**Nvidia GeForce GTX 970 / 980 Review Thread**

[andybird123]

Yes exactly, my comments are all based on 20nm as all the rumours are that titan2 will be on 20nm
We dont have to agree to disagree as you are disagreeing with something I never said

Have you been channeling Charlie by any chance?

 

[CAT-THE-FIFTH]

I agree with this. Since the GPC are arrange in clusters of 4 SMM and since the gtx 980 is made of 4 GPC, you have 16 SMM or 128x16 = 2048 shaders. SO something scaled up should contain additional blocks up of 1 GPC or 4 SMM more. Unless they change the configuration again like they did with the gtx 750 ti(which had 5 SMM in one GPC).

So scaling up and considering the current die size(398mm2), they should be able to fit 6 GPC or 24 SMM or 3072 shaders, 96 ROPS, 384 bit bus if they tighten the transistor density a bit. They will need to do this because these large chips they focus on double precision performance, which adds a tad more to the size. So I am guessing a slightly bigger maybe 620mm2 chip, with 30-40% more performance than this chip. But of course 1200 dollar price tag or something ridiculous. It's going to be hard to get good yields from a chip like this and we are likely to see something with just 20-25% percent more performance from the cut down version meant to be sold to most people.

Yes,but you are both forgetting that the GM200 will have better DP performance which adds to the transistor budget massively.

The GK110 had twice the number of transistors as the GK104 and was nearly 90% larger in surface area. Yet it was only around 50% faster at 2560X1600 according to TPU.

Maxwell is already more massively more efficient at compute performance over kepler, Kepler had bad SP performance and really awful Dp performance, it will take much less work to convert small maxwell in to a DP monster than it did with GK104 to GK110

The GM204 still has poor DP performance though:

http://images.anandtech.com/graphs/graph8526/67748.png

Remember its a 33% bigger die and a 1.5 billion more transistors(42% increase) over the GK104.

With 40% transistors the GK110 is getting close to DOUBLE the DP performance of a GM204.

The GM204 only appears to have a 10% improvement in density over the GK104(or a bit higher).

It does not still change what I said though. Adding DP performance adds a HUGE amount of transistors,and ultimately Nvidia will want a BIG improvement too with the GM200.

You forget Intel is encroaching onto their turf with MIC(and AMD has even seen their pitiful marketshare in the segment grow a reasonable amount).

So if the GM200 is on 28NM it will be the biggest GPU in history. A real monster on all levels IMHO and it will be bloody fast.

Yet a GM204 is still more efficient as a gaming GPU as it does not need to do all the DP stuff as well,and hence you can optimised far better towards gaming performance and better gaming performance/mm2.

I'm assuming Titan2 will be 20/16nm, they would be mental to try to do it on 28

GM204 has 1/32 DP performanceover keplers 1/24, hence the small improvement, but it wouldnt take a huge amount to improve that, so yes improving dp performance will eat in to mm2 allowance, but not by as much as it did with gk104>gk110

Ifnyou put a Titan in to full DP mode it has 1/3 dp performance, but maxwells SP performance is so good you could trim that back to 1/6 and still have a beast of a card

Even at 1/6 it would still be worse overall on 28NM. You are still not realising that the GM204 is 33% larger than a GK104 and has 42% higher transistors. Maxwell has slightly higher transistor desnity than Kepler,but overall DP performance has increased primarily down to them just using more transistors and higher clockspeeds(actually closer to 50% since I looked at the graph again).

A GK110 with 40% more transistors and 40% higher die area has nearly double the DP performance of a GM204. These are based on published figures.

Every single higher end Nvidia GPU from the GF100 onwards has shown the same thing. Adding additional DP performance eats into the transistor budget,and adds more power consumption.

The GF100,GF110 and GK110 all did not scale uniformly in gaming performance when compared to the GF104,GF114 and GK104. Yet the GPUs below the latter three did. The big difference is the "big" GPUs added functionality which is not required for gaming and this eats into power consumption,TDP,die area and transistor budget. You cannot change the laws of physics within the same uarch using the same process node.

We will have to agree to disagree.

Edit!!

On 20NM its a moot point. You get reduced power consumption and much higher transistor density anyway,so you can offset the ineffiencies of having added non-gaming relevant functionality added.

Yes exactly, my comments are all based on 20nm as all the rumours are that titan2 will be on 20nm
We dont have to agree to disagree as you are disagreeing with something I never said

Have you been channeling Charlie by any chance?

Yes exactly, my comments are all based on 20nm as all the rumours are that titan2 will be on 20nm

Have you been channeling DM by any chance?

I don't need to channel DM,I have my own ranty style which is patented!!

