How Much Is "A Lot"?
AMD officially launched their Radeon Fury X last week, and in a few more weeks, we should see the air-cooled version (though exact specs for the vanilla Fury are not yet clear). With our initial Fury X review, we didn't have time to test everything as thoroughly as we'd like, but we've updated the original benchmarks—no more factory-overclocked EVGA 980 Ti comparisons, and everything is running in the same standardized test bed:
The change in test beds and comparison to the EVGA 980 Ti didn't affect the review, as we had already factored in those items, but we were also able to play around with the Fury X a bit more and run some overclocking tests. We've seen reports of up to 10 percent GPU core overclocks on Fury X around the web, but our particular sample proved unstable even with a mere eight percent overclock, despite maintaining low core temperatures. We ended up with an even less impressive seven percent stable overclock, which, as you might imagine, means real-world performance is only 3–5 percent faster than stock.
There are two important limitations with the current state of Fury X overclocking that need to be addressed. First, there are no utilities (yet) that allow adjusting the core voltage. Bumping the voltage a bit is often the difference between almost stable and fully stable overclocks, not to mention given the liquid cooling and low temperatures, Fury X should still have room to go faster with additional voltage. The second is that there isn't currently a way to adjust HBM clocks. With 512GB/s of bandwidth it might not seem like a bottleneck, but considering Fury X has 45 percent more computational power than 390X but only 33 percent more memory bandwidth, that assumption may not be entirely correct. Interestingly, there are other posts circulating the Internet suggesting the HBM memory can be overclocked by as much as 20 percent, and that it actually makes a difference.
These two limitations mean that this is only a preview of what Fury X overclocking may offer, and we are really hoping to see a lot more in the future. AMD needs to deliver better performance in order to compete with Nvidia's 980 Ti, as it only ties or trails the competition (depending on benchmarks, resolutions, and settings) currently, and that's only the stock-clocked 980 Ti. Factory-overclocked 980 Ti cards have existed since launch, with EVGA boasting a 10 percent core overclock initially, and now the Classified model comes with a nearly 20 percent factory overclock, and that's only that beginning. Zotac just sent us a sample of their 980 Ti AMP! Extreme, which comes with a 25 percent factory overclock on the core, and an extra three percent GDDR5 overclock for good measure. But we already managed a 20 percent core overclock on the reference card, and 11 percent GDDR5 overclock as well, so it will be interesting to see how far we can push some of these other 980 Ti cards.
If you've wondered about the reasoning behind the initial review scores, overclocking is certainly a factor. Nvidia's Maxwell 2.0 cards (GTX 960/970/980 using GM204 and GTX 980 Ti/Titan X using GM200) have all proven very capable overclockers. For high-performance enthusiast cards, getting great out-of-box performance is expected, but being able to tack on another 15–25 percent through overclocking really pushes things over the top. Conversely, a product that offers competitive stock performance but only a small overclocking potential is typically less desirable, assuming all other factors are equal. With that said, we felt it would be enlightening to take another look at the Fury X vs. 980 Ti, this time with the best stable overclocks we could manage on each card.
As usual, we want to note that nothing is guaranteed with overclocking, other than perhaps factory overclocks, so these results are more of a baseline than something you will experience. Your choice of case, power supply, and case fans can all play a role, not to mention ambient temperature, drivers, and VBIOS. We're far more confident in our 980 Ti results as being representative of the GPU than we are of Fury X, and hopefully AMD will help unlock the full potential of the card in the coming months. For now, however, despite the similar price, overclocking in particular will heavily favor Nvidia.
Two Men Enter, One Man Leaves
Note: 97th percentile average used for Heaven minimums.
At stock clocks, the Fury X and 980 Ti are at least reasonably matched. AMD claims a couple of wins and two more ties at 4K, and on average we're looking at less than a 10 percent delta between the two GPUs—and an even smaller difference at higher resolutions. The average of our eight core games gives Nvidia a ten percent lead at 1080p, seven percent at 1440p, and only three percent at 4K. Minimum frame rates favor Nvidia a bit more—18, 14, and six percent at 1080p, 1440p, and 4K—likely thanks to the 6GB VRAM on the 980 Ti. And, of course, AMD is doing this with liquid cooling and lower temperatures, while still using a bit more power; they're running beta drivers as well, but that's how things currently stand.
Throw in overclocking and the comparison becomes far more lopsided. At present, the best we're able to manage with the Fury X is a seven percent overclock, while on the reference 980 Ti we're able to hit +225MHz on the core and +900MHz (7,910MHz effective) on the GDDR5. That represents a 23 percent core overclock and a 13 percent memory overclock. And if you're wondering, the Zotac 980 Ti AMP! Extreme goes even further; we were able to hit 1,300MHz on the core and 8,300MHz on the memory—stay tuned for the full review.
Looking at the overclocked results, it's a clean sweep for the 980 Ti right now. Even games that traditionally favor AMD (e.g., Hitman: Absolution) are ruled by the overclocked 980 Ti. The smallest lead we recorded for the 980 Ti OC is around five percent, in Hitman at 1440p and 4K. In games where the 980 Ti already held a lead, there can be a chasm between the two cards; Heaven, for instance, has a 28 percent lead at 4K and a 45 percent lead at 1080p. Ouch. On average, the 980 Ti OC leads the Fury X OC by 26 percent at 1080p, 21 percent at 1440p, and 15 percent at 4K in average fps; looking at minimum fps, right now 980 Ti OC leads by 39/32/24 percent over Fury X OC. About the only consolation prize is that the overclocked 980 Ti has peak power use for the system that's about 20W higher than the overclocked Fury X—426W vs. 408W—but that's partly due to the rest of the system working harder to provide the increased performance, and overclockers aren't usually ultra-concerned with needing a bit more juice to keep their system running at top speed.
To the Victor Goes the Spoils
And there you have it: AMD has plenty of work to do if they want Fury X to match the GTX 980 Ti consistently. Improving performance by 10–15 percent is certainly possible, and we've already seen Nvidia do that with driver updates since the launch of the Titan X. However, at least right now, AMD seems to be taking the Titan X approach to clock speeds with Fury X, meaning all of the cards are full reference models running reference clocks; at least the pricing is also consistent, though finding the Fury X in stock is another matter. The 980 Ti, on the other hand, is open for experimentation, and factory core overclocks of 10–25 percent are readily available, often at only a minor price premium. Even if you stick with the reference cards, it looks as though nearly all 980 Ti GPUs will happily run at 1,200MHz base, with boost clocks hitting 1,400MHz or more in some cases.
What we want to see from AMD in the coming weeks are two things. First, we want drivers that perform better than the current launch/beta 15.15 drivers. All indications are that AMD could net 10 percent or more performance with additional tuning, and that would be enough to match 980 Ti performance in most cases. Second, considering the "overbuilt" liquid cooling, we want to see voltage unlocks and HBM overclocking. If HBM can indeed be clocked 20 percent higher (600MHz base instead of 500MHz), and that results in another 10–15 percent improvement in performance, then we'd be looking at parity. That may not be the most exciting result if you want to argue about which graphics card is best, but stronger competition from Team Red can only help consumers in the long run.
But until/unless these improvements come to pass, the GTX 980 Ti is the best choice for gaming enthusiasts. It's fast, it's available, and it can overclock like a Swiss watchmaker hopped up on muesli.
Follow Jarred on Twitter.
