After absolutely swearing I would skip Titan, save thousands of dollars, and jump off the deranged “super enthusiast train”, I went ahead and bought 3 Titans. Yes, I need help. I am taking recommendations for multi-step programs that can cure me of this sickness before it leads to a life of ruin and (more) regret. In the mean time, I have 3 GTX680 4GB cards up on eBay and have been running benchmarks 🙂
First, a Refresher
Just to recap, the overall system configuration is as follows:
- Motherboard – Asus X79WS
- CPU – Intel Extreme 3960x @ 4.6Ghz on Corsair H70
- RAM – Corsair Dominator 1866Mhz 16GB matched kit
- Storage – 2 x Crucial M4 512GB 6Gb/s SATA 3 in RAID 0 (Intel RAID), LG BluRay writer
- Audio – Creative Labs X-Fi Fatal1ty Titanium Pro PCI-E
- PSU – Thermaltake Tough Power 1200
- Displays – 3 x Viewsonic VX2268WM 120Hz 3D Vision
- Case – Lian Li PCA77F, modded for 2 x 70mm side intake fans, all fan bays full
- Miscellaneous – Logitech G510/G13, Razer Naga, 3DVision 1.0, 2 x Scythe Kaze Master Pro fan controllers
- OS – Windows 8 Professional, new install
The outgoing video subsystem was a trio of 4GB GTX680’s – 2EVGA, 1 Zotac. In honor of the new Windows 8 install, and as a “final hurrah”, I wanted to retest them and report the results here to compare. I will be comparing the following games in both 2D and 3D surround:
- Metro 2033
- Crysis 3
- Farcry 3
My methodology was to use FRAPs benchmarking for comparison for Crysis and Farcry, and use the built in demo mode for Metro 2033 (all FPS games should have that IMO… Kudos to the Metro team). Before getting into the gaming results I’ll spend just a brief time on the actual Titan GPUs and the physical build.
Why Hello There Titan (x3) !
The first thing about Titan is that you really have to be very deliberate about not dwelling on the obnoxious pricing. Starting to think about all of the other things you can do with that money, or about the starving children in the world, will send you directly back to the store. So I wrote another check to UNICEF, took a nice stiff drink, carefully avoided looking at the invoice, and dove into the physical build. Incidentally I got these at MicroCenter on sale for $999 (store pickup).
With no further adieu, I present the worlds most ridiculously expensive (and fastest) single GPU’s!
As the photo clearly shows, I ended up with 2 SC cards and one standard. Oddly enough all $999 and MC. So a great deal on the SCs, and a pretty much normal deal on the regular despite all being “on sale”. Incidentally, the cards look absolutely identical in every way besides that “SC” sticker including the boxes and pack ins. Of course I would have opted for 3 x SC if they had them. I put the standard card at the bottom of the stack (figuring the SC cards should be at least from a better bin and hold up better to heat), and they all ended up overclocking equally at an 89C thermal target, 103% power target, 50Mhz clore offset and 50Mhz RAM offset.
Unboxing the cards, I discovered that $1000 doesn’t buy you any better packaging or swag. That’s no problem really as this stuff just ends up in the basement anyway until it is time to sell, but it would have been neat to have gotten some sort of spruced up presentation:
Next I tore out the GTX680s and slapped in the Titans. Quick and painless. They occupy the same amount of room, use the same power connectors, and basically seem pretty much exactly the same. Many have spoken of the Titan build quality, and they are a bit nicer, but the 4GB GTX 680 boards were fairly decent as well so it isn’t like stepping from a Honda Civic to a Rolls Royce Silver Ghost by any stretch of the imagination. A quick shot of some case pr0nz:
First Up – Metro 2033 – “The Widowmaker”
Metro is just a horrible, horrible workload. It seems to be one of those cases where there will simply never be hardware that will truly be able to just brush this game off at the highest settings and resolutions. Maybe it was just far ahead of its time, or maybe it was just really poorly optimized. Either way, surround resolution in Metro is just brutal even in 2D. Here is how the 680’s did in 2D:
This is actually decent performance and is the way I play Metro. Realistically those catastrophic minimums seem to only happen in the Frontline demo and, when framerates do drop off during real game play, they are momentary. The overall subjective real-world experience is pretty smooth.
