20 replies to this topic

[Bad Karma 308]

UPDATED: Nvidia is supposedly readying a GeForce Titanium video card for February.

"said to offer about 85 percent of the performance of a dual-GPU GeForce GTX 690."

http://www.maximumpc...a_graphics_card


31-Jan-13 News Update:

Nvidia's upcoming GeForce Titan could end up faster than a GeForce GTX 690.

http://www.maximumpc...usly_though2013

Edited by Bad Karma 308, 31 January 2013 - 06:52 AM.

[Barbaric Soul]

Anyone looking to buy a XFX Core Edition 7970 towards the end of next month?

Seriously though, 2560*1600, yeah I can use the extra muscle, plus this thing will be a number crunching beast, which will help my WCG results. I'll probably buy one once they become widely available, and throw my 7970 in my crunching rig. Not sure what to do with the GTX560 in that right now though. Probably sell it.

[Catamount]

Barbaric Soul, on 23 January 2013 - 01:03 PM, said:

Anyone looking to buy a XFX Core Edition 7970 towards the end of next month?

Seriously though, 2560*1600, yeah I can use the extra muscle, plus this thing will be a number crunching beast, which will help my WCG results. I'll probably buy one once they become widely available, and throw my 7970 in my crunching rig. Not sure what to do with the GTX560 in that right now though. Probably sell it.

Considering what this thing will probably cost, I wish I had the money to just say "yep, I'm just gonna go snatch one"

I think I'll just stick to my heavily OCed 7970

[Barbaric Soul]

Catamount, on 23 January 2013 - 02:48 PM, said:

Considering what this thing will probably cost, I wish I had the money to just say "yep, I'm just gonna go snatch one"

I think I'll just stick to my heavily OCed 7970

Around the $800 mark. Tuition reimbursment from college is the only way I'll be able the afford one. Considering I have a family history of cancer, alzphimers, and lukemia, it's worth it for the added performance in cure research(WCG crunching). I won't be buying it for it's gaming performance, that'll just be a plus.

Edited by Barbaric Soul, 23 January 2013 - 03:54 PM.

[Catamount]

Do you really think a Kepler-based GPU will come anywhere near Tahiti, given the massive GPGPU disparity between them? I'm having a hard time believing that it'll be even close to as fast as the 7970 you already have (and if GPGPU-supporting number-crunching apps support Crossfire, which I admittedly don't know anything about the state of, wouldn't a second 7970 be a far better investment?)

Edited by Catamount, 23 January 2013 - 04:55 PM.

[Bad Karma 308]

Catamount, on 23 January 2013 - 04:54 PM, said:

Do you really think a Kepler-based GPU will come anywhere near Tahiti,

The current GK104 architecture on the GTX680 trades blow for blow in different benchmarks with the 7970 GHZ editions (or heavily overclock 7970). The GHZ edition was a stop gap measure to push the 7970 up to the same level with the 680; only it sells for a $100-$200 less. What you need to look at is that the GK110 chip in the article is the next generation. I'd almost expect them to call it the GTX780 and build a new line off of the GK110 platform. If their expectations are correct then this is a 35% jump in performance over the last gen; which is quite healthy and respectable.

However, I think ATI has a contender already waiting in the wings. They've only recently announced the mobile units so far but the AMD 8000 series is apparently on the way. http://www.tomshardw...hmark,3382.html

You have to remember that the competition is pretty high between ATI and Nvidia. Usually when a new flagship is released from one company the other waits a bit to play with tweaking the specs, and then releases theirs, and usually just a shade faster.

We use GPGPU very heavily at work. And we've found that Nvidia's enterprise level product tend to perform significantly better in that arena. Now that may change once Microsoft's "Accelerator" mainstreams and pushes both CUDA and OpenCL out and level the playing field. Our usage of OpenCL left us thinking that AMD entered the GPGPU market late and are still playing a bit of catch up. But we routinely retest as new products are released, so who knows, maybe the 8K series will be a winner.

Edited by Bad Karma 308, 23 January 2013 - 07:30 PM.

