AMD's Fiji architecture has been one of the most hotly anticipated GPU launches for quite some time, representing a concerted push into the next-generation of high-end by leveraging an innovative new technology to try to shape the direction the gaming industry as a whole takes.
Fiji Pro Unveiled
Fiji is a major shift in strategy for AMD. They’ve moved into the arena of the large die once again, with all the risks that come with it. Potentially we could have high power consumption decreased efficiency and lower overall yields.
Fiji brings a lot of major and minor improvements that have helped to maximize TSMC's already aging 28nm process. The most notable and certainly the most talked about being the introduction of a new memory technology that could potentially provide a substantial advantage in memory bandwidth limited scenarios. And in reality it does have the potential to provide a healthy boost in gaming performance, provided it's properly developed for.
High Bandwidth Memory has been able to provide a significant leap in memory bandwidth compared to traditional memory technologies, though currently it doesn't have as much of a direct impact as we'd like to see. AMD had said during our interview with their chief memory designer at their Fiji release event that memory management itself could be the key to unlocking that potential, and certainly for cards with less than 4GB of VRAM.
The Fury is perhaps the most interesting product launch, being an only slightly cut-down version of full Fiji, but at the same power target but with air cooling instead. The Fury has 3,584 stream processors running at 1000MHz compared to the Fury X running at 1050MHz with 4096 stream processors. The Fury also has 224 texture units as opposed to 256 for the full fat Fury X, though they share the same number of ROP's. The memory subsystem is the same, using four modules of 4-Hi HBM generation 1 that gives a 4096-bit wide bus capable of 512GB/s of bandwidth.
|WCCFTech||Radeon R9 Fury||Radeon R9 Fury X (Water Cooled)||Radeon R9 Nano||Radeon R9 290X|
|GPU||Fiji Pro||Fiji XT||Fiji XT||Hawaii XT|
|GCN Compute Units||56||64||64||44|
|Render Output Units||64||64||64||64|
|Texture Mapping Units||224||256||256||176|
|GPU Frequency||1000 MHz||1050Mhz||TBA||1000Mhz|
|Memory||4GB HBM||4GB HBM||4GB HBM||4GB GDDR5|
|Effective Memory Speed||1Gbps||1Gbps||1Gbps||5Gbps|
|Cooling||Sapphire Tri-X, 7 Heatpipes, 3 90mm Aerofoil Fans||Liquid, 120mm Radiator||Air, Single Fan||Air, Single Blower Fan|
|Performance (SPFP)||7.2 TFLOPS||8.6 TFLOPS||TBA||5.6 TFLOPS|
|Power Connectors||Dual 8-Pin||Dual 8-Pin||8-Pin||6+8 Pin|
Most would probably look at the 4GB of memory and scoff. And that's a valid concern, in the days where many modern games have very large assets that fill up that room quite quickly, not to mention the graphical effects that also take VRAM to be applied properly.
Again, how those assets are called into memory and managed could be a more viable long-term solution than just increasing memory capacity of graphics cards and letting things run wild. Certainly it could be better for our wallets. HBM isn't cheap, and neither really is GDDR5. AMD is also currently restricted in how much on board memory that can be placed there by the virtue of how 4-Hi HBM works. It's stacked as high as it can and has the highest capacity available currently in each module, which allows it only 4GB max. HBM 2 will bring higher capacity and even larger bandwidth, though that isn't quite ready for production yet.
But the question of whether memory capacity is ever enough will remain forever at the forefront of our minds, regardless of the different methods used to lessen its impact. After-all, moar memory automatically means more efficiency, right?
The Fury is placed at a price-point that doesn't necessarily directly compete with anything. At a base price of $549, it's less than a 980 Ti yet more than a 980. The price performance ratio should theoretically put it beyond the 980 and even perhaps trailing the 980 Ti, though not by very much. As we also told you, the Fury consists of cut-down Fury X dies, meaning that it's entirely possible to enable those disabled compute units using a tool, though your results may vary. But you might have a hidden Fury X inside, which increases the value immensely.