Now let’s look at 3D:
This is a lot uglier. We can see the mins are really apocalyptic at 5 and the average is now down to 33fps. Predictably about half of 2D which makes sense since it is twice the pixel load. Again, on the subjective side, I do fine that 3D is playable for most of the game. It is noticeably less smooth than 2D obviously, and the hitching is more pronounced with clear slow down during fire fights, but it isn’t like old school 2D shooter slowdown on an NES. Still, it definitely leaves you wanting more and feeling sort of let down after spending the big money for a 3 x 4GB GTX 680 rig. Incidentally on Metro 2033 I see VRAM usage just topping 2GB at about 2.1GB in 3D.
With the 680 testing complete, it is time for the moment of truth! Let’s take a look at Titan numbers first in 2D:
First reaction is WTF!? The minimums stay at an identical 13fps. Something is very, very wrong in Denmark with Metro Frontline. Some day we will have 35,000 core GPUs that will still find a way to fall of a cliff at the midpoint of that benchmark. Clearly something there is broken. Looking at the average though, we are up to a massive 112FPS – 72% more performance than the 680s. This is a really impressive result and is even more impressive in game. From a subjective standpoint, the game goes from perfectly playable, to silk smooth bumping right up against the refresh rate of the 120hz displays. Incredible stuff. So riding high on 2D, let’s come back down to earth and look at 3D:
The good news is we have seen a 50% gain in our minimum! All the way up to 7FPS! All kidding aside, the good news here is that the average has climbed up to a whopping 48FPS. By the numbers this doesn’t seem like a ton, and it remains well shy of the magic 60FPS, but the subjective difference in game is transformational. Where 3D on the 680s was best kept as a novelty and overall was too uneven to play the entire gain through in, on the Titans, there is no reason to turn 3D off (short of a headache!). Real world performance stayed consistently closer to 60fps then the benchmark shows and the catastrophic drop-offs were fewer and farther between than the 680s. Really impressive.
So overall what can we conclude? I can sum up the Titans performance vs the 680s as follows:
- 2D surround goes from perfectly playable to silky smooth and is a noticeable gain, but more a “nice to have”
- 3D surround, on the other hand, is a “must have” and turns 3D from a novelty best sampled occasionally into a modality you can actually use for an entire play through.
- the Titans were rock solid and stayed cool and just a bit quieter than the 680s. Their utilization hovered around 97-98% with the occasional mysterious Metro drop-offs. The CPU was asleep as always in this game
Now are those results worth $3000? VERY hard to say and harder to justify. This is a very individual decision and much like with cars, those last 50HP are the most expensive!
Next Victim – Crysis 3… “Can it play Crysis???” becomes “Can it play Metro?!” Officially!
I have to say that I have been pretty much awed by the amazing coding of the Crysis team. I am not sure how lower end systems are faring, but at the bleeding edge of the high end, particularly 3D surround, these guys have worked miracles. Compared to Metro 2033, Crysis 3 is a breath of fresh air even on the GTX680s. Think about that one for a second! I’ve already spoiled it that the Crysis results are a happier tale then Metro, so let’s get to it. First are the settings used both base level graphics settings, and advanced:
I disable anti-aliasing because at these resolutions, in a fast moving FPS, I just really don’t see a benefit from it. Also, traditionally it isn’t like any of these cards really have horsepower to spare in these scenarios (particularly in 3D). Adjust expectations accordingly if using these results to gauge performance and you like cranking AA up. For the Crysis tests, since I used FRAPs benchmarking, I put together min/max/avg and “time to render” graphs comparing 2D and 3D. Here is how the 680s held up frame rate wise:
Absolutely incredible performance when you consider the visual quality of Crysis 3. What these numbers represent is a run through of the first mission from the point you are given control through after you clear the large multi-level room (big firefight) inside the factory and are about to go back outside. From my experience this first mission is one of the most intense in the game and definitely much more brutal than the outdoor clear weather landscape that comes later.
From a subjective standpoint, both 2D and 3D feel very smooth and playable. The averages are dragged down by the minimums that occur during the hitching that hits in the beginning of that intense firefight in the large warehouse room (after you get the bow). Also, I notice that with Crysis 3 the frame rate comes crashing down during the cut scenes. I tried very hard to skip them immediately, but a second or two of lousy cut scene frame rate made it into these numbers.
