The R6xx

The R6xx

We won’t go into detail on R6xx architecture as we already did this in the test of theRadeon HD 2900 XT. Let’s just remember that it’s a unified architecture, or in other words all calculation units can process anything whether they are vertices or pixels. Contrary to Nvidia, who opted for scalar functioning on its units, AMD stayed with a vectorial model even though they claim the contrary. The calculation units of R6xx chips can therefore be seen as very flexible MIMD 5-way units, but they are still vectorial and the compiler has to assemble a certain number of independent instructions in order to fill these units to the maximum. For Nvidia’s G8x, the maximum use of units is done naturally.

Seen from another way, the architecture developed by Nvidia with fewer ALUs allows similar performances to vectorial architecture which has a lot more ALUs.

For reference, here are first of all the specifications for high end graphic cards:

the RV630

Part of the new mid range GPU line, the RV630, which equips the Radeon HD 2600, is made of no less than 390 million transistors versus "only" 289 million for the G84, which equips the GeForce 8600. Thanks to the 65 nm process, it is nevertheless smaller at 143 mm² versus 169.


It has 24 vec5 calculation units instead of the 32 double pumped scalar ones of the competition, which gives it a largely superior calculation power:


Of course, the G84 architecture is more efficient and this partly compensates for this difference in calculation power. However, the biggest difference probably stems from the fact that while the two GPUs have 8 texturing units, those of the G84 are capable of carrying out two bilinear filterings per cycle. To this we add the greater number of ROPs, which is just four for the RV630 versus the G84‘s eight.

Note that just like the R600, the RV630 has two main L2 caches which allow the hiding of read/write operations. However, its size is reduced to 128 KB (compared to 256 KB for the R600). There is a 128 bit memory bus, just like the GPUs from the same product line.

The RV610

The RV610 is a very small chip at only 85 mm² compared to 115 mm² for the G86, the competition. This time the AMD chip has fewer transistors at 180 million versus Nvidia’s 210 million.


It’s equipped with eight vec5 calculation units instead of the sixteen double pumped scalar units for the G86. Luckily for AMD, the 65 nm size allows to set the RV610 at higher speeds than the G86.


There is an identical number of ROPs, but this time AMD has an advantage in terms of filtering with four simple units at 700 MHz versus four double ones at 450 MHz, the former being more flexible.

There is no L2 cache for the RV610, which has a more classic memory structure, in order to save on transistors. The memory bus also changes to 64 bits compared to 128 bits for the Nvidia chip. This is a strange turn of events, because in previous generations, it was the contrary.

Specifications, the cards

Specifications

Note that AMD indicates 800 MHz for the memory of the Radeon HD 2600 XT GDDR3 in certain documents, and 700-800 in others. This isn’t too clear. When asked about this, AMD told us that they didn’t fix memory frequencies and believed that partners would probably go with memory at 800 MHz. A Radeon HD 2600 XT can therefore have memory that varies between 700 and 1100 MHz, or a difference of more than 50%!

Radeon HD 2600 XT GDDR4

AMD supplied us with a reference Radeon HD 2600 XT GDDR4.


The GPU is set at 800 MHz and memory at 1.1 GHz. All of this is cooled by a copper radiator. The card is silent given that the Radeon HD 2600 barely heats up and so the fan can run at minimum speed.

Powercolor Radeon HD 2600 XT GDDR3

Just in time, Powercolor was able to supply us with a Radeon HD 2600 XT equipped with GDDR3 memory.


The GPU is still set at 800 MHz, but memory is decreased to 700 MHz. This card is also very silent.

Radeon HD 2400 XT

This was also an reference card supplied by AMD.


The design is rather unique and we wonder if it might have been more interesting to have a smaller PCB and do away with the CrossFire connection, which in this case is more symbolic than of real use.

DirectX 10 and antialiasing

DirectX 10

For this test, we decided not to include any available DirectX 10 games, because one was developed with the help of Nvidia and the other with AMD. We prefer to wait to see more clearly on a later date.

Nevertheless, we can now confirm without any doubt that GeForce 8s have big performance problems when geometry amplification is used in geometry shaders. The reason lies in the fact that these GPUs have to keep all data inside the GPU, while geometry amplification might generate tremendous amounts of data. In order for it to remain in the GPU, GeForce 8s have to reduce the number of elements processed in parallel. In order to have an optimal efficiency, a group of eight scalar processors should contain at least 192 elements (vertices or pixels) that should be spread out between the 8192 32 bit registers that are available. With geometry amplification that allows the generation of 1024 32 bit values per polygon, we can easily see that the number of elements will be a lot less than 192, which strongly reduces real calculation power.