The Sapphire Tri-X Radeon R9 Fury
The Sapphire Radeon R9 Fury continues the use of the Tri-X cooler but with improved fans and a different, darker color scheme. They're also selling two variants of it, one with a mild overclock and one at stock clocks. It's unfortunate that even with Fiji's much more efficient design and the Fury X's very competent cooler that it doesn't overclock very well.
The Sapphire uses the stock PCB straight from AMD, meaning it's 7.5" in length and has a six phase power design that can handle well over 300W total in total. Mounted to the PCB is a massive copper and aluminium cooler that has no less than seven heatpipes and three 90mm fans that Sapphire is dubbing their 'Aerofoil' fans. These are capable of moving vast amounts of air with very low noise. And they also have a zero fan idle so that the card is passively cooled if not under load.
With the Tri-X cooler, the entire card measures in at a quite long 12", extending 4.5" past the reference PCB used. The included backplate adds some much needed stability for such a large and substantial card. It's heavy, and certainly without such support the cooler could rip right off! Well, not quite that simply, but it's weighty alright.
Thankfully Sapphire has elected to stick with the stock dual BIOS design, with the same switch on the reference PCB. The stock selected bios allows for a maximum temperature of 75C and a maximum board power of 300W, but interestingly, it seems that the second BIOS ups that to 80C and 350W respectively, just in case you want to use LN2 or try your hand at overclocking. This would allow a lot more options for modders should they wish to experiment. Messing up a BIOS flash is easy to fix, just flip the switch to the other position and boot right back up. This is especially handy for those wanting to try their hands at unlocking the other CU's to see if they are truly just inactive or actually defective.
You'll find a generously sized vent on the end with three DisplayPorts and one HDMI 1.4a port. Power is provided via two 8-Pin PCIE ports on top, with the activity LED's present on the Fury X are also here, though in a magnificent blue.
To test the Fury properly, 4K is an obvious necessity. This is being billed as a 4K GPU, capable of providing a 4K gaming experience, so to that end the testing rig has been updated to include a Dell P2715Q.
|CPU||Intel Core i5-6600K|
|Motherboard||ASRock Z170 Extreme 4|
|Power Supply||EVGA SuperNOVA 1300 G2|
|HDD||SanDisk Extreme II 120GB|
|Storage Disk||Seagate 2TB|
|Memory||16GB Crucial Ballistix DDR4 2400|
|Video Cards||AMD R9 Fury, AMD R9 Fury Nano, GeForce GTX Titan X|
|Operating System||Window 10 64-Bit|
For all the tests MSAA was set to X2 to even the playing field. Battlefield 4 consisted of a play through of a 64 player server on the Siege of Shanghai level. Crysis 3’s benchmark was done by playing through the first level. Fraps was used to capture the framerates of Battlefield 4 and Crysis 3. The rest use the internal benchmarking services that were available.
First let's take a look at what the Fury is capable of when we use a more modest resolution of 1440P, still a common resolution among gaming enthusiasts with high quality panels out there and now those with much higher refresh rates out in the wild.
Of course we'll begin with the ubiquitous 3dmark Fire Strike test, with all three modes tested. This is a synthetic benchmark whose results are not applicable to real-world situations, but of course I know that a lot of people like to compare these scores, so they're included here.
And of course, let's take a look at the API overhead test featured in 3dmark, a useful tool to see how fast a GPU can process draw calls. This data is of no use on its own, but can help us understand how efficient a particular GPU can be with arbitrary graphical data.
The Frostbite engine has come a long way in the many years of its development. Even though it was released in 2013, it provides great insight into the capabilities of the modern GPU. DICE has done a great job making it look fantastic while generally running well on any number of GPU's It's still quite the challenge for even the most potent, however.
