Introduction About a year ago now I put together my first review/testing thread and ended up having a lot of fun with it and getting a lot of good feedback from users here on the forums. The focus of that test was to find out how well the new-at-the-time Kepler architecture scaled at surround resolutions as you added cards and also to find out just how hampering the 2GB frame buffer that the reference 670's and 680's came with was going to be at those resolutions. A year later, we find ourselves fresh off the launch of another series of GPU's. While the popular question last year was 2GB vs 4GB, this year it has turned into 3GB or does GK110 really need 6GB to feed itself? Last year when the VRAM limitation questions were all over the hardware enthusiast community I was limited to having cards to test that were the lower memory versions of the choice in question. This year I come armed with the upper end of the choice in question, the 6GB Titans. Now, you might be wondering why a guy running Titans is posting this review in the 700 series forum. It is being posted here because it is intended to give potential GPU buyers the answer to the question that many people will come to this very forum to find the answer: is the GTX 780's 3GB of memory enough? In addition to that, let's be honest here, Titans and 780's are so dang close once everyone has them clocked up to max clock that I'm not sure the differences in actual GPU power are really even worth talking about. It's the same as the 670 FTW vs 680 ... 680's had a few more shaders, but 670 FTWs could usually clock higher. At the end of the day the difference was minimal and not even worth discussing. In this little writeup I hope to give you guys the tools and information to be able to answer that question yourself. The answer is not entirely definitive and will come down to what you as a gamer prefer. Do you like eye candy, or do you like performance? I will spend some time talking about four different games: Battlefield 3 Online, Crysis 3, Far Cry 3, and Metro Last Light. The resolution in use in this review will be 5760x1080. That is another purpose of this review. Surround users are often left hung out to dry when GPU reviews come out. In the event that some benches at surround resolutions are included, it is usually relegated to it's own page with very little actual discussion taking place on it, and
nobody seems willing to touch on the one thing everyone wants to know about, memory usage at these resolutions.
Without further ado, lets get to the test setup and methods!
Test Setup and Methods Here are a couple of quick pictures of the machine all this data was collected and compiled on:
Test Setup: CPU: Intel Core i7-3930K overclocked to 4.7 GHz
Motherboard: Asus Rampage IV Formula
Memory: 16GB Corsair Dominator Platinum 2133 MHz CL9 Quad Channel Kit at rated speeds
GPUs: EVGA GTX Titan SLI overclocked to 1163 MHz core, 6200 MHz memory, 314.22 Driver, both running PCI-E 3.0 x16
Power Supply: Seasonic X-1250 Gold
Operating System Drive: 256 GB Samsung 840 Pro SSD
Operating System: Windows 7 x64 Professional SP1
Monitors: 3x De-Matted Asus VG248QE at 144hz, 5760x1080
Methods: To record all of this data I used EVGA Precision X and FRAPS. I elected not to use Precision's FPS meter because you can't do so without using the RTSS.exe program that causes so many people problems. I have known it to cause performance decreases for me before, so to prevent this I logged FPS with FRAPS and all other data with Precision X 4.2.
With the exception of Metro Last Light, all data were recorded in actual game play. I used FRAPS to record a 5 minute run of each game on each different setting tested. For Battlefield 3, I used a 64-player Ziba Tower online round. I was able to gather data for all settings in the same round, so I was happy with that as far as limiting the massive amount of variables for a 64-player online game as much as possible. For Crysis 3 and Far Cry 3 I just recorded a repeatable part in the middle of the game somewhere that I encountered a good number of enemies and repeated the same sequence for each setting tested. Metro Last Light I did a little bit differently. I had not actually started the game yet, so I fired up last night hoping to find a good repeatable spot to bench early in the game, which I found doesn't really exist. So, in the sake of time, I just used the benchmark tool for Metro Last Light to record data the same way I did during the game play for the other games.
The methods are nothing fancy but seem to get the data that I'm after. Now let's move on to the results for each game!
Battlefield 3 Battlefield 3 isn't necessarily a new title, but still the game that I play by far the most. This section is as much for me as anyone as BF3 isn't going to be testing the limits of the 780's 3GB of VRAM, but I still wanted to bench it just to see how it performed on paper (plays incredible in feel). I gathered data for this game during the same 1500 ticket round that was played on Ziba Tower with at least 50 players in the server. I realize that a Close Quarters map isn't going to be the most demanding scenario in online BF3, but since I've been on a Close Quarters kick the last month or two I just benched it. The outdoor vehicle maps are going to be a little more demanding, but there is so much variation in performance on those maps due to so many things other than your GPUs. For instance, if I'm looking from RU to US on Firestorm I will be pulling about 130 FPS. If I'm looking the other way, from US to RU I will pull 80-90. Close Quarters seems to be more consistent to me, so in order to limit variations due to the directions I'm looking and things like that, I just stuck with Close Quarters.
Here is a screenshot of the settings used with varying levels of AA tested being none, 2x, and 4x:
And the results:
Now for the graph that I'm really interested in in this review, the performance vs VRAM usage.
As we can see, performance in BF3 with SLI Titans/780s is going to be quite good. I would expect that from a game that is a couple years old now. Also, even with BF3 maxed out, you won't be getting anywhere near the 3GB limit at this resolution on BF3. I can't speak for larger resolutions.
Metro Last Light As I mentioned in the methods section, this game was benched differently in that I just used the canned benchmark for the game since I hadn't played through and found a good place to bench in game.
