GeForce Titan coming end of February

[sontin]

Nice try. GTX480 had more functional units than M2050.

http://www.nvidia.com/docs/IO/105880/DS-Tesla-M-Class-Aug11.pdf

It appears it was based on the GTX470, with 448 CUDA cores. GTX480 was a 480 CUDA core part, with more TMUs and ROPs as well. K20X and Titan are rumored to share 2688 CUDA cores and 224 TMUs. Are you suggesting the ROP count has been neutered from 56 to 48 then, or the card has a TDP of 250 but draws more than that in gaming like the 480 did? Most amusing of all that 250W TDP on the 480 was a useless metric since the card drew 270W+ in actual games.

So, you are telling me that it was possible with the GTX480 to have a 39% higher core clock, another SM and 15% higher bandwidth to get to 250 Watt but it would be not possible for TITAN without more active SMX?

 

[blackened23]

So how did we get to this point? Ugh.

 

[tviceman]

I have been saying this for a long time but it gets ignored. Can someone explain to me how you can increase GPU clocks by 39% (1019 / 732), increase GDDR5 from 5.2ghz to 6Ghz (288GB/sec memory bandwidth) and end up with a 250W TDP card?

Geforce parts always use less power than their equivalent Tesla parts at the same clocks. Non-ECC memory uses less power. The chip itself will have much of the compute features turned off or dialed down, thus considerably less transistors will be switching when under load.

 

[tweakboy]

$899 ?

Seems legit.

Your kidding me 899 ,,, thats just way too high.

Look at video cards from 90's and most of 2000 era

The expensive top of the line card has always been like 600 dollars. hmmmmmm

 

[BallaTheFeared]

Nice job tviceman, you killed the thread!

 

[RussianSensation]

So, you are telling me that it was possible with the GTX480 to have a 39% higher core clock, another SM and 15% higher bandwidth to get to 250 Watt but it would be not possible for TITAN without more active SMX?

Did you miss the entire comparison I made regarding real world power consumption of GTX670 as well? I didn't just discuss it from 1 angle. I looked at it both from K20X's 235W TDP (and if that # is too high, which means it may not be accurate to assess real world power usage), and I also looked at it from the GTX670 angle. You have 2688 CUDA cores which is double that of 670's and real world power consumption only goes up 64% assuming the Titan uses 250W in games but is on the same 28nm node? The 480's TDP and power consumption are irrelevant in this case since the card broke 250W in games. Therefore, it's TDP was in no way related to its real world power usage.

Geforce parts always use less power than their equivalent Tesla parts at the same clocks. Non-ECC memory uses less power. The chip itself will have much of the compute features turned off or dialed down, thus considerably less transistors will be switching when under load.

How can a 500mm2+ chip be clocked at 1019mhz and use just 250W of power but the HD7970GE that is just a 365mm2 chip is clocked at 1050mhz and draws 238W of power in games at TPU? Does that imply that NV's 28nm tech is far superior to what AMD is using for their cards? Does NV use much higher quality / less leaky transistors that they can overclock a 500mm2+ chip to 1Ghz at less than 1.256V of HD7970GE? All I am doing is asking basic questions. What about GTX680 that had just a 294mm2 die and 1058mhz GPU clocks? You yourself said it's more realistic for the clocks to be around 900mhz range than 1Ghz+. GTX680 draws 185W of power in games despite only having 1536 CUDA cores, 2GB of VRAM and a much narrower and less power hungry 256-bit bus. It doesn't hurt to ask and question the specs. I guess I am one of the few that is more skeptical.

I guess I am not buying into the hype like everyone else this time. The specs for next gen flagship NV GPU have been overhyped on this forum consistently: GTX480/580/680. The performance/specs for none of them lived up to their pre-launch hype/leaks. If NV delivers 1019mhz clocks in a 250W max power usage in games and retains all the functional units of K20X (all the ROPs, TMUs), I'll be pleasantly surprised.

 

[SirPauly]

Will see when or if anything is accurate or semi-accurate!

What peaked my interest was possibly a new AA mode!

 

[ocre]

I think if i where nvidia i would take advantage of the time between the original gk104 launch and the gk110 geforce. I would start with swapping a few transistors in the mix as well as other process related improvements to get the TDP down as low as possible. With a whole year, its kinda hard to imagine that there wouldnt be some critical improvements.

I have no doubt that the power consumption will be high for this card (230+ watts). But i also feel like the calculations per watt will be improved over the gk104. With all this time, it is actually an expectation of mine. I dont think this is a bad thing at all. For those who wanted very low power solutions, they have options. But i think there are a many number of people who have been waiting for the power that the titan will bring. They have been waiting since the 28nm cards launched

 

[SirPauly]

The constructive nit-pick to me wasn't the TDP of the first fermi GeForce sku's but the performance per watt -- will see how efficient the Kepler monolith may be for GeForce.