Despite the numbers, Crysis 3 feels smoother than Metro 2033 in both 2D and 3D on the tri-SLI 680s. With all of that said, let’s have a look at the “time to render” graph. What I did here is take the interesting data FRAPs provides which shows the time to render frame by frame. It literally gives you the exact number of milliseconds between frames. From this you can observe all sorts of interesting things like frame 3 to frame 4 taking longer than frame 200 to 201 and, if you analyze where these frames are during play, you could potentially figure out exactly where and why a card is struggling. Interesting data and I think the community is really just scratching the surface of looking at things this way and acting on it. The time to render data is from the same run through as the FPS data. 2D is up first:
The strange thing here are those big lag spikes. Why those occur is something I haven’t figured out yet, but it might be the sampling periods of the actual FRAPs benchmark tool. Either way these spikes do map to the frame rate minimums (obviously), so it does give a sense of where the game will hitch and where the player will have their experience disrupted. The graph also shows us that the majority of the time, the game is humming along within a pretty tight time-window frame to frame (20-30ms). For reference, a steady 60fps would require a constant rate of 16ms frames (1000ms in a second / 60 fps desired target frame rate). Now lets take a look at 3D:
Now this is interesting! Check out the massive reduction of lag spikes switching to 3D. Why? I don’t know. But this does seem to lend itself to my theory that either Crysis itself, or the NVidia drivers, are in some way tuned for 3D performance and that Crysis is able to realize benefits from this that possibly other games have not yet been able to. The (relatively) tight grouping here goes a long way towards explaining how Crysis manages to feel so smooth in 3D. I actually found that it didn’t really feel any less smooth than 2D. On Crysis 3 I am finding that 3D is, essentially, “free” and this is a good trend!
OK, so how did Titan do? First lets look at the frame rates:
In a word: WOW! Just look at the massive slam dunk Titan in tri-SLI effects over GTX680. We’re seeing averages hitting the GTX680 max. That is impressive and the impact of this performance increase is immediately noticeable in game. What’s even more shocking is how close 3D performance is to 2D! More evidence of 3D being free in Crysis and really an astounding accomplishment. Hats off again to the Crysis team for creating an engine that can manage averaging in excess of monitor refresh rate (60hz) in surround sound 3D with details cranked to the roof! Even the minimums don’t tell quite such a horrible story here and in game any pausing is only occurring on that brutal open warehouse firefight and is nowhere near as impactful as on the 680s (which can actually effect your accuracy at points). So how did the time to render hold up for the Titans?
Very interesting that those huge lag spikes are gone. It is definitely possible that those lag spikes were some sort of anomaly with that particular run. Unfortunately the 680s are now pulled so I can’t re-run that test, but I do have some older tests stashed away that I will analyze. As one might image having seen the minimum frame rate, the Titans never exceed 90ms single frame render time and the grouping is quite tight with many frames actually coming in well below 10ms. Just amazing. But how about 3D?
The peaks have increased just a bit (owing to that lower minimum frame rate), but the grouping remains tight and again, a big chunk of frames are completing sub 10ms. The graphs are actually very similar since, shockingly, the final performance numbers were so close!
What is the overall conclusion on Crysis 3?
- Crysis 3 is a real showcase app for next generation hardware without a doubt
- The GTX680s fare quite well and even 3D is nicely playable – a breath of fresh air from Metro
- Titan takes it to just another level. The performance difference over the already solid 680 times is noticeable and pushed up against monitor refresh
- Titan 3D performance is particularly amazing and Crysis 3, in 3D surround, at this level of performance is a sight to behold!
So is it worth $3000 that the upgrade cost? Well I have to say that Crysis 3 starts to make a strong argument in favor of yes 🙂 Of course that’s a bit insane, but hey, the money is already spent and who needs food when you have frames right?
Last Up – FarCry 3
I have a real soft spot for FarCry. I really appreciate their interesting “open world FPS” approach and I think it helps set the franchise apart from most of its competition. All of that rich environment, open world goodness, combined with the fast action of an FPS, makes for a pretty tough workload. Crysis 3 performance was surprisingly impressive, so how did FarCry 3 hold up? First, the settings. Here is what base video options were set to:
And here are the video quality settings:
Once again everything is maxed out except for AA/AF for the reasons outlined earlier. I find diminishing returns at this resolution (especially in 3D), and there isn’t enough horsepower on hand to justify them at the loss of either frame rate or quality. Even with Titan, so far this logic seems to still hold (at least at this nosebleed level of gaming intensity – lesser workloads would not require the tradeoff). For FarCry, the benchmark used was a FRAPs measurement of the first mission in the jungle after liberating the first radio tower. Starting with the end of the conversation with the NPC (at the tents), and following the path around the river to the farm where the task is to gather some plants. Along the way the left fork is taken to cut through the tall brush before rejoining the main road in order to test the impact of dense foliage.
