The launch of the current flagship among AMD's discrete video adapters – Radeon R9 Fury X – has gone more than ever. Fiji graphics processor and its opponent from NVIDIA – GM200 – became the pinnacle of the 28 nm process technology, which is used by both companies for the production of mass discrete GPUs from 2011 onwards. According to the area of the crystal (596 and 601 mm 2 respectively), these chips are already limited only to the photomask used on TSMC equipment, so the outcome of the confrontation completely determined the merits of the micro-architecture – Maxwell from NVIDIA and GCN 1.2 from AMD.
The top products, which were built on the basis of GM200 and Fiji-GeForce GTX 980 Ti and Radeon R9 Fury X, are very similar in their overall capabilities to each other, but differ significantly in certain aspects. NVIDIA sacrificed the speed of double precision calculations (FP64), but otherwise the GM200 turned out to be a well-balanced GPU, equally successful both in FP32 calculation tasks and in 3D rendering. Fiji, on the other hand, has unprecedented processing power of the array of shader ALUs, but in terms of performance of the front-end chip, which is engaged in triangle rasterization, it remained at the level of the previously presented GPU Tonga, which is now installed in medium-level accelerators Radeon R9 380 ). As a result, the Fury X is at least as low as the GTX 980 Ti in game rendering with Ultra HD resolution, but in 1920 × 1080 and 2560 × 1440 modes, when the geometry processing becomes a bottleneck, there can be no question of parity with the NVIDIA flagship.
In addition, the specifications of Radeon Fury X were affected by the limitations of HBM memory of the first generation, which did not allow placing on a single substrate with a GPU more than four chips totaling 4 GB. Although to date, only in exceptional cases, games can fully master such a volume, without first restraining the bandwidth of other components of the GPU, in the near future the situation may change. In addition, 4 GB of RAM already now does not leave a reserve for the Fury X tandem running in CrossFire (because the contents of the adapter's memory are duplicated). Finally, due to insufficient memory, AMD is unlikely to release a professional Fiji-based FirePro accelerator (at least if an intermediate upgrade is not available for HBM of the second generation). More details and performance of the Radeon R9 Fury we studied in the main review maps and in the addition of to it.
As a result, although the Fury X itself is a terrific product, the availability of the GeForce GTX 980 Ti for the same money deprived AMD of victory in the last round and limited the potential audience of Fury X to a few owners of 4K monitors. However, for those who care about this, Fury X has another quality – a compact and extremely quiet SVO (the problem with the wheezing pump in the early samples of the map is already considered solved).
However, this is not the end of the story with the Fiji graphics processor. In the confrontation of the flagship video adapters AMD had to exhaust all the potential inherent in the new GPU, and GM200 still had some reserve. At the same time, working on the younger model based on the partially blocked Fiji chip, AMD could freely maneuver in the price and performance range between the GeForce GTX 980 ($ 499) and the GTX 980 Ti ($ 649). As a result, Radeon R9 Fury "without X" received the recommended price of $ 549. Let's see what we are offered for this amount.
Fiji graphics processor in the Radeon R9 Fury lost 8 of 64 Compute Units. The remaining 56 blocks contain a total of 3584 shader ALUs (stream processors, in AMD terminology) and 224 texture modules (compared to 4,096 and 256, which have a full Fiji). In addition, AMD has lowered the maximum GPU frequency from 1050 to 1000 MHz.
The number of ROPs, the frequency (1000 MHz with 4096-bit bus) and the amount of RAM (4 GB) remained unchanged.
Thus, in the Radeon R9 theory, Fury can be 17% slower than Fury X in tasks that primarily load shader ALUs or texture modules. This is 100% corresponding to the load in the GP-GPU, but, as shown earlier by the Fury X tests, it is not typical for modern games. Users who do not disregard overclocking are likely to be able to play the difference between Fury and Fury X by raising the frequency to the level of the flagship or even slightly higher.