Though this was a multiplayer run, the Fury does well here, with slightly more frames than the 980 but not quite as many as the 980 Ti. It certainly does give the big hitters a run for the money, considering the price difference.
Crysis 3 is still one of the best looking games around. It's a challenge for any video card and almost the standard for which we jokingly ask if something is powerful enough. The CryEngine is a very advanced engine and brings most GPU's to it's knees when everything is turned on, such as here.
We see the Fury edge slightly ahead of the GTX 980 with a 10.7% lead.
Dragon Age: Inquisition
Released in in 2014, Dragon Age: Inquisition makes use of the Frostbite 3 engine in a slightly different manner and perspective. It also represents one of the better looking RPG's out there. Being Frostbite based, it should prove challenging to most GPU's.
Again it's really no surprise that the Fury vanilla is doing well against the 980, coming out on top again, if only slightly. A 6.4% increase over the 980 isn't insignificant by any means.
Shadow of Mordor
Shadow of Mordor is the wild card here, and I didn’t really know what to expect from it. The modified LithTech engine is an almost unknown. It has the ability to provide some nice looking visuals in a visceral world within Lord of the Rings, with great lighting and particle effects.
The Fury beats out the 980 by a high margin in this test, coming close to the 980 Ti even. The Fury is no slouch in this game, even though the LithTech engine isn't quite as highly detailed as some other engines. It still taxes the GPU well, and the Fury does fantastic here.
Civilization: Beyond Earth
As a 4K turn-based strategy game, one wouldn't necessarily think that this is a tremendously graphical intensive game, but it is. It makes heavy use of shader effects to make the map look and feel alive. That, and having the entire gigantic map filled with units does take its toll on the GPU.
All of those textures and the increased use of shader units means that the memory will be close to full for the majority of the test. The Fury, however, seems to do quite well, besting the 980 once again by just enough while being cheaper.
GTA V is a gorgeously rendered game, and very memory intensive at that. It even has a built in memory usage meter so you can see precisely what effects use the most of your precious graphics VRAM. That being said, there are a tremendous amount of options available in game to mess with, and doesn't have any pre-defined graphical levels. High Quality here means everything is turned up to it's highest, we have the softest shadow setting and MSAA is set to 2X on reflections and in the game itself.
The Fury is very much in line with our expectations here. even though having less memory, it takes on the lead among its similarly priced rivals. It's behind the 980 Ti by only ~8 frames, which is quite close indeed.
The Witcher 3: Wild Hunt
This is perhaps one of the most demanding games in our test arsenal. It makes use of some very high resolution textures and nearly every effect under the sun that you could possible think of to provide a very emotional and engrossing experience. It's a fantastic looking game and could very well be the new standard for GPU performance. This was benchmarked using version 1.5.
It seems to be a very close race for every card here, likely being CPU bound. But again, the Fury doesn't necessarily disappoint, and certainly provides a great experience.
And of course, below is a compilation of games to show off how the Fury does in games that are either older, or perhaps not quite the most popular release at the moment. All of the games were run at 1440P with the highest settings possible.
Now let's take a look at just how well it can perform at 4K considering it's seemingly limited by it's RAM amount. Does the HBM make a difference, and can it be faster than some of it's rivals?
Dragon Age: Inquisition
Shadow of Mordor
Civilization: Beyond Earth
The Witcher 3: Wild Hunt
4K gaming is still not necessarily a reality for many people, though the price of 4K monitors is certainly falling, even if they aren't necessarily of the best quality. This means that 4K gaming is accessible, but maybe not something that can be done on a budget.
The R9 Fury, even with 'only' 4GB of VRAM is a very capable card with 4K resolutions. It's consistently faster than the 980 and provides a very playable experience in all games except for The Witcher 3: Wild Hunt. A constant 30FPS, while a 'cinematic' experience, isn't necessarily the best for an action RPG. But I digress, the Fury makes an outstanding showing with 4K content. The best part is that it can provide that experience at a decent price as well. HBM, while not a magic bullet by any means, seems to be an effective solution to providing a playable experience at 4K.