 

[tviceman]

How can a 500mm2+ chip be clocked at 1019mhz and use just 250W of power but the HD7970GE that is just a 365mm2 chip is clocked at 1050mhz and draws 238W of power in games at TPU?

Back your roll up there, I said I don't think Titan will be running at 1019mhz. And you agreed with that. My power draw statement was pointing out that it's not apples to apples to compare Tesla to Geforce.

It's also not entirely apples to apples to compare different chips from different companies, especially when you start talking about die sizes.
Cayman - 389mm^2 - 185 watts peak, 287 watts max power draw
GF110 - 520mm^2 - 226 watts peak, 317 watts max power draw

GF110 was 33% bigger, but only had a 22% higher peak power draw and 10% higher maximum power draw vs. cayman. See what I did there? I made Nvidia's chip look like VOODOOtm magic based on your fallacy die size and different chip architecture argument.

EDIT: Nvidia's Tesla m2090, which had the same operational units as the gtx580, ran at 650mhz and had a tdp of 225 watts. gtx580 ran at 772mhz and had a 244 watt tdp. That is a 19% increase in clock speeds, and only an 8% increase in TDP.

 

[f1sherman]

big Tesla TDP vs small GeForce TDP:

Entirely different workloads, disabled DP parts, much more stringent requirements, ability to run 24/7, bigger safety margin/tolerance (must not fail no-matter-what)

more on Tesla/Geforce/Titan TDP. Take it or leave it:

The K20c is rated 225W, which it hits only when running Linpack. A single-precision only load results in a power draw of ca 120W. Since the Geforce Titan will almost certainly have the DP units disabled, and assuming the silicon is stable at 1GHz, a 50-60% increment on the GK114/GTX680 is entirely plausible.

http://semiaccurate.com/forums/showpost.php?p=177570&postcount=82

 

[RussianSensation]

BIt's also not entirely apples to apples to compare different chips from different companies, especially when you start talking about die sizes.
Cayman - 389mm^2 - 185 watts peak,
GF110 - 520mm^2 - 226 watts peak,

GF110 was 33% bigger, but only had a 22% higher peak power draw and 10% higher maximum power draw vs. cayman. See what I did there? I made Nvidia's chip look like VOODOOtm magic based on your fallacy die size and different chip architecture argument.

GTX580 to 6970 is not the same as the point I made for a couple reasons:

1) GTX580's GPU clock was 772mhz vs. 880mhz for the 6970. Here I am comparing nearly identical GPU clocks of 1019 and 1050.

2) GTX580 not only had less VRAM, but it was clocked much lower than 6970's. HD6970's GDDR5 was 5.5ghz not 4Ghz. The Titan goes from 3GB of VRAM of 7970 to 6GB and retains the same 6Ghz GDDR5.

Notice what you said above. 33% larger die size and 22% higher power draw. 1019mhz 502mm2 Titan would imply 38% larger die size and a 5% increase in peak load power in games (250W vs. 238W). BTW, I never claimed that you agreed with the 1019mhz clocks. I mentioned that earlier

EDIT: Nvidia's Tesla m2090, which had the same operational units as the gtx580, ran at 650mhz and had a tdp of 225 watts. gtx580 ran at 772mhz and had a 244 watt tdp. That is a 19% increase in clock speeds, and only an 8% increase in TDP.

19% increase in GPU clocks and 8% increase in GDDR5 speed of 2090 vs. 580 is not comparable to a 39% increase in GPU clocks and a 15% increase in GDDR5 speed the Titan supposedly carries over the K20X.

big Tesla TDP vs small GeForce TDP:

Entirely different workloads, disabled DP parts, much more stringent requirements, ability to run 24/7, bigger safety margin/tolerance (must not fail no-matter-what)

more on Tesla/Geforce/Titan TDP. Take it or leave it:

http://semiaccurate.com/forums/showpost.php?p=177570&postcount=82

Geforce parts always use less power than their equivalent Tesla parts at the same clocks. Non-ECC memory uses less power. The chip itself will have much of the compute features turned off or dialed down, thus considerably less transistors will be switching when under load.

Thanks!! Those was actually a great explanation with more technical details. :thumbsup:

 

[BallaTheFeared]

And GF100 ran at less than 1.05v, typically even below 1v (such as my 6+ samples).

6970 shipped at 1.175, Fermi's max voltage was 1.087v, that's max allowed without a bios flash.


Honestly RS, I have no idea why you're trying to draw parallels between Nvidia and AMD uArch wise when the only similarity really have is fab process.


Do you have any evidence that GDDR uses a lot of wattage?

http://originus.samsung.com/us/business/oem-solutions/pdfs/Green-GDDR5.pdf

^ Googled that, I have no clue personally!

