Shootout: 780 Lightning vs 290

[ICDP]

If you could clarify this, that would be great. Because i've read this at more than a couple of websites.

Powertune dictates how much power draw the GPU can use before it starts throttling, not the actual voltages. So upping powertune to 150% simply tells the driver to be less strict about TDP limits (within reason). This reduces the chance of throttling and in all honesty anything over 20% is mostly pointless from my testing.

This link explains powertune 2.0 very well.
http://semiaccurate.com/2013/12/16/amds-powertune-2-0/

To change voltage we can simply use Precision X, GPU Tweak or MSI Afterburner. Ironically they upped the voltage on the GTX780 as this can only be done using one of the above utilities, yet chose not to up the voltage on the R9 290. This test cannot be used to determine max OC potential of the R9 290/X. My own testing shows that ~1170 should be the average on air and ~1220 on water using the reference PCB.

 

[wand3r3r]

There seems to be some confusion about 780 scaling, hopefully despite being at 1080p I can clear that up a bit.

900/1500 @ 1v 260w peak

1100/1700 @ 1.1v 330w peak

23% core OC netted 21% increase in performance. This was with Rome 2, which was necking on me despite not going higher with the OC. An OC which is actually lower than what this 780 GHz boosts to out of the box (but still has stock vram).

Nvidia doesn't just pick a random memory frequency, the cards are fairly well balanced. Increasing core without increasing memory will yield poorer scaling than increasing both.

Are you implying a valid conclusion can be made for "780 scaling" based on a single data source from a single game?

Not only is that impossible to extrapolate with the single data source, there are variations as clocks go higher along with memory scaling and ratios. Cherry picking comes to mind as well (whether or not that game happens to scale better then usual, making broad statements based on a tiny set of data is questionable)

I'm not saying I don't value the data, rather that it doesn't put anything to rest.

 

[ICDP]

There seems to be some confusion about 780 scaling, hopefully despite being at 1080p I can clear that up a bit.

900/1500 @ 1v 260w peak

1100/1700 @ 1.1v 330w peak

23% core OC netted 21% increase in performance. This was with Rome 2, which was necking on me despite not going higher with the OC. An OC which is actually lower than what this 780 GHz boosts to out of the box (but still has stock vram).

Nvidia doesn't just pick a random memory frequency, the cards are fairly well balanced. Increasing core without increasing memory will yield poorer scaling than increasing both.

The problem with the vast majority of recent GTX780 cards is that they use Elpida VRAM which seriously reduces OC potential. I tried 2x MSI Gaming GTX780 and 160-180MHz offset was the maximum VRAM overclock possible without causing lock-ups. OC of 18% gave only 12% real world performance increase.

Tomb Raider 2560x1600 Ultimate preset
Stock: 1047/6008: Avg FPS 52.7
OC (voltage mod 1.212v) : 1241/6372: Avg FPS 59.6

This max OC only just brought my GTX780s up to stock R9 290X performance. My R9 290X overclocks to 1170/1400 17% (Core OC 12% VRAM OC) gives me ~16% real world performance.

Tomb Raider 2560x1600 Ultimate preset
Stock: 1000/1250: Avg FPS 61
OC +100mV: 1170/1400: Avg FPS . 71.2

Of course the silicone lottery applies and none of us should be talking in absolutes when it comes to OC potential. I have seen many folks claim as absolute fact that GTX780 will overclock better than R9 290/X. Not when we consider the fact that almost all GTX 780s are coming with Elpida VRAM and that means the chances of getting a great overclocking GTX780 are vastly reduced.

 

[BallaTheFeared]

Are you implying a valid conclusion can be made for "780 scaling" based on a single data source from a single game?

Not only is that impossible to extrapolate with the single data source, there are variations as clocks go higher along with memory scaling and ratios. Cherry picking comes to mind as well (whether or not that game happens to scale better then usual, making broad statements based on a tiny set of data is questionable)

I'm not saying I don't value the data, rather that it doesn't put anything to rest.

My card at 1320/1775 is faster than the Lightning used in this thread at 1440/1675 in Valley.

My card at 1306/1775 is faster than the Lightning used in this thread at 1333/1650 in 3Dmark11.

