When two models from the Volcanic Island series came out last month, we were a bit disappointed with the fact that the model with a brand new chip and code name Hawaii was delayed for the end of October. In the end, they kept their promise, and we’ve received Radeon R9 290X. With this newest addition, AMD went back to the top when it comes to gaming performances and that’s quite an accomplishment. The newest chip is bigger and more complex than ever, and it’s the best GPU that AMD ever made.
Record number of transistors
AMD had to work hard in order to make something that would put them on par with NVIDIA’s server chip named GK110. NVIDIA had this GPU for a long time, but it wasn’t cost-effective to immediately switch from K20X server card to models intended for commercial use. Therefore, it comes as no surprise that prices are extremely high – about 600 euros for GTX 780 and completely unreasonable 1000 euros for GeForce GTX TITAN. On the other hand, Hawaii was developed from the beginning for commercial uses and, as such, it’s optimized for end users. Even though many assumed that this will be the first 20 nm graphics processor, that wasn’t the case. AMD stated that sticking by 28-nm lithography was a calculated move. Even though it had some consequences. TSMC’s 28 nm production process is reliable and guarantees a huge yield (a number of correct chips per buffer). That results in great profitability and affordable end price, which is more important to the market than leading technology. Therefore, in this aspect, nothing has changed, and that leads to conclusion that the only two ways for improving performances are increasing the number of transistors and reorganizing the existing architecture. That’s what AMD did, so Hawaii can boast an impressive amount of 6.2 billion transistors. That’s a 50% increase than with Tahiti which had about 4 billion. Of course, the surface of the chip was increased, but not as drastically. With reorganizing the architecture, AMD managed to fit all those transistors on a surface of 438 mm2. Bearing in mind that Tahiti had 365 mm2, a difference of about 20% is far from significant. Still, in order to achieve impressive performances in games, AMD followed the competition's lead in designing Kepler. Performances in GPGPU was decreased, so the users of BitCoin will probably be disappointed, because the performances of this noticeably more powerful chip will be limited and not that much different than Tahiti. On the other hand, everyone else can rejoice.
In terms of architecture, Hawaii chip is still based on GCN, but this are organized a bit differently and more efficiently. The base is composed of four Shader Engines, input units, Media accelerators, L2 cache and memory controllers. We'll leave Shader Engine for later, and we'll explain right away everything else. Input units include ACE (Asynchronous Computing Engine) units and Graphics Command processors, which send various data over Global Data Share cache towards the Shader Engine, which processes the data. Located next to it are the media accelerators which include UVD and VCE units that deal with video processing and its acceleration, as well as new TrueAudio DSP. Also, there are controllers for CrossFireX, Eyefinity and DMA engine. Lastly, a large part of the chip is dedicated to memory controllers, of which there are now eight, thanks to new design and aside from supporting 512-bit bus, and they take up only 20% more surface when compared to such controllers in Tahiti. Aside from that, another important part in the chip area is L2 cache of 1 MB capacity that is shared among all four Shader Engines. Now, let's take a peek into white lies in the Shader Unit, the most important part of Hawaii GPU.
Hawaii chip has four Shader Engines, and each one is a separate functional entity that has Stream processors, local registries, L1 cache, ROP and texture units. Therefore, virtually each one has the functionality of a small GPU that communicates with the rest of the chip. Hierarchically speaking, there is a Geometry Engine at the entrance that is still a separate entity. It has a geometry vertex assembler paired with one tessellation unit. Since Hawaii has four Shader Engines that means that it has a total of four geometry engines. Therefore, when compared to Tahiti, it’s twice as powerful in terms of geometric processing, but more importantly – tessellation power has been doubled. Of course, we’re basing this assumption that these are the same tessellation units as in Tahiti, since AMD didn’t provide any information regarding this. After data processing, the data is sent into Compute Units (CU), which is next in line. This is the smallest part that has Shaders, local cache of 64 KB each, L1 cache with 16 KB as well as scalar and vector registries. Of course, there are also the 64 vector units which AMD classifies as Stream processors. Each of the 11 CU has 64 Stream processors, meaning there are a total of 704 per Shader Engine. We’ve said a couple of times that there are four of them, so it’s clear that total number of Stream processors amounts to 2816. Since it’s intended for large resolution, including the upcoming 4K monitors, AMD increased the number of ROP units. So far, there are 64, which is a yet another item that’s been doubled when compared to Tahiti. This is very important, because this way Hawaii got twice the amount pixel fill rate, which is very important for monitors with high resolution. Higher resolution means a large number of pixels that needs be displayed, and graphics card with a small number of ROP units will certainly have difficulty processing them at high speed. Therefore, in terms of specifications, it’s clear that Hawaii is an excellent chip, but that’s not all. There are a few more novelties which tell us how Radeon will look like in the future.