If Titan II is 20NM,it makes me wonder how long it will be before release and how many they can get out initially.

Good old Apple and companies will be hogging 20NM as much as they can,since Intel is trying to flood the market with cheap Atoms,which is annoying.

Edit!!

But honestly next time try not to do the strawmans and try and keep to what I said,since it is based on Nvidia figures and sites like Anandtech not made up ones. You knew very well the comment I was answering was talking about 28NM.

I also own a Nvidia card and had more of them this gen than AMD ones.

It also does not change the fact that DP has always added significantly to transistor budget and its been replicated by the GF100,GF110,GK110,Tahiti and Hawaii. History is on my side on this.

You are basing your calculations on non-DP optimised parts and it has never added up for ANY of the last few generations. Hence they are rather pointless.

A scaled up GM204 at 620MM2 would probably be a faster gaming GPU than a GM200 at 620MM2,providing you don't hit ROP or memory walls.

DP does not add to gaming performance or gaming performance/mm2. It hinders it meaning you need to spend more transistors and more die area at the same node to get similar gaming performance.

FFS,even the arrangement of the compute units will be different as gaming and compute workloads are different.

This is why what Nvidia is doing is clever. The GM204 is a fine example of a GPU optimised for gaming and consumer workloads not DP stuff which is more for supercomputers and certain commercial work.

 

[CAT-THE-FIFTH]

Repeat post.

 

[almighty15]

Guys call me stupid but which 970 do I order if I plan on using a full cover block? Can't find any reference to a 670 with a reference PCB ??

 

[opethdisciple]

970 is tempting.

670SLI 2Gb vs single 970?

 

[Freddie1980]

Guys call me stupid but which 970 do I order if I plan on using a full cover block? Can't find any reference to a 670 with a reference PCB ??

Wait until full covers are available and us EK's cooling Configuration tool to check which cards are compatible or take a risk and buy the most reference looking card available.

 

[andybird123]

snip

I was the first in the chain to suggest that Titan2 could be 50% faster than Titan1
I did this on the assumption that Titan2 would be 20nm

it wasn't clear if you other 2 were talking 20nm or 28nm, as soon as you mentioned 28nm I interjected to point out that my original post assumed 20nm

I didn't create a strawman, I simply clarified that I had always been talking 20nm (again, because all of the rumours point to 20nm for big maxwell)

All of the "history" you are looking at is comparing same process, with a die shrink I think they can get a bump in performance without busting through die size budgets
I don't see the point in continuing to debate 28nm for big maxwell when I'm working on the assumption that it will be 20nm, if 28nm then you are absolutely right, they will have to dump DP performance as a requirement if they want to improve gaming performance, which would end up with a totally seperate GTX vs. Tesla part, which is something they've never done, so if they want a new king of the roost tesla part, it will have to be 20nm

I don't know what the point of the "I have an nvidia card" comment was, I never mentioned AMD?

 

[Noxia]

Really looking forward to breaking my dual GPU cherry, hope it's not a disappointment.

 

[CAT-THE-FIFTH]

The GTX670 SLI will have more peaky performance. Potentially it might beat the GTX970 in some games and in ones where SLI scales poorly it will be slower.

 

[ihatelag]

Seems like a worthy upgrade from my 680 SLI

Not unless you upgrade the rest of your rig, it will be more of the same. Your 680 GTX is still good to go for a few more years and not breaking a sweat.

If you have an upgrade itch that you can only scratch with a 980 then get it.

 

[Freddie1980]

I probably going to get bashed for going against the flow here, but seriously? We are getting excited by the 980...why? I mean putting the impressive low power consumption aside...these are "next gen card", have our expectative really sinken so low in terms of performance increase?

Yes the 980 does look to be faster than the 780/780Ti, but while I looked at the bench comparison, it looks to me that the 980 rely too much on the high stock clock to outperform the old cards. I mean especially comparing the frame rate of 780 and the 980, it look as though had the 780 been actually on the same clock as the 980, it wouldn't really be behind it much, if at all.

In a way, it is a bit like comparing the i7 4770K to i7 4790K, and then going crazy because the 4790K is faster (due to higher stock clock) and lower temp. I simply don't get it.

But yes I think that the low power consumption would benefit those who want to SLI/Crossfire but don't have a powerful enough PSU previously, but for those that built their PC with SLI/Crossfire in mind from the start, they'd most likely already got a powerful enough PSU already.

The way I look at it is if Nvidia can achieve the performance of a 780ti with GTX680 power consumption what performance can they achieve with a TDP equivalent to that of the 780Ti or even a new Titan.

 

[Silent_Scone]

Meh, I've bought one for now. EK store has no stock of blocks yet though.

