Earlier this week I posted a review of the NVIDIA GeForce GTX Titan Z graphics card, a dual-GPU Kepler GK110 part that currently sells for $3000. If you missed that article you should read it first and catch up but the basic summary was that, for PC gamers, it's slower and twice the price of AMD's Radeon R9 295X2.
During that article though I mentioned that the Titan Z had more variable clock speeds than any other GeForce card I had tested. At the time I didn't go any further than that since the performance of the card already pointed out the deficit it had going up against the R9 295X2. However, several readers asked me to dive into overclocking with the Titan Z and with that came the need to show clock speed changes.
My overclocking was done through EVGA's PrecisionX software and we measured clock speeds with GPU-Z. The first step in overclocking an NVIDIA GPU is to simply move up the Power Target sliders and see what happens. This tells the card that it is allowed to consume more power than it would normally be allowed to, and then thanks to GPU Boost technology, the clock speed should scale up naturally.
Click to Enlarge
And that is exactly what happened. I ran through 30 minutes of looped testing with Metro: Last Light at stock settings, with the Power Target at 112%, with the Power Target at 120% (the maximum setting) and then again with the Power Target at 120% and the GPU clock offset set to +75 MHz.
That 75 MHz offset was the highest setting we could get to run stable on the Titan Z, which brings the Base clock up to 781 MHz and the Boost clock to 951 MHz. Though, as you'll see in our frequency graphs below the card was still reaching well above that.
Click to Enlarge
This graph shows clock rates of the GK110 GPUs on the Titan Z over the course of 25 minutes of looped Metro: Last Light gaming. The green line is the stock performance of the card without any changes to the power settings or clock speeds. While it starts out well enough, hitting clock rates of around 1000 MHz, it quickly dives and by 300 seconds of gaming we are often going at or under the 800 MHz mark. That pattern is consistent throughout the entire tested time and we have an average clock speed of 894 MHz.
Next up is the blue line, generated by simply moving the power target from 100% to 112%, giving the GPUs a little more thermal headroom to play with. The results are impressive, with a much more consistent clock speed. The yellow line, for the power target at 120%, is even better with a tighter band of clock rates and with a higher average clock.
Finally, the red line represents the 120% power target with a +75 MHz offset in PrecisionX. There we see a clock speed consistency matching the yellow line but offset up a bit, as we have been taught to expect with NVIDIA's recent GPUs.
Click to Enlarge
The result of all this data comes together in the bar graph here that lists the average clock rates over the entire 25 minute test runs. At stock settings, the Titan Z was able to hit 894 MHz, just over the "typical" boost clock advertised by NVIDIA of 876 MHz. That's good news for NVIDIA! Even though there is a lot more clock speed variance than I would like to see with the Titan Z, the clock speeds are within the expectations set by NVIDIA out the gate.
Bumping up that power target though will help out gamers that do invest in the Titan Z quite a bit. Just going to 112% results in an average clock speed of 993 MHz, a 100 MHz jump worth about 11% overall. When we push that power target up even further, and overclock the frequency offset a bit, we actually get an average clock rate of 1074 MHz, 20% faster than the stock settings. This does mean that our Titan Z is pulling more power and generating more noise (quite a bit more actually) with fan speeds going from around 2000 to 2700 RPM.
At both 2560×1440 and 3840×2160, in the Metro: Last Light benchmark we ran, the added performance of the Titan Z does put it at the same level of the Radeon R9 295X2. Of course, it goes without saying that we could also overclock the 295X2 a bit further to improve ITS performance, but this is an exercise in education.
Does it change my stance or recommendation for the Titan Z? Not really; I still think it is overpriced compared to the performance you get from AMD's offerings and from NVIDIA's own lower priced GTX cards. However, it does lead me to believe that the Titan Z could have been fixed and could have offered at least performance on par with the R9 295X2 had NVIDIA been willing to break PCIe power specs and increase noise.
UPDATE (6/13/14): Some of our readers seem to be pretty confused about things so I felt the need to post an update to the main story here. One commenter below mentioned that I was one of "many reviewers that pounded the R290X for the 'throttling issue' on reference coolers" and thinks I am going easy on NVIDIA with this story. However, there is one major difference that he seems to overlook: the NVIDIA results here are well within the rated specs.
When I published one of our stories looking at clock speed variance of the Hawaii GPU in the form of the R9 290X and R9 290, our results showed that clock speed of these cards were dropping well below the rated clock speed of 1000 MHz. Instead I saw clock speeds that reached as low as 747 MHz and stayed near the 800 MHz mark. The problem with that was in how AMD advertised and sold the cards, using only the phrase "up to 1.0 GHz" in its marketing. I recommended that AMD begin selling the cards with a rated base clock and a typical boost clock instead only labeling with the, at the time, totally incomplete "up to" rating. In fact, here is the exact quote from this story: "AMD needs to define a "base" clock and a "typical" clock that users can expect." Ta da.
The GeForce GTX Titan Z though, as we look at the results above, is rated and advertised with a base clock of 705 MHz and a boost clock of 876 MHz. The clock speed comparison graph at the top of the story shows the green line (the card at stock) never hitting that 705 MHz base clock while averaging 894 MHz. That average is ABOVE the rated boost clock of the card. So even though the GPU is changing between frequencies more often than I would like, the clock speeds are within the bounds set by NVIDIA. That was clearly NOT THE CASE when AMD launched the R9 290X and R9 290. If NVIDIA had sold the Titan Z with only the specification of "up to 1006 MHz" or something like then the same complaint would be made. But it is not.
The card isn't "throttling" at all, in fact, as someone specifies below. That term insinuates that it is going below a rated performance rating. It is acting in accordance with the GPU Boost technology that NVIDIA designed.
Some users seem concerned about temperature: the Titan Z will hit 80-83C in my testing, both stock and overclocked, and simply scales the fan speed to compensate accordingly. Yes, overclocked, the Titan Z gets quite a bit louder but I don't have sound level tests to show that. It's louder than the R9 295X2 for sure but definitely not as loud as the R9 290 in its original, reference state.
Finally, some of you seem concerned that I was restrticted by NVIDIA on what we could test and talk about on the Titan Z. Surprise, surprise, NVIDIA didn't send us this card to test at all! In fact, they were kind of miffed when I did the whole review and didn't get into showing CUDA benchmarks. So, there's that.