DSP, CrossFireX, PowerTune and more...
AMD improved a couple more aspects that aren't directly related to Radeon R9 290X's performance, but which will improve the user experience. First of all, there's the new DSP that belongs to the TrueAudio technology. By implementing this new DSP, Hawaii GPU is capable of dealing with sound processing, in order to enhance the gaming experience and better define spatiality of sound, even while using a stereo headset. Developer teams will have an increased number of options for adapting and improving the sound in games, which AMD notes as being an important element that is often forgotten in the graphics race. Also, there’s the new CrossFireX technology, which has been improved, so there’s no need for the well-known bridge. This cable will probably be forgotten starting with next generation, but it’s important to remember who started the revolution. AMD guarantees that performances will remain the same, or be even better, because now GPUs in CrossFireX mode have even better communication than before. Since Hawaii is a pretty large GPU and demanding in terms of energy, PowerTune had to be redesigned. AMD has improved the optimization by implementing new internal logic of the card along with implementation of IR 3567B which deals with monitoring and voltage changes. With this new algorithm, all the important statistics of the card are noted and a balance is trying to be achieved. PowerTune now tracks the temperature, desired voltages, consumption and based on this data sets the fan speed, frequencies and real voltages. The range available is from 0 to 1.55 V, and the settings made in steps of 6.25 mV which gives it a fine grit so that even the last W, mV and the degree of Celsius are maximally exploited. In short, that would be it, and now we can finally talk about the video card itself. Meet the AMD’s fastest video card with one GPU. Ladies and gentlemen, meet Radeon R9 290X.
AMD Radeon R9 290X 4 GB
We’ve received the test model directly from AMD, meaning that it’s a reference model intended for testing. That also means that we didn’t receive any other equipment, not even the packaging box. Card was delivered in an antistatic bag, which we’ve hurriedly tore apart in order to get to this piece of hardware. In terms of dimensions Radeon R9 290X isn’t the least dramatic. Actually, it looks like it’s no longer than Radeon HD 7970, and if it is, the difference is just a couple of centimeters. Radeon HD 7990 is definitely the reigning king in terms of this. Design is characteristically AMD’s, with slight changes. Color scheme remained unchanged – red and black combination with somewhat more creative emphasis on red details. First important detail is the red turbine, which is the same size when compared to Radeon HD 7970 GHz Edition. However, according to Frank Hering, the turbine is now spinning to a maximum of 5000 RPM and we hope it will be silent. Under the plastic hood there’s a simple cooler made out of thick aluminium ribs. At first, we were a bit puzzled with the absence of heat pipes, but then we realized that the basis for the cooler is the vapor chamber which practically functions like a large heatpipe. Since it lacks backplate, AMD decided to install an aluminium cooler that, asides from cooling the vital parts of the card, is used to eliminate any bending. Underneath the cooler there’s the Hawaii GPU surrounded by as many as sixteen memory chips. Those are Hynix chips rated for use at 5 GHz with voltage of 1,35 V, or rather 6 GHz at 1,5 V. AMD had to lower the memory frequency in order to secure a lower TDP. Maybe even the more important reason is that it would provide the lower price, because 512-bit bus increases the cost. Power unit is surprisingly simple, so the voltage to the GPU is filtered in just six phases. Powering it requires one 6-pin and one 8-pin molex connectors, but also a solid amount of power. Bear in mind that Radeon R9 290X is declared at 250 W, but we believe that its TDP is higher in practice, as is the case with HD 7970 GHz model. On the front panel, there’s the standard set of video outs that include two DVIs and one HDMI and DisplayPort. In terms of clock rate, AMD managed to retain the rate of 1000 MHz, despite the large chip, which is commendable. In situations such as this, manufacturers usually resort to decreasing the rate, but it’s obvious that AMD completely mastered the 28-nm production process. As for the memory, 4 GB GDDR5 RAM works at 1250 MHz, which effectively amounts to 5 GHz. It doesn’t look like much, but when you take into account that it’s a 512-bit interface, the result is a respectable bandwidth of 360 GB/s. Truly, AMD didn’t make any compromise in designing Radeon R9 290X. Simply, when you take into account 2816 SP, 176 texture and 64 ROP units, along with the frequencies, it’s clear that 290X has tremendous power reserves and we only hope that developer teams will be able to harness all this potential.