[Barbaric Soul]

Two 7970 cards won't be faster than a Titan and a 7970. Remember, the 7970 will go in my other crunching comp with a i7 860

[Catamount]

I don't know, unless you're using something that's very specifically tailored to Nvidia's architecture (that 5% or so of applications where both AMD and Nvidia cards work, but Kepler doesn't get trounced), two 7970s would be faster than a 7970 and a Titan, because the Titan would have to be something like 50-100% faster than a 680 to make up the GPGPU disparity (just to break even), which I posted in some detail on back in May. As of that post, the 7970 (NON-GE, mind you), either trounced the 680 or cleanly held its own against it. In the few the 680 did better in, it was by much tinier margins that were effectively more like ties, unlike the 7970's performance victories that were generally by very large margins. Of all the links I gave there, tests where the 680 won by any kind of real margin comprised a tiny, tiny fraction of what was out there, and on average, across all the tests, the 7970 (non-GE) more or less handed the 680 its own rear end on a platter. Since the cards themselves haven't changed, I highly doubt this situation has changed any in the past eight months.

Besides, for an $800 expenditure, you could have three 7970s (maybe even four, if you were very good with the used market) instead of a 7970 and a Titan, again admitting that I don't know the state of Crossfire support in GPGPU (but I assume multiple GPUs aren't a problem for the software; many GPGPU applications lends themselves to almost limitless parallelization).


Don't get me wrong, Kepler GPUs are great (almost bough a 670FTW before I saw a deal I couldn't pass up on a 7970), but despite Nvidia's massive early lead in GPGPU, and the fact that they do have better software support still (afaik), Tahiti just has gobs more actual GPGPU horsepower than any GPU Nvidia has out right now. Hell, if I'm not mistaken, Fermi is vastly superior to Kepler in GPGPU most of the time. One of those review sites I linked to noted that the 680 is only about as fast overall as a 560TI in those applications. So unless you're getting something that's actually going to massively favor Nvidia, AMD is just the way to go right now. You'll get far more bang for your buck, and frankly, if you are spending that kind of money, and do need an Nvidia GPU, a 590 would probably still end up being the vastly better investment, since it would probably kick the Titan's *** into next week (again, strictly speaking in terms of GPGPU).

Edited by Catamount, 24 January 2013 - 09:02 AM.

[Catamount]

Keep in mind that all this is admittedly based on eight month old data, so if something has changed since then, then I'd love to see it.

[Bad Karma 308]

I remember those article from when they came out. And I absolutely agree with you about the current GPGPU generation. But I did stipulate my comment with

" And we've found that Nvidia's enterprise level product tend to perform significantly better in that arena."

We use massive scale GPGPU enterprise systems for modeling and simulation. Other than a few Fermi and Firepro cards in workstations, we're talking about cluster banks of 1U Tesla's filling 48U racks. From an engineering standpoint it isn't feasible to house the comparable amount of HTPC systems just to accommodate 2-3 AMD GPUS in a 4U chassis compared to up to 4 in a 1U Nvidia chassis. If I'm populating 100 racks then I can get 4-5 times more computing power from Nvidia than from AMD per rack. So until AMD comes up with a similar scale enterprise solution then it's not worth looking at for us.

Also, from our perspective, Nvidia has the win thanks to CUDA being so easy to develop for, that and the already massive developer base to pull from.

We do have a several FireProS10000 laying around for a few project we contracted on and they do work admirably. But like I said should AMD come up with a decent enterprise solution then my firm would be heavily interested in the product.

Edited by Bad Karma 308, 24 January 2013 - 03:00 PM.

[Catamount]

Well I think AMD is seeing enough success to finally be looking towards that kind of thing, but I don't blame them for not going gung ho with developing such products up to now, since, well, let's face it, their GPGPU products sucked. All of their professional grade products were considerably behind their Nvidia counterparts for a long time (though maybe Evergreen based ones were good; I can't say I use them enough to make comment).

Now that they can compete, real Enterprise grade products it might be worth doing for them (perhaps with the 8000 series, or the new 2014 products; and knows how the new cards next year from Nvidia and AMD will compare, GPGPU wise).


I agree, of course, that for your situation it's still more worth going with Nvidia (but then, are your GPUs Kepler based, or Fermi based?); I was commenting strictly on home crunching machines. The Geforce Titan or whatever it's called will be a beastly single-GPU card for gaming (and probably the basis for one HELL of an Asus Mars card), basically a super high-end solution for people who don't want to do SLI/Crossfire, but it's not a good home GPGPU card. Kepler just isn't fast enough for a small number of them to compete with a small number of Southern Island GPUs. Even the 7870 is apparently faster in GPGPU than the 680.

Edited by Catamount, 24 January 2013 - 05:03 PM.