My card at 1215/1775 is faster than the Lightning used in this thread at 1320/1650 in BioShock Inf.


There are only so many comparisons I can make at 1080p with a 4.8GHz Haswell before cpu limitations start to devalue the data. Even Rome 2 was limited by the cpu with the OC.

 

[wand3r3r]

Thanks for the comparison and for trying to remain impartial to the brands! I'm sure it took considerable time and effort. It's always fun to have more data to consider and compare.

 

[BallaTheFeared]

25% increase in Tomb Raider.

Edit: 1280/1775

42% increase over 900/1500, and I believe 900/1500 is higher than the euro site gets with their non "uber" 780.

almost all GTX 780s are coming with Elpida VRAM

I just got two different 780 GHz both were B1 and both had Hynix. Pretty sure most of the higher end boards are coming with Hynix, while the 780 Ti's are soaking up probably all of the Samsung chips.

 

[Imouto]

Are you locking the frequency to 900 Mhz? Because that way the GTX 780 would start even lower in the charts.

 

[BallaTheFeared]

Boost is disabled, it does what I tell it or it gets the hose again.

 

[Zanovar]

haha

 

[BallaTheFeared]

I thought 900/1500 was a good baseline, these websites are testing with the reference 780s they got a long time ago.

 

[ICDP]

I just got two different 780 GHz both were B1 and both had Hynix. Pretty sure most of the higher end boards are coming with Hynix, while the 780 Ti's are soaking up probably all of the Samsung chips.

I'm not disputing your results, I'm disputing the idea that they are standard for anyone who purchases a GTX780.

1047 Boost clock at standard
Card 1: ~1217 Boost clock with modded voltage BIOS
Card 2: ~1240 Boost clock with modded voltage BIOS

That is a 16%-18% overclock and it gives 10%-12% real world performance increase.

It should also be noted that disabling boost and proclaiming a 42% increase is disingenuous as no GTX780 operates that way. The only valid metric is how much of an OC you get on a card by card basis based on that actual cards original stock boost clocks.

In order to achieve those 900MHz non-boost clock you actually have to underclock your card. Underclocking by ~20% and then overclocking by 40% from that point is not proving anything other than you deliberately under clocked your cards well below stock.

If I were to underclock my R9 290X to 800MHz then overclock to 1200 so I can claim a 50% overclock would that be acceptable?

 

[jj109]

I thought 900/1500 was a good baseline, these websites are testing with the reference 780s they got a long time ago.

Testing with GPU boost is super slippery. HardOCP's 780 didn't exceed its 900 MHz boost clock for some reason. TPU's 780 boosted up to 1007 MHz in some games.

 

[ICDP]

I thought 900/1500 was a good baseline, these websites are testing with the reference 780s they got a long time ago.

Reference GTX780 boost clocks vary from ~960 MHz - 1GHz at stock. The point is that 900 is not a good baseline, it is well below reference speeds.

http://www.anandtech.com/show/6980/nvidia-geforce-gtx-780-overclocking-results

Notice how the stock boost clock was actually 1006MHz?

http://www.altesc.net/2013/06/15/gtx-780-overclocking/

Again notice how the stock boost clock was 1005MHz?

http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_780/30.html
Boost clock up to 1007MHz. Seeing a pattern yet?

It has become increasingly common to measure GTX780 OC results compared to a baseline number that no card released ever drops to. I know from experience GTX780 cards are potentially great overclockers, that is not a gaurantee and many here imply that it is (this is not aimed at you).

 

[BallaTheFeared]

I'm not disputing your results, I'm disputing the idea that they are standard for anyone who purchases a GTX780.

Based on what, exactly? Bought any recent revision cards? Any HoFs, Classifieds, Lightnings? Or did you buy the cheapest 780 on the market as a point of reference?

1047 Boost clock at standard
1241 Boost clock with modded voltage BIOS

That is an 18% overclock and it gives 12% real world performance increase.

12% more than you had!

It should also be noted that disabling boost and proclaiming a 42% increase is disingenuous as no GTX780 operates that way. The only valid metric is how much of an OC you get on a card by card basis based on that actual cards original stock boost clocks.