Here is how the GTX680s look in 2D and 3D in terms of frame rate:
Looks like the FarCry engine isn’t quite up to Crysis levels, or perhaps the jungle setting is more intense than the Cell Factory level. Both games look fantastic. Subjectively, the performance is excellent in 2D, but feels a bit choppy in 3D. NVidia reports that the 3D Vision support is “fair” and there are some noticeably anomalies with depth and object rendering in the 3D plane. Unlike the flawless Crysis 3 experience, the Far Cry 3 3D experience needs some refinement. Mins are rough in both cases, but there is a likely misconfiguration snafu that caused this which I will discuss below. In any event, they aren’t really felt in 2D whereas in 3D they seem to have more impact. The time to render analysis should tell the tale there. First 2D:
Interesting results from the render time analysis for 2D. Looking at the graph the bulk of the frame rendering is occurring in less than 25ms (about 40fps). There are some huge spikes early on, but I believe these are likely the result of how FarCry reacts to starting the FRAPs benchmark. Turns out F11 is mapped to quick load, which is also the default setting to start benchmark testing in FRAPs. As a result, FRAPs starts benching as the game starts loading. After 10 seconds or so it does settle in however. In any event, the frame rate delivery isn’t very spikey overall which does explain the subjectively smooth experience. Now lets see how 3D held up:
Ignoring the initial spike, 3D is still much more spikey with lots of peaks and valleys and a lot of time spent in noticeably poor frame rates (above 50ms). This explains the subjective rough feeling of playing in 3D. I would say that Far Cry 3 and Metro 2033 feel fairly similar on the 3D side. Playable, but not exactly pleasant in the intense moments and not something you would do for the entire game. 2D surround, on the other hand, is fine. Is there room for improvement? Sure. Let’s see what Titan is able to do:
Once again the minimum data is likely corrupt as a result of the F11 mishap. The render times will provide a lot more insight. Looking at the averages, however, we see some really healthy gains moving from 680 to Titan. More importantly, the subjective impact is huge. 2D surround feels extremely smooth and any dropouts are completely unnoticeable while 3D actually becomes something you could consistently play. 3D on Titan feels ballpark, subjectively, with 2D on GTX680 and that is an impressive thing indeed. So how do the render times look? First up 2D:
No surprises here. Ignore the first few seconds as the game reloads and we see a really tight grouping. So tight in fact that something interesting jumps out which is the 3 spikes. What those represent, I am not sure. It might be FRAPs benchmark sampling or something else that was going on in the background on the test system. Either way, the spikes were not noticeable during play and that is really the main thing. Did 3D hold up as well? Let’s take a look:
Not as tight, of course, but still pretty smooth. Fewer dropouts than on the GTX680 and less spikey overall accounting for the smoothness. Once again a strong showing for Titan.
So what can we conclude from all of the testing? Well I think there are a few points that jump out:
- Titan scales well in tri-SLI – as well as GTX 680
- Gains over single card scale accordingly – roughly 20-50% gains single card vs single card translate to the triple card setup
- There is room for developers to differentiate even at this level – one look at the amazing Crysis performance proves that. C3 has become a performance champion in my opinion
- Even Titan could use some more horsepower when it comes to 3D surround at 5040×1050, but Titan makes even the brutal scenarios tested here playable
Is it worth it? If you already have thousands sunk into a premium gaming system and are chasing a specific experience like 3D Vision Surround then I would say Titan is absolutely worth it. Why? Because it can make things that are not playable become playable. At a near 100% increase in price card for card vs a 4GB GTX680 it is a tough pill to swallow, but it does deliver the numbers and that is what matters.
I guess the main point is that I am not a pro-reviewer who has these things on loan. I bought these cards and the register rang up a sobering $3000. After all is said and done, the 680s are on eBay and I have formalized my move to Titan. I think in the final analysis, that really says it all! Thanks for reading and expect continued testing of Titan and hopefully some game play movies!