Today, we are reviewing NVIDIA Titan RTX for overclocking, gaming, thermal and acoustic performance, seeing the first of two cards in the lab. We have a third card that arrives to be exchanged with one broken unit, working around the 1350MHz clock lock that we found, but it won't arrive after this review is displayed. Titan RTX is worth $ 2,500, higher than the RTX 2080 Ti about 2x, but only allows 4 additional streaming multiprocessors. With 4 more SMs and 256 more channels, there isn't much performance that can be obtained in the game scenario. The big advantage is in memory-bound applications, because Titan RTX has 24GB GDDR6, a noticeable increase of 11GB on the RTX 2080 Ti.
The use case example can be machine learning or deeper, or more traditional, 3D graphics rendering. Some of our internal Blender project files use so much VRAM that we have to render with slower CPUs (rather than CUDA acceleration), because we will run out of 11GB framebuffer too fast. The same applies to some of our Adobe Premiere video editing projects, where our graphic overlay becomes very complicated and high resolution so that it exceeds the 1080 Ti memory limit. We did not test one of these use cases today, and instead focused our efforts on the gaming market and were enthusiastic. We know that this is also a big market, and many people want to buy these cards just because "it's the best," or because "the most expensive = best." We will see how much difference you actually get. , with special interest in thermal performance in accordance with blower removal.
Finally, note that we are stuck at 1350MHz with one of our two samples, something we have done with NVIDIA for research. The company now has our broken card and has traded us with a functioning card. We bought the damaged Titan RTX, so it was a "real" retail sample. We just want to help NVIDIA solve the problem, and the company is now working with it.
The testing methodology has completely changed from our latest GPU review, which is possible for the GTX 1070 Ti series card. Especially, we have overhauled the host's test bench and have updated with new games. Our game choices are very careful: Time is limited, and after analyzing our previous testing methodology, we identify shortcomings where we end up wasting time testing too many games that don't provide data that is very different from the other titles we tested. To optimize our better available time and test "smarter" (than "more," which is one of our previous goals), we have chosen games based on the following criteria:
- Game Machine: Most games run on the same machine group. By choosing one game from each of the main engines (eg Unreal Engine), we can ensure that we represent a broad game that only uses default engine level optimization.
- API: We have selected the DirectX 11 and DirectX 12 API integration groups selected, because this is the most common today. We will include more Vulkan API testing because more games are sent with Vulkan
- Popularity: Is this something that people really play?
- Longevity: Regardless of popularity, how long can we reasonably expect a game to run without updates? Updating games can damage comparative data from previous tests, which affects our ability to cross-compare new and old data, because old data may no longer be compared to post-patch
Game graphics settings are defined in their respective charts.
We are also testing most games at all three popular resolutions – at least, we are for the high-end. This includes 4K, 1440p, and 1080p, which allows us to determine GPU scalability across multiple monitor types. More importantly, this allows us to start pinpointing the reason for performance uplift, rather than just saying there is performance uplift. If we know that performance boosts harder at 4K than 1080p, we might be able to call this indicative of a ROPs advantage, for instance. Understanding why performance behaves the way it does is critical for future expansion of our own knowledge, and thus prepares our content for smarter analysis in the future.
For the test bench proper, we are now using the following components:
GPU Test Bench (Sponsored by Corsair)
NVIDIA Titan RTX Overclocking
Overclock step appears first. You must know this information to follow our game benchmarks. The Titan RTX overclocking process follows the same steps as Ti 2080, and becomes choked by the cooler immediately under the out-of-box fan curve built into VBIOS. Under complete stock settings, we see an average frequency of around 1800MHz at TimeSpy Extreme, with a peak frequency of 1920MHz. Power is measured around 280W through GPUZ, even though we have separate measurements later. The fan seems to stick around 1515RPM to maintain a thermal target of 75 degrees Celsius.
The maximum power target is 114%, so we have limited space to improve performance. We end up at an average of 1830MHz, up to 30MHz, and that is without core frequency offsets. We also slowed down instantly at 88 degrees, causing the clock to bounce between 1780MHz and 1850MHz. A few steps later, we blow up the fan speed up to 3700RPM, or maximum speed, and see the performance increase significantly. This is an illustration of how limited stock coolers. Now, traditionally, some people have complained to us that we are too hard on NVIDIA blower coolers for this type of card, because their idea is you stack lots of blower cards into narrow boxes to be rendered, in this case more effective and remove heat from the system than axial coolant. This time, now that NVIDIA has replaced the cooler, the argument is no longer valid. This is only worse than that provided by any partner, but it doesn't even offer the only benefit a blower will offer, which is better performance in a scenario with multiple cards facing each other. We finally fell with a 200MHz offset, then found a clock core stability of 175MHz offset. The maximum frequency is 2085MHz when around 62 degrees Celsius, finding a resting point at the peak of 2040MHz below 66 degrees for the core. The average frequency landed at 2030MHz, with the maximum stable memory frequency offset at 1080MHz, which we think is one of the memory binders.
