a Quadro in 95% of cases. I won't be surprised, nor will I get bit.

[Manpowre]

Originally Posted by TheMawn

Alright. I stand corrected.

I still fail to see the justification in two thousand dollars more for a GPU but if someone else thinks it's worth it, so be it. To be fair, I didn't know about the ECC memory on the professional cards.

I still won't be surprised when someone finds out the 780 Ti can be successfully hacked into a Quadro in 95% of cases. I won't be surprised, nor will I get bit.

Remember that the old hack, was that someone found a way of using the quadro drivers which also have a API that allows grabbing frames and send to video boards. This is what the company I work for is doing.

The quadro cards are clocked lower due to slower ECC memory, and error correction through the GPU. also, 80% of the cost of the board is not the board itself, it is the quadro drivers which is way different than the GTX drivers. the quadro drivers include a full board API aswell.

The hack didnt turn a gtx card into quadro, it made the gtx card report that it is a quadro with similar HW which allows quadro drivers to be used instead.

[LaurV]

Originally Posted by Manpowre

The hack didnt turn a gtx card into quadro, it made the gtx card report that it is a quadro with similar HW which allows quadro drivers to be used instead.

Wanna say that. You were faster. The "hack" was in fact cheating the drivers into believing a quadro is inserted where a gforce was. It was not a "performance change" under no circumstances (in fact, the performance of quadro is totally lousy, they sacrifice a lot of performance for stability) but some people feel better to lose a bit of stability to get more performance, and this was what the "cheat" did.

[TheMawn]

Originally Posted by Manpowre

The hack didnt turn a gtx card into quadro, it made the gtx card report that it is a quadro with similar HW which allows quadro drivers to be used instead.

My understanding was the drivers for the GTX did not allow for full use of the DP hardware whereas the drivers for the Quadro did.