3DMark Timespy Titan V Benchmarks

TimeSpy gave us more trouble than Firestrike or any of the games, and refused to launch the second graphics benchmark without bringing memory back down to a 150MHz offset, and core down to 175. Everything else was fine at 200MHz offset for each application.

TimeSpy measured our graphics score at 12308 points for the stock Titan V, placing it 22% ahead of the stock Titan Xp at 10092 points. Overclocking the MSI 1080 Ti Gaming X got us the closest to the Titan V, at 10798 points, marking the stock Titan V as 14% ahead of the 1080 Ti Gaming X overclocked card.

Comparing the framerates from each of the two graphics tests will help us understand where Volta is doing better. Volta holds approximately a 23% lead in GFX test 1, with about a 20-21% lead in GFX test 2. Graphics test 2 has about three times as much tessellation as graphics test 1, and relies more heavily on volume ray casting, which dynamically traces rays through the volume for each pixel. Graphics test 2 tends to be significantly more memory sensitive, and will crash faster under unstable VRAM clocks.

This gives us a starting point for where Pascal and Volta diverge.

3DMark Firestrike Benchmarks – NVidia Titan V

Looking at Firestrike, let’s start with Firestrike Ultra graphics scores for a baseline. This one shows less of a lead than Timespy, which uses newer, lower level programming techniques than Firestrike. With Firestrike Ultra, we’re actually seeing our Titan Xp overclocked card outperform the Titan V by about 5.3%. Overclocking the Titan V gets it to 8.3% ahead of the overclocked Titan Xp, which isn’t all that impressive, particularly considering how much better it did in Timespy. Versus an overclocked 1080 Ti Gaming X, the overclocked Titan V is about 16% ahead.

To understand why this behavior occurs, we can look at GFX 1 and GFX 2 scores. Graphics test 1 loads the GPU with polys and heavy tessellation, but doesn’t apply much of a compute load. GT2 increases compute workloads and can stress memory more, similar to Timespy GT2, but with less tessellation focus.

Looking at these numbers, we see somewhat considerable gains in the Titan V with GFX test 1, ranking at 48FPS AVG stock, as opposed to the Titan Xp’s 40FPS AVG stock. The Titan V maintains a significant 20% lead, but manages to fall behind in GT2: the V is at 26FPS versus 26FPS for the stock Titan Xp. Overclocking each gets both to 29FPS for GT2, despite a considerable 22% lead in GT1 for the Titan V, at 55 versus 45FPS.

Quick shout-out to the CrossFire Vega 64s, here – TimeSpy still likes multi-GPU testing, and it shows in this test.

Looking at this data, it appears that the Titan V has stronger potential in tessellation and geometry-heavy scenes, but might be becoming somewhat memory bound in GT2. We aren’t sure yet. The advantage in Timespy also suggests improved asynchronous compute performance for Volta over Pascal, which would be useful in newer, low-level APIs like DirectX 12 and Vulkan. We can build this hypothesis by exploring games which are known to leverage asynchronous and low-level programming techniques.

DOOM Vulkan GPU Benchmark with Titan V

Starting with DOOM at 4K and with asynchronous compute enabled, the stock Titan V places at 132FPS AVG. For average framerate alone, we’re about 41% ahead of the Titan Xp stock GPU. Overclocking the Titan Xp – actually, let’s give it some help and defer to the overclocked Hybrid Xp – gets it to 113FPS AVG, allowing the Titan V stock GPU to hold a 17% lead. The stock Titan Xp is largely choking on its cooling, for starters, but power limitations also restrict its performance. Either way, once we account for a Titan V overclock, we’re at 157FPS AVG versus 113FPS AVG, or back to a 39.5% lead over the Titan Xp. Scaling is pretty linear between stock and overclocked tests.

The Titan V does lag behind in one key area, though, and that’s its frametimes. We think part of this has to do with drivers or with a DOOM update; in fact, if you look at our old Titan Xp data from around May, you’ll notice that its low frametime performance was actually better, as was also the case with the older 1080 Ti data. Looking at modern data sets, the Titan V and Titan Xp are closer together in lows.