[Bad Karma 308]

Catamount, on 24 January 2013 - 05:02 PM, said:

I agree, of course, that for your situation it's still more worth going with Nvidia (but then, are your GPUs Kepler based, or Fermi based?);

Right now we have a mixed bag of of both Kepler and Fermi. We've been playing with some pre-release vendor equipment based on the GK110, it has proven quite a jump for our particular applications so we already have a large order in for hardware as soon as it comes off the manufacturing line in enough qty.

My firm runs a distributed simulation environment for the DoD, so whenever people want to use or work with our resources, or contribute to our environment, we usually wind up designing and spec'ing datacenter sized computational facilities. Many of our contracts can run in the 10 figure range so a lot of the manufacturers and vendors eagerly provide us early access to pre-release engineering samples of hardware. A lot of times our inputs and desires as well as design funding go into the hardware.

Edited by Bad Karma 308, 24 January 2013 - 06:43 PM.

[Barbaric Soul]

Catamount, on 24 January 2013 - 08:55 AM, said:

I don't know, unless you're using something that's very specifically tailored to Nvidia's architecture (that 5% or so of applications where both AMD and Nvidia cards work, but Kepler doesn't get trounced), two 7970s would be faster than a 7970 and a Titan, because the Titan would have to be something like 50-100% faster than a 680 to make up the GPGPU disparity (just to break even), which I posted in some detail on back in May. As of that post, the 7970 (NON-GE, mind you), either trounced the 680 or cleanly held its own against it. In the few the 680 did better in, it was by much tinier margins that were effectively more like ties, unlike the 7970's performance victories that were generally by very large margins. Of all the links I gave there, tests where the 680 won by any kind of real margin comprised a tiny, tiny fraction of what was out there, and on average, across all the tests, the 7970 (non-GE) more or less handed the 680 its own rear end on a platter. Since the cards themselves haven't changed, I highly doubt this situation has changed any in the past eight months.

Besides, for an $800 expenditure, you could have three 7970s (maybe even four, if you were very good with the used market) instead of a 7970 and a Titan, again admitting that I don't know the state of Crossfire support in GPGPU (but I assume multiple GPUs aren't a problem for the software; many GPGPU applications lends themselves to almost limitless parallelization).


Don't get me wrong, Kepler GPUs are great (almost bough a 670FTW before I saw a deal I couldn't pass up on a 7970), but despite Nvidia's massive early lead in GPGPU, and the fact that they do have better software support still (afaik), Tahiti just has gobs more actual GPGPU horsepower than any GPU Nvidia has out right now. Hell, if I'm not mistaken, Fermi is vastly superior to Kepler in GPGPU most of the time. One of those review sites I linked to noted that the 680 is only about as fast overall as a 560TI in those applications. So unless you're getting something that's actually going to massively favor Nvidia, AMD is just the way to go right now. You'll get far more bang for your buck, and frankly, if you are spending that kind of money, and do need an Nvidia GPU, a 590 would probably still end up being the vastly better investment, since it would probably kick the Titan's *** into next week (again, strictly speaking in terms of GPGPU).

You can't really use the GTX680 for a comparison since the GTX680 was made strictly for gaming applications. It was not designed for number crunching, which is why the 7970 beats it so bad in GPGPU. For a good comparison, look at the K20 and K20X Tesla GK110 GPU card.

Also, the main deciding factor whether I actually do get one is reviews of the card after it's release.

[Catamount]

Somehow, I think the $450-$500 GTX680 will be a closer comparison to what is basically an $800 680 Ultra (that should also be gaming-focused) than a $3500+ professional GPU. Remember, even if this new card is massively more powerful than a 680 (like 50-100%), it would still only roughly break even with the 7970, for twice the price. To be notably better, it would have to be an ungodly leap in power over the 680, and I still see all of zero chance of it having a GPGPU advantage that would equal its price tag.

As is always the case, we'll just have to see, but I don't see this card being anything notable in GPGPU, let alone $800 notable.

Edited by Catamount, 24 January 2013 - 09:14 PM.

[Bad Karma 308]

The real question for people on this forum is how well it will handle MWo.

And what we should all be really happy about is this is signalling that the next generation of cards are on the way, which should also move current GPU inventories into better price points for those wanting/needing to save a few bucks.

Edited by Bad Karma 308, 24 January 2013 - 09:20 PM.

