Tried it: AMD Radeon HD 7870/7850 - dangerous pit bulls

The first representatives of the GCN architecture, namely the AMD Radeon HD 7970 and HD 7950, have already been introduced. Tahiti is proven, but now here’s Pitcairn which specifically targets high-performance hungry players at a slightly more affordable price. They are the Radeon HD 7870 and HD 7850. These two models came from both GIGABYTE and ASUS, so in addition to various measurements and tuning, we looked at CrossFireX. Let's get started!

 AMD Radeon HD 7800 - the gamers dream? 

If we wanted to characterize the Pitcairn chip in the shortest possible way, we would just say that the Graphics Core Next (GCN) architecture has broken to the top. The tile shows outstanding parameters from both a technical and user’s point of view, so it definitely deserves a few lines, but let’s start with a little recall of the past.

In late 2009, early 2010, the AMD Evergreen series began its conquest. The product family was completed in half a year and was built on four graphics processors. At the time, it looked like it was going to be a hard-to-beat pace. We now know that it was a walking gallop compared to the Southern Islands family. The company came up with three GPUs in just 3 months, which allowed it to attack both the middle class and the upper segment. Now let’s get acquainted with the last piece of the puzzle.

In a point-breaking approach, let’s take two Cape Verde GPUs and knead them together. The mixture thus obtained is called Pitcairn. This makes things a little more complicated, and we also want to repay an older debt.
Pitcairn has achieved a brutal transistor density thanks to TSMC’s 28nm manufacturing technology, with 212 billion transistors per 2,8 square millimeters. The most powerful member of the Radeon HD 7800 series hides 20 GCN arrays, equivalent to a total of 1280 stream processors. This is roughly halfway compared to Tahiti and Cape Verde tiles, but we’ll see Pitcairn more like its big brother - both the frontend and the backend are good examples of this. The chip received two rasters and also two geometric motors.

The latter is not much different from the previous solutions, however, the engineers performed a number of optimizations, which resulted in improved performance and efficiency. The two ninth-generation tessellator units are, of course, capable of working in parallel. Incidentally, their work is facilitated and accelerated by a separate buffer. Thanks to this, and a number of minor enhancements, the Radeon HD 7800 can brutally humiliate the HD 6900 series during tessellation. With this, we essentially managed to get on the same level with the NVIDIA GeForce GTX 500 series, which is quite a big deal, as the rival in this area has not saved much on transistors. In connection with the frontend, we should also mention the Command processor, which is responsible for load balancing and scheduling.

Each CU (Compute Unit) has four texturizers, so the full Pitcairn GPU can manage a total of 80 texturizers and 320 Load-Store units, respectively. These appear to be optimal values ​​despite a significant setback compared to Tahiti graphics processors. The component hierarchy ignores radical innovations and changes, but this was not really necessary.

Because AMD was primarily looking to meet the needs of the gamer society with the Radeon HD 7800 series, Pitcairn embarked on a major cleanup of the graphics processor, as performance under GPGPU applications is no longer a significant factor here. The size of the second-level cache dropped from 2 MB to 512 Kbyte, and the peak computing power dropped from 947 GFLOP / s to 160 GFLOP / s for double accuracy. It is a significant setback for the big brother, but not for the NVIDIA GK104 processor anymore, as it has similar technical characteristics. Regardless, the GPU is well acquainted with the DirectCompute 11.1, OpenCL 1.2 C ++ AMP APIs, so there is nothing to worry about. Seeing weirder information than usual in official marketing material. According to the company, the time has come to replace the ATi Radeon HD 5800 series controllers with a member of the AMD Radeon HD 7800. True, the former series can no longer be called the king of video cards, but no matter how we twist and twist, even this change would be a strange step. On paper, there is a significant difference in terms of idle consumption, but marketers have slipped a bit at both extremes, so under sharp conditions we would be greeted by a different picture. The main weakness of Cypress chips is the tessellator unit, but this is currently encountered in few games. For obvious reasons, we would not comment on the benefits of the PCI Express 3.0 interface. 3D Mark 11 was left at the end. Obviously, this choice could have been made because the difference could be more significant here. In this synthetic measurement program, we can talk about a difference of roughly 30 percent. Looking at the price tag of just over 100 thousand forints, this is all but convincing. In terms of performance, we do not recommend switching at all, no matter how AMD wants it to be.

Turning back to architectural dissection. The biggest surprise came in the backend area. Pitcairn contains 8 ROP clusters in the same way as Tahiti. Each of these “arrays” contains four ROP units and 16 Z samplers, so the number of pixels calculated in one second at the operating frequency of the HD 7870 1 GHz has developed very favorably. The paper interface has already entered the memory interface. The four 64-bit memory controllers have a total capacity of 256 bits. The default video memory size is 2048 MB. The 4GB package is also theoretically feasible, but it wouldn’t make much sense. The strong backend capacity can create quite interesting situations.