 

[RussianSensation]

Honestly RS, I have no idea why you're trying to draw parallels between Nvidia and AMD uArch wise when the only similarity really have is fab process.

Not comparing the underlying architectures. Just looking at same nodes, GPU die sizes, their clock speeds and real world gaming power consumption. I'll stop talking about the power consumption now. No point dragging the same horse. I'll wait until specs

1019mhz with 2688 cores is 69% more shading power over 680. Under water at 1210mhz this card would have double the shading power of a 680. EPIC. I wonder how many GTX680/670s will be unloaded in the used market. Balla, this might be your chance to ditch those 470s!

 

[BallaTheFeared]

Yeah, sadly after getting laughed at for the better part of a year for calling the 680 a mid-range product it seems I'm going to have to settle for a used one on the cheap because this thing is priced way outside my comfort range. /\

Ah well, maybe I'll just wait some more and get a GK110 next year when Maxwell's mid-range thumps it... That will probably cost $500 too, hooray...

 

[Crap Daddy]

Yeah, sadly after getting laughed at for the better part of a year for calling the 680 a mid-range product it seems I'm going to have to settle for a used one on the cheap because this thing is priced way outside my comfort range. /\

Ah well, maybe I'll just wait some more and get a GK110 next year when Maxwell's mid-range thumps it... That will probably cost $500 too, hooray...

Yep. We've been tricked. Kepler vs. Fermi, about 80% increase in performance with about 100% increase in price.

 

[f1sherman]

Everyone agreed that it was mid-range part.
<300mm^2 / 256bit, 118% of GTX 580 transistors (GTX 460 has 40% more tran. than GTX 285)

But it makes no sense to call it mid-range if it's top-performer and NV has nothing better for another year.
The only thing that remains mystery is at what point NV decided to make gk104 top dog. And was that the plan all along or forced decision.

 

[BallaTheFeared]

Everyone agreed that it was mid-range part.
<300mm^2 / 256bit, 118% of GTX 580 transistors (GTX 460 has 40% more tran. than GTX 285)

But it makes no sense to call it mid-range if it's top-performer and NV has nothing better for another year.
The only thing that remains mystery is at what point NV decided to make gk104 top dog. And was that the plan all along or forced decision.

It makes perfect sense to me.


The rest is just arguments waiting to happen, so I won't beat that old war drum.

Good Day, Sir!

 

[Cookie Monster]

Not comparing the underlying architectures. Just looking at same nodes, GPU die sizes, their clock speeds and real world gaming power consumption. I'll stop talking about the power consumption now. No point dragging the same horse. I'll wait until specs

Thing is that the underlying architecture is probably the most important part of the discussion when trying to analyse and speculate about its potential power consumption.. and it happens to be the very subject where most forum-goers lack the knowledge in to really talk/speculate things that are meaningful.

On top of that, I tend to think that the less meaningful part of the discussion is the die size because all it tells us is that there could be more transistors on the chip and/or its not a densely packed layout and/or theres a difference in transistor density even though its on the "same" node etc (A reason why comparable AMD parts during the HD4800 days had considerably small die sizes as they were able to fit them very efficiently given the die area budget). Whether or not if affects the power consumption is another story, a good example of this is R600 vs G80 where the latter was clearly/physically the larger part.

Anyway.. not sure where the upgrade itch came from but I feel like I might just go buy this card depending on how it performs

 

[maniacalpha1-1]

Anyone seen this ?

http://videocardz.com/39536/nvidia-geforce-gtx-titan-to-be-released-on-february-18th

 

[tviceman]

GTX580 to 6970 is not the same as the point I made for a couple reasons:

1) GTX580's GPU clock was 772mhz vs. 880mhz for the 6970. Here I am comparing nearly identical GPU clocks of 1019 and 1050.

2) GTX580 not only had less VRAM, but it was clocked much lower than 6970's. HD6970's GDDR5 was 5.5ghz not 4Ghz. The Titan goes from 3GB of VRAM of 7970 to 6GB and retains the same 6Ghz GDDR5.

Notice what you said above. 33% larger die size and 22% higher power draw. 1019mhz 502mm2 Titan would imply 38% larger die size and a 5% increase in peak load power in games (250W vs. 238W). BTW, I never claimed that you agreed with the 1019mhz clocks. I mentioned that earlier

19% increase in GPU clocks and 8% increase in GDDR5 speed of 2090 vs. 580 is not comparable to a 39% increase in GPU clocks and a 15% increase in GDDR5 speed the Titan supposedly carries over the K20X.

Thanks!! Those was actually a great explanation with more technical details. :thumbsup:

Yeah I think it'll be 900 boost clocks with a 250 watt TDP. That is my guess.

 

[RussianSensation]

Anyway.. not sure where the upgrade itch came from but I feel like I might just go buy this card depending on how it performs

Well said, well said. All those other things you mentioned like transistor density and how those transistors work in particular chip do matter. I should have considered those characteristics in how they impact the power consumption.