For AMD, a main L2 cache which allows the hiding of reading and writing operations is used to give the GPU a virtually unlimited number of accessible registers more or less efficiently. This lets the GPU continue to function in an optimal way by writing the values generated by the geometry shader in video memory as soon as there aren’t enough "hardware" registers.

The performance gain can be enormous and we noticed differences of up 50 times higher (not %!) between AMD and Nvidia. Of course, it’s now up to developers to see what they can do with this possibility. Nvidia won’t of course incite them to do this too much.

Note that the Radeon HD 2400 does not have an L2 cache. Accesses to video memory, therefore, aren't hidden which should reduce performances. At any rate, as you may have noticed, this GPU is really not equipped to handle DirectX 10 games with advanced functions.

Antialiasing

We won’t come back to the CFAA which we already wrote here. One small simple clarification will do. The Radeon HD 2400 doesn’t support MSAA 8x and is limited to 4x.

Antialiasing in the Radeon HD 2900 has spilled a lot of ink because its implementation is partly bugged. This is also the case for the Radeon HD 2600 and 2400. It will probably be of use here to describe once again the problem, because there is a lot of confusion on the subject.

There is no problem with the quality of antialiasing and only performances are reduced. In a GPU, as a general rule, the ROPs contain a fixed unit responsible for mixing samples taken from an MSAA buffer to make the final pixel. In the current Radeon HD 2000, this unit doesn’t work. So what happens? The shader core is used to make this mix, which has the advantage of having access to a programmable unit. However, performances are reduced, because general calculation power is used. DirectX 10 makes GPU manufacturers to allow developers to do the same operation but here we can suppose that AMD uses a more effective method in its drivers to carry out this operation. The only thing that remains is the unavoidable cost on performances.

Note that this has nothing to do with shader based antialiasing such as that used in S.T.A.L.K.E.R. For AMD, a real MSAA buffer is used in an adapted resolution, while in S.T.A.L.K.E.R., there is only a filter applied to a classic image based on a certain amount of data.

HD Video

Vidéo

Contrary to the Radeon HD 2900, the Radeon HD 2600 and 2400 have UVD, an engine devoted to decoding video, which completely supports decompression of H.264 and VC-1 formats at the maximum speeds authorized by HD-DVD and Bluray. These are capacities similar to those introduced by Nvidia with the GeForce 8600 and 8500, except that Nvidia only partially manages VC-1 decoding.



We tested the performances of the different cards in reading HD videos. To do this, we use two HD-DVDs. The first, « Babel » is encoded in H.264, and the second, « King Kong », uses the VC-1 format. The CPU here was a Core 2 Extreme X6800. For some unknown reason, the rendering of videos wasn’t fluid and was corrupted on the Radeon HD 2600 XT GDDR3.


It was no surprise that the GeForce 8600/8500 and Radeon HD 2600/2400 showed excellent performances with an advantage to Nvidia. CPU use rose once reading started on the other GPUs. Also, for Nvidia the GeForce 7600 GT suffered from fluidity problems, while the GS version had so many problems that viewing wasn’t possible.


In VC-1, all Nvidia GPUs showed similar performances, with the GeForce 8600 and 8500 not entirely supporting the decoding this format. Therefore, here the advantage went to the Radeon HD 2600 and 2400.

Power consumption, the test

Power consumption

As usual, we measured consumption with the different cards. Measurements were taken at the wall socket, in order to have the total power consumption of the power supply, in this case, an Enermax Galaxy 850W. We give you the figures obtained in Windows desktop, then in load with the Pixel Shader test in 3DMark06 combined with Prime95. The latter allows us to have constant processor use independent of the performances of the graphic card. We also added the system’s power consumption while playing the two HD-DVDs.


In idle, while consumption of the Radeon HD 2400 XT in stand-by is very low, this isn’t the case for the Radeon HD 2600 XT GDDR4, whose values are unreasonably high. Is this due to the GDDR4 ? It’s the same in load, and therefore, the Radeon HD 2600 does not have lower consumption than the GeForce 8600 in 3D.

When playing video, the Radeon HD 2400 XT is more economical. While the Radeon HD 2600 XT GDDR4 has CPU use that is less than the GeForce 8600’s in VC-1 reading, its power consumption here is nevertheless higher.