[Barbaric Soul]

Catamount, on 24 January 2013 - 09:12 PM, said:

Somehow, I think the $450-$500 GTX680 will be a closer comparison to what is basically an $800 680 Ultra (that should also be gaming-focused) than a $3500+ professional GPU. Remember, even if this new card is massively more powerful than a 680 (like 50-100%), it would still only roughly break even with the 7970, for twice the price. To be notably better, it would have to be an ungodly leap in power over the 680, and I still see all of zero chance of it having a GPGPU advantage that would equal its price tag.

As is always the case, we'll just have to see, but I don't see this card being anything notable in GPGPU, let alone $800 notable.

You missed my point.The GTX680 has a GK104 GPU in it. the Tittan has a GK110 gpu. The Tesla K10 has two GK104 gpus, the K20 and K20X has one GK110 gpu. There is your comparison. GK110, while not as fast as two GK104 gpus, it isn't far behind it.

http://hexus.net/tec...k110-wheres-at/

Edited by Barbaric Soul, 25 January 2013 - 03:58 AM.

[Dr Killinger]

Damn, it costs more than my entire rig. I hope some smaller cards come riding in on it's coattails, I have a house to pay off!

[Catamount]

Barbaric Soul, on 25 January 2013 - 03:57 AM, said:

You missed my point.The GTX680 has a GK104 GPU in it. the Tittan has a GK110 gpu. The Tesla K10 has two GK104 gpus, the K20 and K20X has one GK110 gpu. There is your comparison. GK110, while not as fast as two GK104 gpus, it isn't far behind it.

http://hexus.net/tec...k110-wheres-at/

Yes, I can see the comparison. Nvidia took a $3500 professional GPU and crammed more double-precision processing power in there because it was sorely lacking. Architecturally, the card has no real difference from the GK104 (according to this very source).

I still see no evidence that a gaming version of the GK110, for $800, will magically be 4+ times as powerful as the 680 in GPGPU, based on design decisions for Tesla cards. You're forgetting that the K10 was already a GPGPU beast, in its own right, yet none of the K10's capabilities seemed to make it into the 680. Why should one expect the K20's GPGPU capabilities land in the 680 Ultra?

Edited by Catamount, 25 January 2013 - 05:56 AM.

[Apoc1138]

Catamount, on 25 January 2013 - 05:55 AM, said:

Yes, I can see the comparison. Nvidia took a $3500 professional GPU and crammed more double-precision processing power in there because it was sorely lacking. Architecturally, the card has no real difference from the GK104 (according to this very source).

I still see no evidence that a gaming version of the GK110, for $800, will magically be 4+ times as powerful as the 680 in GPGPU, based on design decisions for Tesla cards. You're forgetting that the K10 was already a GPGPU beast, in its own right, yet none of the K10's capabilities seemed to make it into the 680. Why should one expect the K20's GPGPU capabilities land in the 680 Ultra?

no one is saying it will be 4 times the performance

I can find GPGPU comparisons of multiple GTX680's vs a K10 (which as already pointed out is 2 x GK104 cores on a single card)... 2 GTX680's are roughly equal to a single K10... if a single core K20 also matches a K10... it is not a massive leap to assume that a GK110 chip will match another GK110 chip regardless of the package it comes on (seeing as the GTX680 packaging isn't limiting it either)

the price difference is the level of support you get along with the card, not the actual card itself

Edited by Apoc1138, 25 January 2013 - 08:15 AM.

[Catamount]

Apoc1138, on 25 January 2013 - 08:11 AM, said:

no one is saying it will be 4 times the performance

I can find GPGPU comparisons of multiple GTX680's vs a K10 (which as already pointed out is 2 x GK104 cores on a single card)... 2 GTX680's are roughly equal to a single K10... if a single core K20 also matches a K10... it is not a massive leap to assume that a GK110 chip will match another GK110 chip regardless of the package it comes on (seeing as the GTX680 packaging isn't limiting it either)

the price difference is the level of support you get along with the card, not the actual card itself

But that's precisely my point. Even if the GK110 based Geforce card was that big a leap over the 680, it would just make this card little more than a an $800 7970 in terms of GPGPU. That's hardly a good investment for a home crunching machine. To be worth it as a GPGPU card, for $800, this card would have to have 4 times the 680's capabilities, give or take. I just don't see that happening. I'm also not convinced that every GPGPU based design decision of the k20 will make it into a gaming card. Sure, you pay mostly for support for a professional product, but that doesn't mean the K20 wasn't purpose built just for GPGPU.

Edited by Catamount, 25 January 2013 - 09:55 AM.