The big take-away here is, as indicated by Timespy, that the Volta card seems to have improved performance specifically in asynchronous compute titles. We can dig into that theory further with Sniper, and then reinforce it with D3D11 titles.

Sniper Elite 4 GPU Benchmark – NVidia Titan V

Sniper Elite 4 is our next title. We run this game at 4K with High settings, DirectX 12, and async compute enabled. Stock, the Titan V operates at 115FPS AVG, with lows at 97 and 91FPS for 1% and 0.1%, respectively. This places it about 27% ahead of the Titan Xp stock GPU, posting massive leads over the predecessor. This is after retesting the Titan Xp with the latest drivers, too, so what we’re seeing is a legitimate performance uplift in the Titan V, despite its clock deficit.

Before digging into this data more, let’s analyze it a bit: We’ve learned in the past that Sniper Elite tends to be surprisingly shader intensive. This is true when comparing the Vega 56 and 64 cards and the GTX 1080 and 1070 Ti cards. The game likes shaders, and that makes sense, as it’s built to asynchronously queue render jobs into those shaders. Having more available means more simultaneous in-flight render jobs. Sniper tends to be particularly sensitive to GPUs in general, but more to shaders than we see in other games.

Back to the chart, overclocking the Titan V to its +200 offsets gets us 41% ahead of our overclocked Titan Xp Hybrid card. This is a tremendous gain, and shows that the card becomes more constrained by its stock clocks in this particular game. The takeaway is that the extra shaders help, but they need to be fed with higher frequencies to really engage fully.

Ashes of the Singularity – Dx12 Benchmark with Titan V

Ashes of the Singularity is our final non-Dx11 title. This one posed a problem for us: We didn’t plan to ever test cards of this caliber, and so our standard 4K test with High graphics immediately demonstrated a CPU bottleneck:

As you can see here, all the cards basically equalize at around 97-99FPS AVG. We cannot draw conclusions here, as we’re now bumping against external limits.

For this reason, we reran a few tests with 4K and Crazy settings, further applying an 8x MSAA option to force load onto the GPU. We don’t have as many numbers for this, as it was unplanned, but you get the idea:

The Titan V operates at around 82FPS AVG when overclocked, a lead of 9.5% over its stock configuration. Against the Titan Xp, the stock Titan V has a lead of about 10.2%, from 74.85 to 67.9FPS AVG. The low 1% and 0.1% lows, representing frametime consistency, are also close by.

This number sounds a lot more down to earth, and demonstrates the extent to which game development impacts scoring of this class of card. We cannot extrapolate wide-reaching results from other titles across all titles. Although Ashes is DirectX 12, it does not leverage the Titan V to the same extent as some previous tests; either that, or it does not choke on the Titan Xp pipeline enough to demonstrate those deltas.

Hellblade 4K – Titan V vs. Titan Xp & 1080 Ti Benchmark

Moving to another chart of limited tests, we have Hellblade, which is our Unreal Engine representative. Hellblade has our Titan V stock card at 67FPS AVG, while our overclocked variant operates at 74FPS AVG (an improvement of 10.5% over stock). Versus the Titan Xp, we’re at 49FPS AVG stock and 54FPS overclocked, putting us at 37% improved when comparing only overclocked Titan V and Titan Xp numbers. Our improvement is 37% from stock to stock.

These gains are larger than we’d anticipated, considering that this is a DirectX 11 title and several key functions of the Titan V are unutilized in games. Let’s look at Ghost Recon for another one.

Ghost Recon: Wildlands – 4K Benchmark Titan V

At 4K and Very High settings, the Titan V stock GPU operates a 62FPS AVG, with lows at 54FPS and 50FPS. This is a stark difference from the last few games: The 1080 Ti GPUs have no trouble keeping up with the Titan V stock GPU, with the overclocked 1080 Ti SC2 somewhat proving that point by achieving parity. The $750 card shows the importance of GPUs targeted purely at gaming, particularly matched against its $3000, scientific-targeted counterpart. Against the Titan Xp, we’re still looking at a difference of 1-2FPS – nothing exciting. Overclocking the Titan V gets us a bit beyond the Titan Xp overclock, about 6% ahead of the overclocked Hybrid Xp, and demonstrates what’s lacking: Ghost Recon seems to want more clocks, and the additional shaders just aren’t helping. This is why it’s important to have a wide sweep of games, rather than just all games, and then talk about what drives those differences.