Here's a settings screenshot where I tried Very High Tesselation, Very High Tesselation with SSAA, High Tesselation, and High Tesselation with SSAA:
And for the results:
I found out that varying levels of tesselation really don't do anything for performance, so for the performance vs usage graph I just used Very High Tesselation with and without SSAA. This is the latest game in the test, and it is interesting to see how different the curve is and how much lower on the performance side it is than BF3. Also of note, you won't be touching the 3GB VRAM limit of the 780 at this resolution in this game either.
Far Cry 3 This data was gathered somewhere in the middle of the Far Cry 3 map where I drove a series of winding roads back and forth encountering enemies periodically throughout the entire benchmark. It happened to be night in the game when I did these benches, but from what I can tell there isn't really a performance difference between daytime an nighttime in the game.
Here are the settings used with no MSAA, 2x, 4x, and 8x tested:
And the results:
This game, along with Crysis 3 which is next, were much more interesting for the purpose of this review. They were the only two games that approached the 3GB limit of the 780. Even then, I believe Far Cry 3 can be maxed with the 780's 3GB of VRAM. The only times 3GB is exceeded is right after I load into the level and the one 8x anomoly where there was a huge VRAM spike for a split second. As far as loading into the level, Far Cry 3 always seems very sluggish for the first couple of seconds after loading until everything gets going. You can see once it gets going that the VRAM levels out to around 2900 MB where it runs the whole time. In fact, at all levels the VRAM at the very beginning is slightly higher meaning I just should have probably cut that data out as it isn't indicative of the game play. As for the 8x VRAM spike, this happened when I did other runs, so it must be a particular spot in the map. Far Cry 3 is fairly buggy like that in that some spots on the map just don't perform well no matter what you're running.
What is interesting to me is the performance as it approaches 3GB of VRAM. It is becoming less and less playable, and this is where the judgement call comes in. How high of FPS you require will directly relate to how much VRAM you are going to be using. Let's move on to Crysis 3 and see what it tells us.
Crysis 3 For Crysis 3 I recorded data in the middle of the game in a part that I encountered a long series of enemies and a couple of automated turrets in an area that has a little bit of indoors and some outdoors as well.
Here are the settings with 1x SMAA, 2x SMAA, 2x MSAA, 4x MSAA, and 8x MSAA all tested:
And here are the results:
Now this game is an interesting one to look at because it can actually far exceed the 3GB buffer of the 780, but can it really? Look at the performance. This is where the preference comes in, but I'm not cool with 40 FPS. I'll try to sum it all up in the next section.
Summary and Conclusions Here are some summary statistics for you guys to chew on
Here is a graph that I made that puts all the memory usage vs performance from the different games on the same graph to see if we can get the big picture of what is really going on here. This shows all of the games and their performance vs the memory usage. I've also included a line at 3GB to represent the limit of the 780. So really, for potential 780 buyers, I guess you just need to look to the right of that line. Are those the levels of performance you want to play at? If so, then you can surpass the 3GB limit of the 780. I would say most people spending the kind of money we're talking on GPUs aren't doing so for 40 FPS, so for the vast majority 3GB is going to be plenty at this resolution and below.
Another thing that I think is very interesting is the path those lines have taken as games have evolved. We have the 2011 game in BF3, the 2012 games in Far Cry 3 and Crysis 3, and the 2013 game in Metro Last Light. As time has progressed, at least with these four games, the FPS per MB used has dropped. This is interesting because even with the 2012 games, it is pretty clear to me that you will not have the GPU power available to use beyond 3GB of VRAM unless you are fine with FPS in the 30's. That's with Titans OC'ed as high as they'll go in games which I would think would be on the very upper end of the best 780's and the performance they can achieve.
If it is indeed true that as games are updated and new game engines are built the FPS per MB of VRAM used is lower and lower, then buying 6 GB for the future doesn't really make any sense either. The 2014 games, the problem will be even worse. You will have even less performance relative to your VRAM usage, so the GK110's won't be anywhere near powerful enough. The one caveat to this is that Metro Last Light just does not appear to use much of anything for VRAM. I can't image this will be indicative of all 2013 games. The one that I can't wait to get my hands on that could prove all of this moot is ARMA 3, but I don't have my hands on it today unfortunately.
All of this kind of confirms what I've long suspected, and that is that nVidia knows what they're doing and there is a reason the 780 got 3GB just like there is a reason the GK104 cards have 2GB. nVidia builds these things and knows their limitations. I really don't believe you can take advantage of more than 3GB of VRAM with the GK110 chip for the vast majority of people because they don't have the muscle. GK104 chips probably can't really take advantage of anything past 2GB because they don't have the muscle. All I'm saying is there is a reason the reference memory buffers are what they are.
IMO, 3GB is enough for you if: - you run 5760x1080 or lower and require above 40 FPS averages
- you are running two or less cards
You may actually surpass 3GB if: - you are running three or more cards
- you are satisfied with sub 40 FPS averages
- you are running surround 1440P/1600P (even then I think you would have to have 3 or 4 cards to be playable above 40 FPS)
Hope you all enjoyed the read and hopefully it brought some hard evidence to the table and can help people answer the question that many have for themselves. I may rerun the tests at a lower clock (1054 or so) to see how the results change since that is presumably a level of performance almost all 780's could achieve, but that will be dependent on time. Thanks for reading and bring on any questions or comments!
post edited by thebski - 2013/06/20 14:42:59