Anyone seen this ?

http://videocardz.com/39536/nvidia-geforce-gtx-titan-to-be-released-on-february-18th

If GTX680 didn't exist, and NV was able to get GK110 out on time, they would have increased GTX480's (~GTX570) level of performance more than 2.5x (172% vs. 67%) in 1 generation if this card comes in a couple of percent under a 690. Impeeessiiiiiiiiiiiiiiive. This is nearly a repeat of 8800GTX (Titan) vs. HD2900XT/3870 (HD7970 Ghz), minus the $900 price tag.

2 or 3 of these in SLI is money in the bank for people who love super-sampling/downsampling. It's going to be a while before any single GPU can topple this monsta.

Very convenient that the new Crysis 3 and a pair of Unigine benchmarks are all coming this month too. This is going to be the month of benchmarking FTW! NV's large die strategy seems unstoppable. I can't even imagine now what the flagship 500mm2+ Maxwell is going to give us on 20nm if NV goes at this pace (2x on top of Titan by 2015 and then we'll be at 5x the performance of a GTX480 from spring 2010). In the CPU sub-forum they are excited about 10% IPC increase on quad-core Haswell. hehe. GPUs is where all da fun is at nowadays! :biggrin:

 

[BallaTheFeared]

I wonder if the features to do work without the cpu will be in Titan and if they'll actually reduce cpu overhead in gaming?

Ugh!!!!

You know it's going to cost a lot, specially designed cooler, no non reference, only EVGA and ASUS will carry it.


Probably have locked voltage too, Nvidia's hand is heavy this generation.


Hopefully AMD can get within 20-15% of this, for 1/2 the price or less, really stick it to Nvidia and this elitism I'm sensing.

 

[parvadomus]

I think people will have to wait for HD8000 series to get titan at a reasonable price

 

[RussianSensation]

I think people will have to wait for HD8000 series to get titan at a reasonable price

I am not so sure. I don't think HD8970 has any shot to hit 50-60% faster than HD7970GE unless HD8970 is 20nm Volcanic Islands. I don't think AMD will even come close to the Titan in a 1 die card until they hit 20nm. I would think the type of people buying $900 GPUs are also not the type who'll want to wait 12 months for something with similar performance for less money. Just look at the depreciation of cards like GTX280, 480 or 580. You can now get $499 GTX580 level of performance in an HD7850 OC for $150. This card is targeting those who want the best of the best as soon as it comes out. The same gamers are likely rocking 2 GTX680s/690, or similar in which case selling them and getting the Titan won't seem like a significant financial cost to side-grade to escape the SLI scaling requirement. Plus a lot of people just like to play with new toys. I expect some nice discounts on GTX680s in the used market shortly.

Hopefully AMD can get within 20-15% of this, for 1/2 the price or less, really stick it to Nvidia and this elitism I'm sensing.

How? Their official road-map has nothing until Q3 2013. 20% faster HD8970 by Q4 2013 won't matter anymore since the Titan will have ruled the market for more than 7 months. I think AMD should keep themselves busy on drivers and game bundles. They have no chance of touching the Titan unless they pull some desperate official HD7990 SKU as a counter. Their best bet is to focus on the $100-300 GPU market segments where the competition is the most fierce rather than waste 100s millions of dollars attempting to counter something as impressive as the Titan. The Titan is like an i7 3970X since it targets a very niche super-high end market segment. With AMD's engineering resources and financial situation, it doesn't make sense to take such risks with such limited financial rewards when the competitor has so easily beaten the best they have in the HD7970GE. Even if they do try to make some 500mm2 HD8000 chip, it might be a total failure, and they would have wasted 6 months of time and resources. AMD doesn't have the luxury of the $5 billion ATI to pull stunts like that. The risk is just too high. What they should do is focus on price/performance in lower segments under $500 with HD7000 since that's what their strengths are. Seeing where Titan lands must be a scary thought for them since if they aimed at increasing performance of HD7970GE by 75% with their 20nm parts, those GPUs might easily get rolled over by Maxwell if a 28nm Titan comes in 60% faster than HD7970 Ghz (!!). Releasing drivers on time, fixing DX10/11 stuttering with GCN memory mgmt rewrite, things like that still matter for HD7000 series since it's targeting the key mainstream market segments.


----
Videocardz reports that the GTX Titan would have computing power of 4.5 TFLOPS, with the GPU clock range of 800-900 MHz core clock. The memory may not reach the 6GHz GDDR5 mark. Paper launch on Feb 18th, with reviews the next day. They also state ROPs will be just 48. I thought GK110 with 14 SMX would have 56 ROPs since 8 SMX in GK104 is 32 ROPs (or 4 ROPs per SMX).