$3000 GeForce GTX TITAN Z Tested, Less Performance Than $1500 R9 295X2
Vigile writes: NVIDIA announced its latest dual-GPU flagship card, the GeForce GTX Titan Z, at the GPU Technology Conference in late March with a staggering price point of $2999. Since that time, AMD announced and released the Radeon R9 295X2, its own dual-GPU card with a price tag of $1499. PC Perspective finally put the GTX Titan Z to the test and found that from a PC gamer's view, the card is way overpriced for the performance it offers. At both 2560x1440 and 3840x2160 (4K), the R9 295X2 offered higher and more consistent frame rates, sometimes by as much as 30%. The AMD card also only takes up two slots (though it does have a water cooling radiator to worry about) while the NVIDIA GTX Titan Z is a three-slot design. The Titan Z is quieter and uses much less power, but gamers considering a $1500 or $3000 graphics card selection are likely not overly concerned with power efficiency.
I'm so glad that I got the gaming bug out of my system when a ridiculously-priced video card was $300, and mainstream cards were in the $90-160 range...
This is ridiculous.
Do not look into laser with remaining eye.
These cards should have been tested from the perspective of high performance computing or scientific application.
don't underestimate the beauty of a quiet powerful computer.
I won't buy a $3000 gpu anymore than I'll buy a $1500 one, but I did buy the GTX 780 over the cheaper but somewhat more powerful R9 250 solely on the basis of it being cooler.
I'm god, but it's a bit of a drag really...
Do games these days typically take full advantage of such setups? I haven't really paid too much attention to gaming/hardware in the past few years, but it seemed as if support for dual GPU's was less than stellar.
IE, the only true advantage was an increase in the memory available to apps -- computationally, very few games took advantage of the additional gpu.
Has this changed, or (equally likely) I am completely off base on the state of afairs ?
The Titan shouldn't be considered a top-end gaming card. It should be treated as a budget Tesla card - even at $3k, it's the cheapest card in Nvidia's lineup with full double-precision floating point performance (which no game uses, but is common for scientific computing, Tesla's market). And on tests using that, the single-gpu Titan and Titan Black outperform the 295X2 by a large amount. AT hasn't gotten to test a Titan Z yet, but you can tell it's going to wipe the floor with the 295X2.
Yes, Nvidia advertised the original Titan as a super-gaming card, and to be fair it was their top-performing gaming card for a while. But once the 780 Ti came out, that was over, and since everyone expects a 790 dual-GPU gaming card to be announced soon, buying any Titan for gaming is a fool's choice.
Nvidia seems to still be advertising it as a top-end gaming card, presumably trying to prove the old adage about fools and their money. It just comes off as a scam to me, but anyone willing to spend over a grand without doing some proper research probably deserves to be ripped off.
The titian line in not a purely gaming GPU! The higher price comes for leveraging it's GPGPU CUDA technology. It's like buying a server hardware and complaining it doesn't run your games and well as an i7 which costs less. Game enthusiasts always ruin hardware news with their one golden spec, the frames per second! "That said it’s clear from NVIDIA’s presentations and discussions with the company that they intend it to be a compute product first and foremost (a fate similar to GTX Titan Black), in which case this is going to be the single most powerful CUDA card NVIDIA has ever released. NVIDIA’s Kepler compute products have been received very well by buyers so far, including the previous Titan cards, so there’s ample evidence that this will continue with GTX Titan Z. At the end of the day the roughly 2.66 TFLOPS of double precision performance on a single card (more than some low-end supercomputers, we hear) is going to be a big deal, especially for users invested in NVIDIA’s CUDA ecosystem." - AnandTech
They don't. What they need this for is ghetto floating point development hardware. This is cheap by those standards and offers far more precision than consumer grade GPUs.
ASIC makers have done if for the government. GPU mining is a thing of the past.
Prices still haven't come down though.
People don't buy the highest performing video cards for gaming, they buy them mining virtual currency.
Keep that in mind when you see that great price for a used high end card. The card probably ran for an extended period of time over clocked to just under its "melting point" and just got replaced by an ASIC miner.
Double precision floating point hardware, designed to do things like physics simulation. This thing has no ECC and some other similar tradeoffs, so it's fairly cheap at only 3k.
Here's an example of a non-ghetto version: http://h30094.www3.hp.com/prod...
Man I wish I could be that dumb.
People that mine either mine scrypt style currencies that still run better on GPUs or they are using ASIC miners for at least a year already. Used high end cards are either NVidia which are sold because the gamer wants something new or is short on cash, or AMD when the owner wants a faster GPU for either gaming or scrypt coin mining. For scrypt coin mining on AMD, overclocking the GPU doesn't work, in general you have to clock down a bit unless you are lucky and you can overclock the memory enough to maximize output with the GPU at standard clock rates.
The highest performing video cards tend to cost so much more for their performance, that miners get slightly less performing cards for half the price of the top of the range. You can get the R9 270 and R9 290 cards for much less money than the R9 290x and now the R9 295x2. The amount of coins you can mine for the purchase price and power consumption are such, that you don't want to buy those "highest performing video cards" if all you do is mine.
The reason you can buy relatively new video cards from miners is because the prices fell after MtGox fell. The get rich quick thing didn't work out and they all need money to pay their power bills. These cards aren't burnt up technically, the miner's wallet is empty and he needs some way to recuperate part of his loss. Sure, some of those are highest end cards, because the miner didn't pay attention to the price/performance thing when he bought them, but most will be high in the midrange, especially since we've seen some new high end cards come out since the prices of crypto coins fell.
I was promised a flying car. Where is my flying car?
I would just like to point out that the 295X2 has superior absolute gaming performance and superior fp32 performance but, just like most gaming NVidia products, the fp64 is crippled at 1/8 fp32 rate at configuration in order to create a profit margin for the costlier "pro" products. The hardware itself is capable of 1/2 fp64 rate and should be superior to the Titan Z if AMD decides to offer "pro-level support".
As proof, consider the fp64 rate of the single-chip AMD W9100, sold at ~$4000, which is 2.6 TFlops (http://www.amd.com/Documents/FirePro_W9100_Data_Sheet.pdf), versus the 2.7 TFlops of the Titan Z (1/3 fp32 rate, see http://en.wikipedia.org/wiki/G...). AMD could unlock the 295X2 at its full potential 5.2 double precision TFlops and release it any day if they want, crushing the Titan Z.
Honestly, instead of the Titan Z, I'd rather buy the AMD W9100 for $4000 and get equivalent double precision compute rate, better perf/W and, most importanty, certification for pro applications and ECC memory. That is certainly worth the extra $1000 in this product segment.
You're arguing with the antecedent. I'm saying "if you care about X, the Titan is good", and you're accusing me of cherry-picking because the Titan is bad at Y and Z, even though I specifically called it out as not being good for anything except X in a performance-per-penny measure.
I am saying that one of the principal reasons to buy a Titan is if you have a heavy double-precision compute load. I then provided a benchmark showing that a Titan beats the 295X2 in such a load. It would be cherry-picking if I picked the one double-precision benchmark that showed the Titan in a good light, but a single-precision benchmark does not invalidate that.
If you are accusing me of cherry-picking, please provide a benchmark that shows a 290X beating a Titan in a double-precision workload. AFAIK the only double-precision benchmark Anandtech uses is the F@H benchmark I linked to originally.
I am not at all arguing that the results in the double-precision benchmark somehow invalidates the single-precision or integer results. If your workload isn't mostly double-precision, the Titan is not for you. But if your workload *is* mostly double-precision, the Titan is a viable card.