|AMD Radeon R9 290X|
|GPU frequency||1000 MHz|
|Number of stream processors||2816|
|Number of ROP / textural units||64 / 176|
|Memory||GDDR5 / 4 GB / 512 bit|
|Memory frequency / effective [MHz]||1250 / 5000|
After testing, it wasn’t that hard to draw conclusions regarding this impressive video card. This is the best one-processor Radeon so far and a huge improvement over Radeon HD 7970 GHz. Also, when compared to the competition, which is of huge importance to fans, you’ll be delighted to now that R9 290X is about 10% faster than GTX 708. Actually, the differences between these two models vary from 0 to 15% in Radeon’s favor. The differences increase along with resolution, and video cards such as this one are intended for playing the games above 1080p resolution. That also means that R9 290X is alarmingly close to the Titan that, until now, had the reputation of an untouchable video card and that costs 1000 euros. Not to mention, Radeon’s biggest trump card is the price. The expected price is “only” 460 euros, which makes R9 290X and AMD winners on the market. But, not everything is perfect. In case you decide on buying Radeon R9 290X, wait a bit for the newer models with alternative cooling system. Whether Vapor-X, Dual-X, DirectCU II, TwinFrozr IV or something else. Reference cooler isn't really that good, because despite having two modes (BIOS) that are changed with a switch, neither is particularly quiet. Even the quiet mode, which doesn't allow the cooler to have more than 2000 revolutions per minute, can be heard in the next room under heavy load. The reason is that the turbine system that requires extremely complex cooling profile in order to get good results with a small amount of revolutions. Simply, the card under heavy load raises the temperature up to 95 degrees of Celsius, which is extremely high and achieved easily with the reference cooler. Therefore, if you look a bit, we’re sure that you’ll find a model with alternative cooling that will provide significantly better experience. In terms of performance and price, it’s impossible to find flaws in Hawaii chip and the Radeon based on it.
|Test results||AMD Radeon R9 290X||NVIDIA GeForce GTX 780|
|1920x1080 / 2560x1440 4xAA I 16xAF|
|Futuremark 3DMark (Fire Strike) GPU Score||8184 / 5073||7398 / 4489|
|Futuremark 3DMark 11 (Extreme preset) GPU Score||4214 / 2689||4165 / 2456|
|3DMark Vantage (Extreme preset) GPU Score||27713 / 18090||26673 / 16230|
|Unigine Valley 1.0 (DX11, Ultra, tesselation off) [fps]||63.1 / 40.7||70.2 / 44.1|
|Unigine Heaven 4.0 (DX11, Ultra, extreme tesselation) [fps]||57.6 / 34.9||55.3 / 35.7|
|Crysis Warhead (DX10, Enthusiast) [fps]||79.7 / 55.1||72.4 / 47.1|
|Crysis 2 (DX11, Ultra, HD Textures) [fps]||97.4 / 65.9||95.7 / 58.4|
|Metro 2033 (DX11, very high, tesselation) [fps]||70.8 / 50.2||64.5 / 44.8|
|Metro Last Light (DX11, very high, tesselation) [fps]||69.9 / 46.7||63.3 / 43.3|
|AvP DX11 (DX11, max) [fps]||97.8 / 61.6||83.3 / 52.2|
|Hitman Absolution (DX11, Ultra) [fps]||72.6 / 51.9||53.2 / 43.1|
|Bioshock Infinite (DX11, Ultra) [fps]||133.1 / 72.6||115.1 / 74.8|
|Dirt Showdown (DX11, Ultra) [fps]||93.1 / 75.7||84.2 / 59.1|
|Sniper Elite V2 (DX11, Ultra) [fps]||41.4 / 25.4||40.3 / 25.2|
|Sniper Elite V2 (DX11, High) [fps]||65.6 / 41.1||66.3 / 42.1|
|Sleeping Dogs (DX11, Ultra) [fps]||67.7 / 42.1||55.2 / 31.9|
|Test configuration: Core i5 2500K @ 4.5GHz, ASUS P8Z77-V, 2x4GB AMD Multimedia Edition DDR3 1600MHz, Seagate Barracuda 1TB, Windows 7 64bit, ForceWare 326.41, Catalyst 13.11 Beta V5|
At a price of 460 euros, Radeon R9 290X doesn’t have any competition, because GeForce that’s slower about 10% costs 120 euros more, at best. Need we say more? NVIDIA will certainly have to lower their prices after this, and the only question is, will they be able to lower it enough. On the other hand, Radeon R9 290X will have an increasingly better position, because we’re only about to see the better performances when newer versions of driver come out, and as you well know, each model’s price drops after a month before its appearance on the market. We believe that all enthusiasts will carefully wait for that moment and secure themselves the magic of Hawaii.