Let's take a look at some compute benchmarks and see how that single precision throughput translates into the real world. I’ve decided that Einstein@home as it's updated frequently and uses CUDA and OpenCL very efficiently. I’ve also included the benches form a novel little benchmark known as ViennaBench. ViennaCL is a a linear algebra library that runs in OpenCL and even CUDA. The creator, Karl Rupp, has even added a nifty benchmark that measures performance in a number of different mathematical tests.
Einstein@Home conducts some very complicated math to determine whether or not a pulsar is located in the area being surveyed. It uses single precision math, and it's surprising that the Fury is the fastest of these cards, despite theoretically being slower.
The Tri-X cooler that Sapphire uses is quite robust and at the very least looks as if it's able to dissipate quite a bit of heat. For the load temperature, Einstein@Home was used to load the GPU with complex max at 100%. After two hours the temperature was recorded using HWiNFO64.The massive size is not just for show, however, and the design is actually quite efficient when paired with Fiji.
It's more than cool, and certainly kept below the temperature target as prescribed in the BIOS itself. It's not quite at the level of the Fury X, but it's better than the stock cooler on the 980 and 980 Ti, and certainly better than the old stock cooler used on the 290X.
Noise levels were measured using a Triplett Mini Sound Level Meter pointed at an open case 100cm away. This measures entire system noise because it would be quite the challenge to have everything else passively cooled just for this type of test. One phenomenon you'll notice is that Sapphire's Fury is the lowest at idle. That's due to the zero fan speed at idle that it's able to accomplish. The fans aren't moving at all, so that's simply a measure of the entire system noise, because it's literally making none.
Even better are the load noise numbers. It's simply very difficult to hear over the rest of the system and doesn't seem to be much of a nuisance from a frequency spectrum point of view either.
Last but not least, we have some power consumption numbers for you. Power consumption was measured while the entire system was at 100% load with a Belkin , so this is as much a measure of system load as it is if the GPU itself. Again, we use Einstein@Home as the program of choice, because it contributes to science as much as it contributes to my power bill.
Slightly higher than that of it's competition, the idle power consumption is nonetheless still quite low, and certainly lower than Hawaii. The zero fan idle probably also plays a role here as the fans do indeed take at least some power even at a low idle RPM.
The internal revisions and improvements seem to be paying off when in a non-power virus situation. The power consumption is much improved over Hawaii in regards to the performance metrics that the Fury is capable of. The gap between big Maxwell is slowly closing, though not quite there as of yet. It's a good effort, certainly, and is lower than Hawaii by a substantial amount.
As a 4K gaming card the Fury is able to provide a very decent frame rate in most modern games. With everyone turned on it might not be able to keep up with big Maxwell (which itself isn't exactly that fast), nor does it provide superior FPS, but one could perhaps turn-off MSAA and even turn down some effects to provide a very playable and thus enjoyable experience.
In terms of it's competition, the R9 Fury is faster than the less expensive GTX 980 ($70 in my case) in almost all situations, though it doesn't quite match the slightly more expensive 980 Ti. The 980 might be the better buy depending on your budget. It allows for better overclocking at the moment and uses far less power, which might be a real concern. At 4K, however, that gap lessens even with the 980 Ti, so the Fury could very well be considered in that situation. 1440P is good performance, and is between 6-18% better depending on the application, however.
All in all, it's a great performing GPU with a lot of potential. It does well at all resolutions and gives even the Titan X a run for its money in certain scenarios. If only it were a better overclocker. It's certainly recommended for those that are looking for a solution to 4K gaming. AMD is making a great showing with Fiji, despite the naysayers, it's a good card that can hold its own in many different situations. It's cheaper than the 980 Ti, which makes it a curiously good buy.