Ghost Recon is more old-school in its DirectX 11 implementation; if For Honor shows us similar results, we can start building a case about certain Dx11 titles and Titan performance.

For Honor 4K Benchmark – Titan V

Here’s For Honor, also made by Ubisoft. At 4K, the stock Titan V has us at 83FPS AVG, but with lows ranging somewhat sporadically between 29FPS 0.1% and 40FPS 0.1%, with 1% lows ranging from 57 to 74. This range is significantly wider than we see on other cards, and potentially indicates driver-level frame pacing issues or fundamental challenges with running a scientific card. The Titan V still manages to chart-top, but it’s only a couple FPS ahead of overclocked 1080 Ti cards. Versus the stock Titan Xp, the Titan V is about 14% ahead, though lows are somewhat difficult to compare, given their seeming randomness.

Destiny 2 Benchmark & Frametimes – Titan V

Moving on to Destiny 2 for another DirectX 11 title, starting with 4K/High, we get another scenario where the Titan V stock GPU only marginally outpaces the Titan Xp GPU. The difference is about 4.1%, with the Xp showing significantly stronger frametime consistency than the Titan V. We are not presently sure if this is a driver optimization layer problem or a fundamental behavior when deploying a scientific card in a gaming scenario. Overclocking the Titan V gets it to 125FPS AVG, providing a substantial 21% lead over the stock Titan V, but the card still struggles with frametime consistency.

Frametimes for 4K/High are shown above. We encountered more frequent spikes in our frame-to-frame intervals with the Titan V card, which coincides with similarly encountered performance in preceding benchmarks.

Performance is comparable to the Titan Xp and 1080 Ti when stock, minimally, and illustrates where those low-level API gains stop.

At 4K and Highest, we only have a few cards present, as the new Destiny drivers wiped our charts out. The Titan V tested at 88FPS AVG, with the overclock granting an 18% uplift. It would appear that Destiny cares more about clocks than other titles, though the Titan V also carries a 13.6% lead over the Titan Xp stock GPU when under the Highest settings.

Conclusion: What the Titan V Teaches Us

We’re entering territory of informed speculation. Please be aware that, from this point forward, we’re using our data to fuel conjecture on possible outcomes for Volta.

Purely observationally, based on the data we have presently collected, it would appear that the Titan V has two primary behaviors: (1) Applications which are built atop low-level APIs and asynchronous computational pipelines appear to process more efficiently on the Titan V; (2) the Titan V appears to host more cores than some of these applications (namely D3D11 titles) can meaningfully use, and that is demonstrated fully upon overclocking.

Given that overclocks in D3D11 applications produce performance uplift of ~20% (in some instances), it would appear that the high core count becomes more of a burden than a benefit. The GPU needs the faster clocks, and can’t access or leverage its high core count in a meaningful way. The result is that the Titan V begins to tie with the Titan Xp, and that the 1080 Ti closes-in on the Titan V. In lower-level API games, however, the Titan V pulls away by large margins – 27% to 40%, in some cases. The gains are big enough that we retested numerous times on numerous cards, but they remained. Our present analysis is that these applications are better able to spin-off multiple, simultaneous, in-flight render jobs across the high core count, whereas the tested Dx11 titles may function more synchronously.

As for the Titan V specifically, it can certainly be used for games -- but only in the context of, "I bought this thing for work, and sometimes I play games." If you're just gaming, clearly, this isn't the right purchase. Even for those users who have non-scientific uses for their scientific cards, the Titan V does appear to have some frame pacing problems that need to be worked out. We are not yet informed enough on the Volta architecture to root-cause these behaviors, and would suggest that it's either drivers or related specifically to the Titan V.

That’s what we think right now, anyway, and that may change. This is still early in Volta.

More soon.

Editorial, Testing: Steve Burke
Video: Andrew Coleman