Its been a long time since AMD last updated their entire product stack and today is the day when we finally get to pull back the curtain and see what’s been happening behind it. The product launch today encompasses the full lineup of the Radeon R-300 family, which includes the Radeon R9 Fury and Fury X, AMD’s two flagship cards made ready to do battle with NVIDIA at the highest performance levels. While you may see articles like these and others on the net, including reviews for the Radeon R-300 series, the launch for Fury, and the review embargo for those cards, lifts a little later this month. What you’ll be seeing today is a preview of the Fiji architecture and performance levels, as well as details of where the rest of the cards sit in relation to the Radeon R-200 series. Lets get into it!
UPDATE: Math is hard. I’ve fixed up the calculations for the increases in hardware over Hawaii’s specifications.
Yes, its true – AMD’s top card is called the Radeon R9 Fury X. These are two separate products that feature the Fiji GPU – the other is the Radeon R9 Nano – and both have their pros and cons to their selection. Fury X is the one we’ve been seeing so many leaks of online, with the attached water cooler and the new shroud holding the water block and pump inside. It is reminescent of the Radoen R9 295X2, which came standard with a hybrid water cooler. It may be tame under the cooler and within its limits, but the Fury as you see it today isn’t its final form. I can’t say too much without going into specific performance numbers, but I will say this: If Fury X in standard mode (with frame rate limiting) is Gohan in Super Saiyan form, then Fury X unleashed (without frame rate control) is Super Saiyan 2. We’ll see more on this in the Fury X reviews going online in the coming weeks.
We’ll begin with a look at the Fiji architecture. It is a massive processor, measuring 596 millimeters squared for the chip itself. Because Fiji implements High-Bandwidth Memory (HBM) as well, which AMD provides with the GPU to its vendors as a complete package, the entire chip, with the memory and the silicon interposer to connect everything up, measures 1011mm². It is pretty big, almost as big as NVIDIA’s GeForce GTX Titan X, which slides in at 601mm². Fiji’s current implementation of HBM tops out at 4GB of memory, but it isn’t of the GDDR5 variety. In fact, there’s almost no other memory technology it can be compared to, because it is so very different from what everyone currently has in their system.
That whole interposer assembly, though, saves an incredible amount of space. From the mass of chips found around most common GPUs to date, to the small size of the Fiji GPU, there’s almost a 3x reduction in the footprint required on the circuit board. That does some interesting things for GPU design – one, you don’t need to make extremely long boards to incorporate more memory on to the board.
Two, you can now fit that GPU into form factors hitherto unthinkable. Like, for example, a six-inch-long GPU that fits into miniITX systems, or a seven-inch long water-cooled unit that fits nearly everywhere. The goal of HBM version 1 is to address primary concerns about memory bandwidth but ultimately it also solves the issue of graphics cards getting ever longer and larger to incorporate bigger GPUs. That’s a relief.
Delving under the hood to see what makes it tick, we discover that Fiji is more similar to Hawaii than anything else, a GPU found in the Radeon R9 290 and R9 290X. Fiji scales things up more, including a total of 4096 stream processors segmented into 64 Compute Cores, as AMD calls them. That leaves us with 64 stream processors per Compute Core, a figure unchanged from the Hawaii family. The L2 cache also doubles in size to 2MB, which AMD expects will help with Fiji only packing in 4GB of VRAM. A larger cache helps with prefetching data, and AMD might be working on some new caching mechanism to keep the beast fed.
Much of the architecture’s back-end is the same as Hawaii, including eight ACE units, six Eyefinity display controllers, VCE, UVD, TrueAudio and the DMA and XDMA engines, the latter which is used for Crossfire without the old bridge, which is an important part in AMD’s HSA hardware-accelerated plan to have VRAM shared between the graphics cards in a single pool as well as cache coherency.
But, oddly, while Fiji is a big chip, it almost looks as if the graph has just been extended downwards, with only four render back-ends per shader engine. While AMD has done a lot of work on optimising this version of GCN for efficiency and power, it is more or less the same as Hawaii, with some lessons learned from Tonga. So, that’s a significant sign of how things are going to play out – it’ll be a strong player in the market and it might even stomp over the competition without breaking a sweat, but there will be power consumption to consider when you configure your system to run a Fiji-based GPU.