 

[melmac]

If you understood the industry, you would understand they are are deep deep trouble.

(r9 285)Tonga is the successor to (gtx 7970)Tahiti the same way the gtx 980 is the successor of the gtx 680.

Looking at their characteristics they are very similar, power, size of die, and even shaders(2048). They starkly difference in performance which is the biggest item of value. The gtx 980 is 75% faster(maybe 64% faster than full tonga). This is bigger than most generational gaps with a node thrown in for good measure.

AMD hawaii(r9 290x) part was AMD attempt to make a monolitic chip good at computing to compete with the gk110(gtx 780 ti).

AMDs new part coming out that is supposed to be their performance part is called fiji and is upscaled Tonga(r9 285). They are based on the same architecture and are part of the same series called pirate islands.

This will be the successor to hawaii or r9 290x. This as you said wouldn't be bad if you didn't care about power or heat and Nvidia didn't release anything else. However Nvidia is releasing an update to gk110 called gm200 that uses the same architecture maxwell. Most rumors have it coming out earlier than Fiji.

So what you think is going to make a better performance part.
A part scaled up from r9 285 or a card scaled up from the gtx 980. And looking at the size and power consumption of both, they both can be scaled up 50%.

This is why AMD is screwed.

This is the type of beating Intel gave AMD 8 years ago with core 2 duo and it cratered their CPU division.

wow, for a person claiming to understand the industry you sure get a lot of stuff wrong.

AMD aren't screwed, the high end isn't even their target market. Look at their business model since the 4xxx series cards, they offer very competitive price/performance at the mid range levels and lower and then put out a really powerful dual card for the extreme high end. The only exception to this was the 7970 because Nvida's GK100 part never made it and their midrange gk104 part became their high end.

The 290x was brought out to compete against the 780.

You got the compute thing all backwards. Second generation Kepler was NVidia trying to counter some of the dominance AMD had in the mining craze. And Maxwell has improved on that again.

You don't understand AMD's GCN either. The 290x is the successor to the 7970.

I don't think you are "teaching" anybody.

 

[mmj_uk]

The high end has always been AMD's target market, the trouble is they are increasingly struggling to compete there meaning they're having to make their money at the lower end. Both Intel and NVidia have done a lot of work on power efficiency (on top of increasing performance) whereas AMD are still stuck in the growth paradigm, they're having to ship their parts with AIO watercoolers which is fairly embarrassing.

 

[Ayahuasca]

People seem to discredit power efficiency as well saying they don't care about it but in the end it saves you money over the hours and hours you use your PC and it also links directly to lower temps and more overclocking headroom.

 

[Marine-RX179]

People seem to discredit power efficiency as well saying they don't care about it but in the end it saves you money over the hours and hours you use your PC and it also links directly to lower temps and more overclocking headroom.

Bottomline is the loss people make on selling their existing 780 or 290 serious plus the extra isn't going to be compensated any time soon with the few pounds saving per years they get with the lower power consumption card.

Anyone that got cards under the 780/290 level these new cards are probably worth the upgrade, but for people that already got 780/780Ti/290/290x it make very little sense for the sidegrade, unless selling their existing card will fully fund the new cards, or people that got only 650W PSU and crossfire/sli 290/780 is might be out of question without upgrading the PSU, then it might worth them consider selling the 290/780 to get a pair of 970.

 

[RavenXXX2]

The GTX480 was not that long a go and look how Nvidia have turned it around.

AMD need to work as hard as Nvidia has to turn things around, Nvidia are making them look stupid currently.

 

[Gregster]

The GTX480 was not that long a go and look how Nvidia have turned it around.

AMD need to work as hard as Nvidia has to turn things around, Nvidia are making them look stupid currently.

Spot on. The 480 was a hot hungry beast and nVidia messed up but learnt their lesson. AMD will bounce back and hopefully soon.

 

[Marine-RX179]

The GTX480 was not that long a go and look how Nvidia have turned it around.

AMD need to work as hard as Nvidia has to turn things around, Nvidia are making them look stupid currently.

It's actually 4 and half years ago already

But I don't think AMD can afford 4 and half years...they will need to do something and something SOON.

May be they could go the opposite direction as Nvidia and go "f power efficient, we go down the performance improvement route instead"? Cause if you look at it objectively beyond the lower power-consumption and temp, reasonble price, performance wise it is actually at a standstill comparing to the 780/780Ti despite being new gen. At stock the 970/980 might seem on par on slightly faster comparing to the 780/780Ti, but when both gen cards are overclocked, the 970/980 isn't really faster than the 780/780Ti...

 

[chaparral]

GTX 980 : Techpowerup > http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_980/

LOL at there last 4 words........