The memory capacity of 4 GB is quite appropriate for a video card competing with the GeForce GTX 980, which carries no more RAM. The only problem is energy consumption. Despite the difference in frequencies and configuration of the active GPU units, TDP Fury and Fury X are the same – 275 Watts. Fury X gets better ASIC with less leakage current. Another important factor is the cooling system. Regular FBO X on Fury X, in addition to being quietly working, is also needed because the leakage current is higher the higher the temperature of the crystal under load. Ordinary Fury does not have such a luxury. As a result, the GeForce GTX 980 with its 165W TDP clearly benefits from Fury's energy efficiency.
|Radeon R9 290X||Radeon R9 380||Radeon R9 390X||Radeon R9 Fury||Radeon R9 Fury X|
|The graphic processor|
|The code name||Hawaii XT||Antigua PRO||Grenada XT||Fiji PRO||Fiji XT|
|The number of transistors, million||6200||5000||6200||8900||8900|
|The technical process, nm||28||28||28||28||28|
|The clock speed, MHz: High State / Boost State||– / 1000||970 / –||– / 1050||– / 1000||– / 1050|
|The number of stream processors||2816||1792||2816||3584||4096|
|The number of texture blocks||176||112||176||224||256|
|The number of ROP||64||32||64||64||64|
|The width of the tire, bit||512||256||512||4096||4096|
|The type of the microcircuits||GDDR5 SDRAM||GDDR5 SDRAM||GDDR5 SDRAM||HBM||HBM|
|Clock speed, MHz (throughput, Mbps per line)||1250 (5000)||1425 (5700)||1500 (6000)||500 (1000)||500 (1000)|
|The input / output bus||PCI Express 3.0 x16||PCI Express 3.0 x16||PCI Express 3.0 x16||PCI Express 3.0 x16||PCI Express 3.0 x16|
|The conclusion of the image|
|Interfaces||VGA, DL DVI, HDMI 1.4a, DisplayPort 1.2||VGA, DL DVI, HDMI 1.4a, DisplayPort 1.2||VGA, DL DVI, HDMI 1.4a, DisplayPort 1.2||HDMI 1.4a, DisplayPort 1.2||HDMI 1.4a, DisplayPort 1.2|
|The recommended retail price, $ (USA, without taxes)||549 (at startup)||199||429||549||649|
Radeon R9 Fury X exists only in the form of a reference board, although in the future, probably devices of original design may appear. R9 Fury from the very beginning is produced by partner firms using original cooling systems on a reference or on its own printed circuit board. Only two AMD partners – ASUS and SAPPHIRE – participate in the launch of the card, and they are still the only manufacturers at the moment. Apparently, AMD is not yet able to saturate the market with enough Fiji chips.
Today we will deal with Radeon R9 Fury in the SAPPHIRE version. The video card is sold in two versions: one – with reference frequencies (it's in front of us), the other – with a GPU overclocking at 40 MHz. In other respects, both versions do not deviate from the specifications set by AMD. In Moscow online stores SAPPHIRE Tri-X R9 Fury meets at a price of 45-46 thousand rubles.
The device is delivered in a compact box. The scope of delivery is limited to the active DisplayPort – DVI adapter and HDMI cable.
The video card features a proprietary Tri-X cooling system, similar to the one we recently saw on Radeon R9 390X produced by SAPPHIRE. As the name suggests, the cooler includes three 90 mm diameter fans on double ball bearings.
The peculiarity of this version of Tri-X is that most of the radiator goes beyond the limits of a compact PCB. As a consequence, the flow of air from the extreme fan passes through the radiator. Stylistically, SAPPHIRE well beat the outstanding end of the cooling system due to the thick frame supporting the radiator and the rectangular casing.
The radiator of the cooler consists of two large segments, connected by seven heat pipes with a diameter of 6 to 10 mm. At the base of each segment is a polished copper sole: one is in contact with the GPU and memory chips, the other is pressed to the transistors of the power supply system through a thermal pad.
Note that GPU crystals and HBM chips, uncoated on a common silicon substrate, differ slightly in height. However, the base of the cooler in this place is absolutely flat: the space between the metal and the HBM chips is simply filled with a thicker layer of thermal paste, as can be seen from the print. The hardened thermal grease tightly held the cooler on the board even after we removed the fixing screws. Contrary to concerns, the replacement of the thermal paste more liquid and the re-installation of the cooler did not cause any difficulties.
The accelerator is based on a reference printed circuit board from Radeon R9 Fury X. We were afraid to shoot the FBO X with our Fury X, so we see this board for the first time. The power system includes four phases for the GPU and two for the HBM chips under the IR 3567B controller, which AMD has been using since Radeon R9 290X.
SAPPHIRE has saved the LED unit next to the card's power connectors, signaling the power consumption. Only instead of red diodes glow blue.
You can estimate the dimensions of ASIC Fiji in comparison with the NVIDIA GM200 and the more compact processors of the previous generation.
If you notice an error – select it with the mouse and press CTRL + ENTER.