Spec-wise, things are very different thanks to the inclusion of HBM. Made on the 28nm TSMC process, Fiji has 4096 stream processors, 256 texture units, 64 ROPs, a 4096-bit memory bus (1024-bit per HBM stack) and memory bandwidth reaching in the 512GB/s range. Compared to Hawaii, it sees nearly 50% more stream processors, nearly 50% more texture units, and a 60% increase in memory bandwidth. The ROP count stays the same at 64, which means that AMD considers the Radeon R9 Fury X to be suitable for running games at UltraHD 4K resolution.
In the notes handed out to myself and other reviewers, though, AMD is careful not to compare the raw specifications to Hawaii in any percentages. In keeping with AMD’s current movements in the APU segment, R9 Fury X has a configurable power draw, which I’ll detail a bit later.
The rationale for going with a water-cooled stock cooler was pretty reasonable – it keeps the R9 Fury X cool, at around 50° Celsius at a typical gaming load. The cooler itself, together with the radiator, has a thermal capacity of up to 500 Watts, which is almost the same as the R9 295X2. The card is powered by hooking up two 8-pin PCI PEG power connectors, each of which has support for up to 150W of power draw.
The typical board power of the R9 Fury X, then, is higher than I’d initially expected, but lower than I thought it would be. The power draw limit for a typical workload is 275 Watts, while the card supports a 375 Watts power draw limit electrically. That means that while Fiji at stock speeds will be rather easy to tame and keep fed and cool, but overclocked, it could be a monster. There’s a dual BIOS switch that was also present on Hawaii GPUs, which allows you to switch into a dedicated BIOS mode for LN2 overclocking. It can also be used in emergencies as a recovery BIOS so that you can restore the default one, if you ever flash it incorrectly.
On the back of the R9 Fury X, AMD’s done some pretty neat things for enthusiasts. There’s a nine-LED readout on top of the power ports, which give you a status of how much power your GPU is using and how close it is to the TDP limits. Eight of the LEDs are dedicated to the power level readings, while the single green LED is only lit when the card is in ZeroCore mode. ZeroCore is a mode activated when using Radeon GPUs in a Crossfire configuration, where the second, third, and fourth cards go into a idle mode while the primary card throttles clock speeds down as low as possible to conserve energy and heat output. I’m not sure why the power LEDs have a blue colour as an option, that just seems weird to me, unless AMD’s trying to get some Intel love going on here.
Also new is the Radeon logo, which now lights up when the R9 Fury X is out of idle mode and in operation. It is pretty simple, and right now there don’t seem to be any options in the Catalyst control center for setting breathing modes or turning it off entirely, but I’m sure those will be available later when the cards are launched.
Port-wise, the R9 Fury X comes with three Displayport 1.2a connectors and a single HDMI 1.4a connector at the bottom of the card’s port output. Not having HDMI 2.0 on-board is an interesting choice especially considering that the 1.4a standard is limited to running UltraHD 4K monitors at 30Hz refresh rates, while the 2.0 specification update supports 60Hz refresh rates. AMD says that this is because they expect more people to use the Fury X with multiple UltraHD monitors as well as in Eyefinity setups with up to six displays.
You can make use of all the ports, though. All four can be used at the same time, while the card tops out at six displays through the use of Displayport MST hubs, which split one port into two for larger multi-monitor arrays. Some MST hubs can extend that into three displays on one port, but those hubs also drop the maximum resolutions down to just 1080p at 60Hz, which might not be as useful.
The entire water-cooling system can be removed and repaired or replaced if it is ever needed, and in the exploded view the copper block and heatpipes can be seen sandwiched underneath the built-in pump. The hoses do seem to have a proprietary connection on them, though, so there’s no way of ever replacing the tubing with your own custom setup – the only way to tinker with this in any way is to replace it completely with your own water block and tubes. Bummer.