Actually it isn't all all, especially if you manage to follow the train of thought that accompanies it as well as the graphs.

We're discussing non reference 290OC performance being compared to reference 780 performance and what that means for perf/clock disparities. It's also why we're discussing vram in conjunction with core clock.

Reference 780s used in reviews are found below 900MHz, and we're talking about clock speed scaling with GK110, so it wouldn't matter if I was using 500MHz as a point of reference or 250MHz. However 900MHz made sense because that's close to the boost Nvidia was shooting for when using 1500MHz vram.

In order to achieve those 900MHz non-boost clock you actually have to underclock your card. Underclocking by ~20% and then overclocking by 40% from that point is not proving anything other than you deliberately under clocked your cards well below stock.

I'm getting the feeling you have no idea what is actually being discussed.

If I were to underclock my R9 290X to 800MHz then overclock to 1200 so I can claim a 50% overclock would that be acceptable?

 

[ICDP]

I'm getting the feeling you will do anything to prove Nvidia = the best. Now you underclock your cards 20% to prove they can overclock 42% and have the stupidity to believe it is a valid test. My point about underclocking my R9 290X to 800MHz is due to the fact that many reviews showed R9 290X cards throttling to that level in games with quiet mode.

I can't simply take a low MHz number and assume it is a good baseline to prove to myself my card is a great overclocker.

Wow, just wow.

 

[Imouto]

It's easier than that. HardwareCanucks observed an average clock of 927 Mhz for the reference GTX 780. An 1.32 Ghz OC on the GTX 780 Classy netted a 22% performance increase in Crysis 3 and 24% in Tomb Raider.

That 42% performance increase is unrealistic, made up and misleading.

 

[Termie]

Simply put, referring to the published 900MHz "Boost Clock" as the baseline operating speed of a GTX 780 is wrong. That is a MINIMUM boost clock, and will almost never be seen in the real world. It's like saying an R9 290X is going to run at 700MHz.

Balla and Blackened - you're just undermining your arguments by claiming that this is the baseline. There's no doubt that the GTX 780 overclocks well, but it's not starting at 900MHz, and it certainly isn't starting at 863MHz.

On the flipside, ICDP, the claim that all 780s have Elpida RAM is also simply wrong. You certainly had a low overclock of 6350, but many cards are hitting 6800 and above. The only card for which reviews have specifically stated Elpida RAM is used is the Lightning. Most cards do MUCH better in RAM overclocking: http://www.techpowerup.com/reviews/Palit/GeForce_GTX_780_Super_JetStream/29.html

I previously demonstrated a 19.6% gain for my GTX780@1215/6800 (in-game frequency) vs. reference (1006/6000) at 1440p across many games: http://forums.anandtech.com/showthread.php?t=2357336

Several games are CPU-bottlenecked at 1080p, by the way, including Hitman and Crysis3.

 

[blackened23]

Erm, it is the baseline. But that baseline gets adjusted when you buy an aftermarket card which is specifically binned for higher clockspeeds, I would think. For instance, if you buy a reference GTX 780 it will probably be in the 970-980mhz area for out of box boost. Whereas the SC ACX is binned higher so 1000-1100 is the out of box boost. So on and so forth. The nice side effect of this is, your boost is guaranteed even if you get a reference card, it isn't the "up to" boost - you always get that boost in games. As well, let's say you throttle 1-2 bins in crysis 3. You're still higher than the guaranteed spec boost of 900mhz. The only annoying thing is, voltage is not unlocked with GPU Boost 2.0. You can do it through trickery, sure, with afterburner. But that's the only annoyance of GPU Boost 2.0. Fortunately you can get around that with software or by buying some gee-whiz aftermarket card like the classy.

In any case, I don't really care anymore. I think the GK110 has much more overclock potential/scaling, but this thread has largely turned into a shouting match between two sides which are firmly entrenched in their views. Me included of course. Kinda silly that I got caught up in this, and I don't really care, so anyway, back to enjoying christmas food and what not. Hopefully you guys as well are enjoying the day as well even if we enjoy politely shouting at each other now and then.