The test

For this test, we used Windows Vista and eight games in 1280x1024 as this is the basic resolution of current screens. While we usually test high end graphic cards with a very high level of detail, here we used high or medium quality modes.

Test configuration:

Intel Core 2 Extreme X6800
eVGA nForce 680i
2 GB of RAM
Windows Vista
Drivers AMD beta 8.38.9 RC2
Forceware beta 162.15

Quake 4

Quake 4

Here, we saved an action scene. Unlike Doom 3, there are fewer shadows but more characters and textures, which changes the load in terms of rendering.

In high quality mode, the game automatically activates 8x anisotropic filtering. In medium quality mode, there is a simple trilinear filtering and special effects are reduced.




The first thing that we noted with Quake 4 is that lowering the quality has little influence on performances. In this first test, the Radeon HD 2600 XT is behind the 8600 and 7600 GT, but it does show a gain compared to the X1650 XT.

The HD 2400 XT finishes last.


It’s a similar situation with 4x FSAA such that the Radeon X1650 XT is equivalent to the Radeon HD 2600 XT.

Half Life 2 Lost Coast

Half-Life 2 Lost Coast

Here we used an internal demo of Lost Coast that we saved in a way to test Valve‘s HDR, which uses a rather complex rendering format that doesn’t take into account the newer capabilities of the latest graphic cards. It runs on all DirectX 9 cards with MSAA. For the high quality mode, all options were activated as well as 4x anisotropic filtering. In medium quality mode, HDR and anisotropic filtering were deactivated.




The Radeon HD seems to be at ease with Half Life 2, because the HD 2600 XT is at the same level as the 8600 GTS and the HD 2400 X is ahead of the 8500 GT.

Once antialiasing was activated, the Radeon HD’s performances plummet.

F.E.A.R.

F.E.A.R.

We use the integrated demo. Unfortunately, this only gives a whole number score, which can lead to a difference of one unit under the same conditions because of a normal variation of two tenths. For each card, we selected the best of three results.

All graphic options are pushed to the maximum except for soft shadows, which were deactivated because they don’t work with AA. 4x anisotropic filtering was activated via the game. In medium quality mode, options were set to medium and trilinear filtering replaces 4x anisotropic.




In F.E.A.R., the Radeon HD are largely behind and are equivalent to the GeForce 7600 GT and X1650 XT.


Activation of antialiasing makes the situation worse for AMD. The X1650 XT loses quite a bit less in terms of performances than the HD 2600 XT.

S.T.A.L.K.E.R.

S.T.A.L.K.E.R.

We carry out an identical movement and measure the framerate with fraps. The test was done in high quality, complete dynamic lighting, and high detail followed by another session in medium quality, partial dynamic lighting and low detail. S.T.A.L.K.E.R. uses an engine based on differed rendering, which is fundamentally incompatible with MSAA and makes the use of antialiasing impossible. A type of filtering of edges carried out with a shader can be activated but results are mixed. The 1.00001 patch is used.




Nvidia largely dominates here.


Nvidia is still ahead, but its advantage is reduced once antialiasing is activated.

Rainbow Six Vegas

Rainbow Six : Vegas

We measure performances in the introductory scene. In high quality mode, 4x anisotropic filtering is activated via drivers. The HDR mode is activated as it is more or less obligatory, because without it banding is very noticeable. Shadows are set to “low” and setting them to a higher quality lowers performance too much in certain areas. In medium quality mode, all options are set to a minimum and no anisotropic filtering is activated.




In Rainbow Six Vegas ,the Unreal Engine 3 is used (this should be questioned given the graphics and resulting performances) and the Radeon now takes the lead. The HD 2600 XT even surpasses the Radeon X1950 Pro.


Unfortunately, only Nvidia has implemented FSAA support in its drivers for this game, which is something not offered in the game. We informed AMD about this problem and Chuck, creator of the Oblivion patch, was assigned this task. Let’s hope that this will be quickly added to AMD drivers.

Oblivion

Oblivion

We saved a specific movement in order for it to be always identical and the test reproducible. For the high quality mode, we activated HDR, used higher settings and 4x anisotropic filtering was activated in the drivers. In medium quality mode, there was no anisotropic filtering and medium settings were activated.




The Radeon HD don’t do too badly here, even if the GeForce 8600 GTS is slightly ahead. The HD 2400 XT is ahead of the GeForce 8500 GT.


Once FSAA is activated, the performances of the Radeon HD plummet again. You may recall that the GeForce 7 are incapable of applying classic FSAA with the HDR FP16 mode used here.