We can push this card harder with a bigger power offset. With a fan speed of 100%, the noise level cannot be tolerated for most users, but thermal performance is acceptable for overclocking. The problem is, just like the previous RTX cards, we become bound by overly protective power targets and voltage limits. We might try and VBIOS mod to see if we can go beyond this, or maybe just rip one of the shunt resistors from the board and do Buildzoid mod.
However, that is our overclock setting for testing. Overclocking is limited to this one, and that's entirely because of the power limit. GPUZ is reading a power draw at 330W when overclocked, but VRM can handle more than this, as shown by our PCB reference 2080 Ti analysis, and this is the same PCB as that.
Titan RTX vs. 2080 Ti, SLI, & 1080 Ti Benchmarks - Sniper Elite 4
Our game benchmarks will begin with Sniper Elite 4, which is one of the best games with a modern API. Using DirectX 12, Async Compute, and 4K / high settings, NVIDIA Titan RTX ends at 112FPS AVG, with lows at 90FPS and 87FPS 1% and low 0.1%. By comparison, the 1080 Tis on SLI does around 170FPS AVG, or the SLI 2080 Tis does 210FPS AVG. We also observed 2080 Ti at around 108FPS AVG, setting a difference of around 4FPS, or an average of 8.96 ms for Titan frametimes versus 9.2ms frametimes for 2080 Ti. We cannot think of many humans, if any, who can identify a 240-microsecond difference in a frame-to-frame interval.
Titan RTX overclocking to around 2040MHz core makes it an AVG 126FPS, surpassing the overclocked 2080 Ti around 3%. The next thing we need to test is NVLink Titan RTXs, which we will work on soon after this.
The following is the time frame between the Titan RTX stock and the 2080 Ti stock. As a reminder, time frames are the most accurate representation of frame-to-frame pacing, or time intervals from one frame to the next, and are the best way to objectively describe raw experiences without averages. Both cards have almost the same performance. Lower is better and more consistent is the best, but these cards are low in frame and consistent in the frame. NVIDIA has done well with frametime consistency for this generation, in large part, and none of these cards experienced a surge or frametime hang that deserves attention. It will be difficult to distinguish between the two. For reference, 16,677ms is 60FPS with 8 around 120FPS. What we care about here is that there is never more than a deviation of 8ms from the average value, so that the user is not likely to detect any fluidity or interference. Some people call this "smoothness."
Overclocking the Titan RTX introduces more frametime variants, as you will see in the line plot, but nothing is severe. We regularly suppress a 3md increase over a period of more than the baseline, from 7ms to 10m in some cases, but this is still relatively smooth and generally not too obvious to users.
Titan RTX Gaming Benchmark vs. 2080 Ti – F1 2018
F1 2018 gives us a glance at the EGO engine with DirectX 11, moving back to the wider API and moving away from the brilliant example of Dx12, Sniper Elite 4.
For F1 2018 at 4K and very high, we are clearly tied to a GPU with a high ceiling set by the CPU. In our CPU review, you will see 300 FPS with the right CPU, so this is a good way to really test GPU limits and avoid limiting the influence of the CPU. This is further illustrated by the SLI 2080 Tis at 169 AVPS AVG, clearly pulling away from the package. As for Titan RTX, which ends in the 110FPS AVG stock, place it right in front of AVF 99FPS 99GPS 2080 Ti FE, or AVG 105FPS AVG 2080 Ti XC Ultra. To the maximum, the distance is not more than 10%, and decreases to around 5% with the AIB 2080 Ti partner model. Overclocking the Titan makes it 118FPS AVG, right in front of the 2080 Ti FE overclocked on the AVG 114FPS.
NVIDIA Titan RTX Power Consumption
For testing power consumption, we measure between walls and systems for total system power consumption. The test platform is controlled 100%, including control of all minor rail voltages on the motherboard, number of fans, speed and type, keyboard and mouse, and any other parts of the system. Failure to control even some of the voltage on the motherboard will discard this reading and leave it inaccurate. We also log in from time to time, so you can see the right view that includes the peak and lowest, rather than the average number of tests.
This graph with Ashes of the Singularity: Escalation under the workload of 4K / Crazy, pushed the GPU to its limit. Running Titan RTX cards, our system's total power consumption peaked at 480W, outperforming the 2080 Ti Founders Edition system up to 24W. The only thing that beats the Titan RTX card is our highly modified Vega 56 tool, which uses a 250% power target to detonate power through the stock allowance, placing that total system pulling around the peak of 640W. Vega 56 is usually closer to the 350W range for total system power withdrawal.