 

[Termie]

snip

Kinda silly that I got caught up in this, and I don't really care, so anyway, back to enjoying christmas food and what not. Hopefully you guys as well are enjoying the day as well even if we enjoy politely shouting at each other now and then.

Good reminder that this is all just fun and games, and that we should all be grateful that we have the opportunity to "shout" at each other over things that luckily don't matter too much.

Happy holidays everyone!

 

[BallaTheFeared]

I'm not arguing absolute performance, only scaling of the core for performance.

How did that get lost in translation?

That 42% performance increase is unrealistic, made up and misleading.

It's 100% true, easily verifiable via video. Would you like that? Would you like to me post a video?

Tomb Raider is a perfect showcase for the GPU since it has basically no cpu overhead.

 

[Grooveriding]

You guys should take look at the 780ti owner's thread at OCN. The results are interesting as far as memory overclocks, at least in the standard 1080p benchmarks such as Heaven and 3DMark.

The 780 nets some good gains from memory overclocks, with the better clocking ones getting up around 7-7.1Ghz. The 780ti starts at 7Ghz and some of the better ones get up around 7.6-7.8Ghz+ The ti owners are not seeing the sort of gains from overclocking their memory that high that 780 owner's did from their memory overclocks, core clocks are much more important once you have those high 780ti memory clocks. GK110 seems to hit a sweet spot of gains around 7ghz on the memory.

Of course this may not hold true for actual gaming where you are running at 1600p+ resolutions. But as far as something like Valley where memory overclocks were a big deal on the 780, they do not make as much of a difference on the 780ti with its 7ghz baseline on the memory.

 

[Zanovar]

Interesting for sure but OCN overclocks 2 min runs.:S

 

[BallaTheFeared]

40% scaling from 900/1500 to 1306/1775

900/1750

1306/1775

 

[BallaTheFeared]

Simply put, referring to the published 900MHz "Boost Clock" as the baseline operating speed of a GTX 780 is wrong. That is a MINIMUM boost clock, and will almost never be seen in the real world. It's like saying an R9 290X is going to run at 700MHz.

I'm just using one of the websites that has been used quite often in the past.

I noticed they weren't coming up in my other post so I'm hosting them myself.

Also my intention was to show scaling, not perf gains as a whole over all reviews and all 780s.

 

[Elfear]

Say what?. The stock 290 in reference form is basically tied with the 780.

You're right. The review I linked with the reference 290 leading by 10% only had 5 games. With a larger sample of games the 290 only leads by 2% at Computerbase, 5% at Hardware Canuck, and 5% at TPU (1600p). Not all of the review sites listed the average boost clocks observed with the 290 or the reference 780. Hardware Canucks stated "Now, before every run, cards are left to “heat up” for 10 minutes under in-game load which will ensure more realistic benchmark numbers." and "The R9 290’s default setting retuned much better results with consistently high frequencies around the 900MHz mark."

For the reference 780, they observed average boost clocks of 927Mhz. So according to them a 900Mhz 290 is 5-6% faster than a 927Mhz 780. If we assume 80% scaling with overclocking for both cards, a 1300Mhz 780 would be approximately equal to a 1240Mhz 290.

IMO that is superior overclock scaling as compared to the 290/X which is struggling to hit overclocks at around 1125-1175 from every review i've seen.

I think we're using two different definitions of the term "scaling". Sounds like you're talking about max overclocking headroom. In your previous post you said:

That said, this doesn't change the fact that the vast majority of reviews published for overclocking on the 290X/290 vs GK110 heavily favor the latter for maximum overclocks, overclock scaling, and clockspeeds above stock.

Are you talking about the same thing in the bold above? Generally "overclock scaling" refers to how much performance you get from overclocking. So if you overclock by 20% and gain 20% more performance you have perfectly linear scaling (i.e. 100%).

If you mean that GK110 generally has more oc headroom than Hawaii, I agree. If you mean it scales better with overclocking, my testing doesn't show that. Balla did point out that my memory overclocks weren't all that great but then again neither were that of my 290. Hard to compare the two though because they have different bus widths.

What we need is for someone with a 780 to test a slew of games keeping the core clock at stock speeds and increasing memory speeds to see what gains come with more bandwidth.