Age of Empire III

Age of Empire III

For this test, we saved a sequence in which we carry out a specific movement when the game is paused. Aniso is activated in the game as well as HDR in high quality mode. However, we should note that while results are similar for both ATI and Nvidia, it’s a different process. Based on FP16 for Nvidia, it’s in FX10 for ATI. While it only has 2 bits for transparency, the FX10 mode allows staying in the classic 32 bits, which is ideal for performances. In medium quality, HDR is deactivated as well as high quality shadows and trilinear filtering is used.



Despite a less heavy rendering for ATI, Nvidia shows higher performances in this game.


In medium quality, the Radeon HD lose quite a bit with the activation of FSAA. This also happens to the GeForce 7 but for a different reason. Given that they don’t support HDR FP16 and MSAA, a 2.25x supersampling is emulated by developers, significantly reducing performances.

Colin McRae DIRT

Colin McRae DIRT

To test Colin McRae‘s latest opus, we carried out a well defined sequence. The quality modes correspond to those of the game. Note that activation of antialiasing is highly recommended given the way menus are rendered and that post process effects amplify aliasing.



The Radeon HD 2600 XT is equivalent to the GeForce 8600 GTS.


Here, it’s exactly the same except the Radeons lose a little more.

Recap

Recap

Here is an index of performances, which we calculated by giving all games the same weight and based on a score of 100 for the Radeon HD 2600 XT in high quality mode:


Overall, we can say that the Radeon HD 2600 XT GDDR3 is equivalent to the GeForce 8600 GT and the GDDR4 version is slightly above. However, this difference between the two 2600 cards is not enormous despite 50% more memory bandwidth thanks to the GDDR4 1.1 GHz.

As for the Radeon HD 2400 XT, it’s ahead of the GeForce 8500 GT.



Once antialiasing is activated, the Radeon HD collapses, especially in medium quality mode, which is much more sensitive to antialiasing performances. While the average reduction in performances with antialiasing is between 25 and 30% on all other cards, it is at 40-45% on these three Radeon HDs.

The Radeon HD 2400 XT thus finishes last and the Radeon HD 2600 XT doesn’t manage to keep up with the GeForce 8600 GT.

Conclusion

Conclusion

Just like with Nvidia’s DirectX 10 midrange and enre level boards, our opinion on these new Radeons is mixed. We hoped that the transition to 65nm would allow AMD to actually offer an interesting mid-level line. Unfortunately, the 35% more transistors of the Radeon HD 2600 compared to the GeForce 8600 don’t add higher performances and only serve in squandering the additional costs of AMD’s DirectX 10 architecture. Obviously, this architecture has higher performances in geometry shaders, but this doesn’t justify the existence of the 100 million supplementary transistors. AMD’s architecture is overall less efficient which creates the need for more calculation units in order to offer equivalent performances.

But the real disappointment comes from relatively poor antialiasing performances due to ROPs that are not entirely functional. This weakness is felt even more, because traditionally this has been a strong point for Radeons. Besides good DirectX 10 support, we have to look to video to find the interest of this line. Here, it is particularly well adapted for this use with the integration of HDCP support, an audio controller to facilitate the transport of sound in secure environments via an HDMI connection, and efficient decoding of the two flagship HD formats.


To play in 1280x1024 with recent games, the current mid-level won’t entirely be enough to take advantage of the latest effects. You will have to make a compromise and turn to cards like the GeForce 8600 GTS and if your budget allows, you can go even higher or wait for the second generation of mid-level DirectX 10 cards, which should arrive at the end of the year or beginning of 2008 and offer more interesting models in terms of performances.

To be clear, we find the Radeon HD 2600 XT GDDR4 of no interest. The difference in performance compared to the GDDR3 version is reduced, power consumption is much higher, and it is much bulkier. The GDDR3 model is more interesting and should be more attractive if it is found for around 100€ as recent rumors have suggested. Its performances with antialiasng, however, lag and this card is more adapted to 1024x768 than 1280x1024.


In our opinion, the GeForce 8500 GT, Radeon HD 2600 Pro and Radeon HD 2400 XT don’t have any interest either except to fill Nvidia’s and AMD‘s catalogues. These cards have low performances for gamers and the entry level DirectX 10 lines offer the same video capabilities. In other words, if you don’t play games and are looking for a graphic card adapted to HD video, the GeForce 8400 GS and Radeon HD 2400 Pro will suffice, the latter being our preference thanks to its standard HDCP support.