Although performance expectations are anyone’s guess at this point, AMD did share this single slide showing the average framerate of Far Cry 4 running on a single UltraHD 4K monitor at Ultra settings. Without context, that result is still impressive, proving AMD’s point that Fiji is expected to be their premier performance-class GPU for 4K gaming. The minimums are also interesting, because at 43fps that’s an ever-so-slightly better result than the Geforce GTX 980, reviewed here by Techpowerup. Radeon R9 Fury X’s main competition, then, is NVIDIA’s flagship enthusiast-class GPU. That’s pretty impressive.
The Fury X isn’t the only way you can enjoy having a Fiji GPU in your system, though – there’s also the Radeon R9 Nano, a Fury X without the water cooling solution and an even shorter length, allowing it to squeeze inside smaller chassis. The card has the same port output as Fury X and is also a dual-slot cooler, simplifying things very nicely.
The R9 Nano also has a reworked power profile. With only a single 8-pin PCI PEG power connector available on the board, it has a thermal power limit of just 175 Watts, comparable to its competitors in the ITX space like the Radeon R9 285 and the Geforce GTX 960. AMD claims it has up to twice the performance density and performance per watt of its competitors in that space and with similar dimensions. I’m not sure how that works out in terms of how the card’s performance is affected by these changes, but AMD says it should be the same, or similar, to the R9 Fury X.
The rest of the Radeon R-300 family
While the R9 Fury X and R9 Nano are new GPUs in every sense of the word, the rest of the lineup is a mixture of GPU families of varying ages. Its not a straight rebrand, mind you – all of these graphics cards have had their performance tweaked to suit their price points and competition, and all of them have some new tricks up their sleeve. However, upon asking AMD very pointed questions about which families were in use, the only question they were able to answer was that out of the GPUs launching today, only one of them does not support the new XDMA Crossfire (and therefore FreeSync and TrueAudio) technology – the Radeon R7 370.
Which, unfortunately, means that it is based on the Pitcairn GPU family, and is therefore GCN version 1.0. The R7 360 is likely an updated Bonaire-based GPU, while the R9 380 is Tonga and the R9 390 and R9 390X are Hawaii reskins, with the codename Grenada. If AMD has made updates internally to these graphics cards, they haven’t been forthcoming about them, so it is very likely that the lineup today is mostly composed of rebrands.
That isn’t a bad thing, though. Rebranding older GPUs with an updated power profile and a few new software tricks thanks to reworked drivers means that they’ll be coming to you at a lower price point. The R9 390 and R9 390X seem to be the same Hawaii cards, but performance has been improved to allow them to compete with the Geforce GTX 970 and GTX 980.
The Radeon R9 380’s main competition is the Geforce GTX 960, while the R7 370 competes against the Geforce GTX 750 Ti and the R7 360 squares things off against the GTX 750. All of these cards are highly competitive in their own right and AMD’s responses to them are suitable, given the performance on tap and the power profiles as well. The increase in the standard amounts of memory for the R9 380 and the R7 370 are also welcome. It is likely that a future update will see the R9 380X being based on Tonga, but with the 384-bit memory bus and more hardware all-round.
Speaking of software tricks, every GPU launching today in the Radeon R-300 lineup is granted the gift of frame rate control, a setting in the new version of the Catalyst drivers that allows one to set a target frame rate based on the hardware you have on hand. While this is primarily advertised as a power-saving feature in mobile GPUs, frame rate targeting control (FRTC) also allows you to tweak the peak levels of performance so that you don’t have to deal with too much tearing on your monitor with V-Sync disabled.
There’s a very large chance that FRTC isn’t limited to the Radeon R-300 series alone, however. AMD should be releasing the Catalyst 15.5 drivers later today, so you might want to grab those and look for the entry for Frame Rate Control in the settings menu. If it works on these GPUs, it should work on all Radeon HD7000 and R-200 GPUs as well.
That’s a wrap for AMD’s new update! The Radeon R-300 GPU review embargoes go up right this minute, and you should see them appearing on the internet very soon. The Radeon R9 Fury and R9 Nano review embargoes go up later this month and we’ll be able to see then what surprises these cards may bring. Keep your sights on NAG Online, as we’ll have several reviews of the new cards launching today in the coming weeks.