Of course, Pitcairn has DirectX 11.1, PCI Express 3.0, Eyefinity 2.0 and Zero-Core support. The latter procedure has not been implemented by NVIDIA, although thanks to ZeroCore technology, with the display turned off, the entire graphics controller can be de-energized and no active cooling is required in this form. The benefits are convincing: zero noise, 3 watts of power consumption. Consumption indicators performed well as a result of near-perfect balancing. For the Radeon HD 7870, the PowerTune limit is 190, while the typical power consumption is 175 watts. The card is idle below 10W and 3W is also available through ZeroCore. The GHz Edition mark can also be found here, which is intended to indicate that the product has reached an operating frequency of 1 GHz. The Radeon HD 7850 is no shame either. The maximum power requirement of the controller is 150 watts, but it is usually up to 130 watts. Due to similar parameters, idle consumption may be roughly the same. 

Detailing an important area was missed due to lack of time in our Radeon HD 7970 and HD 7950 articles, and this is PRT (Partially Resonant Textures), i.e. hardware virtual texturing. Today’s graphics processors have a major headache in handling large amounts of textures. As the player travels from one section of the track to another, the CPU, graphics card, and data storage are captured for continuous work. Because you have to work with a large amount of information, it can easily happen that loading won’t be smooth - I don’t need to go into detail with Rage. AMD wants to bridge this problem with a hussar cut. The solution is really simple, but great. The point is to consider the VRAM of the graphics card as a texture cache system.

The soon-to-be-used textures should be loaded into the video memory before they are applied, which has already solved the problem. When the GPU wants to use the data, it can access it (for understandable reasons) lightning fast. In a certain sense, this can be considered a kind of texture "streaming" process. Therefore, PRT dynamically loads the selected textures, thus avoiding bandwidth hogs, i.e. clogging of bandwidth, even with large files. Unfortunately, the graphics engine of the given game must be prepared separately for this. We can be sure that Doom 4 will support Partially Resident Textures (PRT) implementation. John Carmack was probably already waiting for this moment with champagne, and considering Rage's strange graphical anomalies, we received it with similar enthusiasm.

UVD 3.0 also offers hardware acceleration for DivX / Xvid, MPEG-4 Part 2 MVC content, and Video Code Engine (VCE) is the AMD equivalent of Intel Quick Sync Video. VCE is stand-alone hardware and is only designed to speed up the transcoding of H.264 videos. The engine is slower than the shader processors in the graphics processor, but much more energy efficient. There are two modes available to users. In the first, only the VCE works, which in itself is faster than most CPUs. In this case, we will not experience a slowdown, we can load the video card or the central unit without any problems. The second option is hybrid mode. The arithmetic-logic units of the VCE and the GPU jump to the task together. This “marriage” obviously has a good effect on the encoding speed, but in that case, don’t be surprised if your favorite game switches to “slideshow” mode. Now it would be really good to see exactly what the system is capable of in sharp conditions, but without the right support it is even further away.

One of the interesting things about Eyefinity 2.0 is that it allows you to conduct multi-monitor conference calls with multi-track audio. The official name of the procedure is Discrete Digital Multi-Point (DDM) Audio.

After a brief introduction to the architecture, we are now paddling into other waters. Since the introduction of the Southern Islands family, AMD has been constantly striving to improve the image quality achieved in games. The first major update concerned the Radeon HD 7900 series controllers. The advent of SSAA (Super-sampling Anti-Aliasing) anti-aliasing under DirectX 10 and DirectX 11 was definitely a joyous phenomenon. In later Catalyst drives, the AutoLOD algorithm can further improve graphics quality, but that's not all, because programmers are already working on the MLAA 2.0 process. The new solution is significantly faster than its predecessor - as confirmed by Tom's Hardware and Anandtech measurements. Another good news is that (on paper) it can be used from the Radeon HD 4000 series. Judging image quality is subjective, but overall it looks like we managed to move on here as well. The second episode of the story is the improved anisotropic filter algorithm. This theoretically completely eliminates the flicker, tremor, and other anomalies we’ve experienced before, which sounds pretty good - we wouldn’t take poison for that. Otherwise, the new algorithm does not require additional buffering, so it does not burden the system resources more.

At both second and third glance, we have to say the Pitcairn chip has become very balanced. Thanks to the optimizations, the chip's size and consumption developed favorably, which practically resulted in the Radeon HD 7950 receiving the strongest competition within the housing. AMD has severely undercut the smaller Tahiti-based card, which has probably completely lost its previous popularity. This seems like an interesting step, but in retrospect it makes sense that the GK104 has become a really successful gamer chip. The engineers on both sides gave their best knowledge. Pitcairn leaves room for maneuver, as the cards built on it produce quite a good level of performance while manufacturing costs are at a favorable level. Different variants of GK104 produce higher speeds but with more expensive production. AMD simply needed such a chip because Tahiti has moved so much towards GPGPU that it can already be beaten on a gamer line. 

The company’s product numbering has now evolved a bit strangely. The Radeon HD 7800 series seems to be a definite improvement over the HD 6800, making the HD 7870 and HD 7950 virtually in a league. The Radeon HD 7700, on the other hand, isn’t a big bang in terms of performance compared to the HD 6700 series, though the latter is just the result of a renaming. In practice, the nearly two-and-a-half-year-old Juniper GPU is barely 20% outperformed by Cape Verde, which we admit is not a convincing step forward at all.