by Alexey Stepin , Tim Tscheblockov, Anton Shilov
09/21/2004 | 04:54 AM
It’s been merely days since NVIDIA Corp. unveiled its new family of mainstream GPUs. Equipped with the PCI Express interface, GeForce 6600 chips are the first mass product to deliver the performance of previously top-end solutions like RADEON 9800 XT or Pro. NVIDIA became the first manufacturer to fill in this niche of mainstream graphics cards for the PCI Express platform, now being rapidly deployed. It became the first, but could hardly hope to remain the only one. ATI Technologies, NVIDIA’s archrival in the PC graphics field, had also been preparing a riposte with a new weapon, codenamed RV410.
<%BANNER[article]%>You may remember that we had already witnessed much the same fencing earlier this year: NVIDIA was the first to release its new flagship product, the GeForce 6800, whereas ATI took a while to disclose its RADEON X800. The same old story recurs once again and now we welcome the new niche-mate to the GeForce 6600 family – the ATI RADEON X700 graphics processing unit!
The RADEON X700 GPU must fill the vacuum between ATI’s existing RADEON X800 and X600 families. The former of the two will remain in the sector of expensive and highest-performing products, while the latter is going to move down into the realm of inexpensive mainstream graphics cards now that the X700 has arrived. The line-up of ATI’s PCI Express solutions is now complete, and any user of that platform can pick a graphics card with an ATI chip according to his/her own tastes and finances. Well, there are still unused smaller names like X400 or X500, which may be employed in the future, too.
By the way, the series of PCI Express products from ATI looks more logical and comprehensive than the RADEON 9xxx nomenclature, in which there were products with a bigger number in the name but worse characteristics in comparison to products with a smaller number. We hope ATI will keep this consistency in the future, so that inexperienced users wouldn’t be confused by the misleading numbers.
Now it’s time to move on to the subject of this review proper – the X700 graphics processor with all its qualities. Good or bad? Let’s find out.
Unlike its elder brother X800, the RADEON X700 is manufactured with a thinner 0.11-micron tech process, which did well in the low-end ATI RADEON X300 processor. The RV410 chip consists of about 110 million transistors – much less than in the NV43 and slightly less than in the RADEON 9800 XT. The use of the new technological process usually leads to higher clock rates and a lower heat dissipation level; coupled with the relatively low transistor count it also potentially reduces the cost of the graphics chip. We’ll find out shortly if any of these truths applies to the X700, but now we are going to examine closely the architectural specifics of the chip.
Like any other PCI Express product from ATI, the RADEON X700 features native support of this new peripheral bus (as you know, this interface is realized through a special bridge in chips of the GeForce PCX and the GeForce 6800 families). The foundation of the new GPU is formed by the successful X800 architecture with all its features: 3Dc, SMOOTHVISION HD, SmartShader HD, VideoShader HD, HyperZ HD, to mention a few. You can refer to our review of the RADEON X800 series for a detailed description of these technologies.
The X700 comes without support of Shader Model 3.0, but of course can execute pixel shaders of version 2.0b. Thus, NVIDIA retains its technological leadership as concerns executing pixel shaders. Anyway, game developers are not rushing yet to add this shader model into their products, and we doubt they will change their attitude in the near future.
As mentioned above, the X700 resembles the X800, but has fewer pixel pipelines (eight) and a narrower memory bus (128 bits). Instead, ATI equipped its new mainstream GPU with six vertex processors, like in the more expensive chips from both ATI and NVIDIA. This is twice the number of vertex processors of the GeForce 6600 chip!
So, the RADEON X700 communicates with GDDR3 memory across a 128-bit bus, like its market rival, the GeForce 6600, does and this narrow bus may become the bottleneck in high resolutions and/or with full-screen antialiasing. The use of a 128-bit bus and fast memory is, however, justifiable with respect to mainstream graphics cards: the price of the fast memory chips, which must be used to ensure a higher memory bandwidth, is going to decrease, while the price of the complex PCB and the die packaging, necessary to realize a 256-bit bus, is going to remain at the same high level throughout the lifecycle of the product.
By the way, the company offers three flavors of the X700: RADEON X700 XT, RADEON X700 PRO and RADEON X700. They only differ in the clock rates and the memory amount, though:
| RADEON X700 XT | RADEON X700 PRO | RADEON X700 |
Core Clock | 475 MHz | 420 MHz | 400 MHz |
Pixel Pipelines | 8 | 8 | 8 |
Vertex Pipelines | 6 | 6 | 6 |
Memory Type | GDDR3 | DDR? | DDR? |
Memory Size | 128MB | 256MB | 128MB |
Memory Clock | 1050 MHz | 864 MHz | 700 MHz |
Memory Bus | 128-bit | 128-bit | 128-bit |
Recommended Price | $199 | $199 | $149 |
The cards will have all the connectors you’d expect them to have: DVI-I, D-Sub and S-Video. Interestingly, the XT and PRO models have the same recommended price, so we will have a choice between a faster graphics card with less of memory and a slower graphics card with a double memory amount. The price being equal, the RADEON X700 XT seems preferable between the two, though, as 128MB of graphics memory is quite enough for almost any application today. The RADEON X700 PRO is unlikely to be shipped with expensive and fast GDDR3 memory, so it won’t probably overclock to the frequencies of the X700 XT. On the other hand, the PCB design of the RADEON X700 XT and X700 PRO is identical, so some graphics card manufacturers may put 256MB of memory on board of their versions of the RADEON X700 XT unless ATI objects to such actions of its partners.
The new family of mainstream GPUs that should press the RADEON X600 family down and become a PCI Express alternative to the RADEON 9800 XT looks good at the first approach, save for the narrow memory bus. We are going to see soon if this 128-bit bus has any negative impact on the performance.
Alas, we couldn’t get a sample of the RADEON X700 XT into our greedy hands to put it under the most merciless tests in our own test labs. It was only a couple of weeks before the announcement, scheduled for September 21 then, that ATI Technologies finished the specifications of the RADEON X700. That was the reason why the company couldn’t make enough samples for the testing media around the world.
So, there are really very few RADEON X700 graphics cards in existence as yet, but ATI was very kind to offer X-bit labs to test the graphics card in their office, taking care about all the organizational issues. We couldn’t miss the opportunity and sent our reporter Anton Shilov – with our own testbed and all – to Munich, Germany, where the European office of ATI Technologies resides.
Anton got all its reports back to our lab in Tallinn, Estonia, to complete this review. The materials he gathered include the results of various tests as well as good snapshots of the RADEON X700 made on the spot. Thus we hope to provide you as much information as we do in any of our reviews we prepare in our labs. So, here it is, the hero of the day, the RADEON X700:
This is the elder model in the family, the RADEON X700 XT. The card is rather compact – the dimensions of its PCB roughly correspond to those of the GeForce 6600 GT, although the PCB itself looks more complex, due to the numerous small elements. You can notice a seat left for a Rage Theater chip, endowing the card with the VIVO functionality, but the chip was missing in our sample. The power circuitry is very simple – simpler and of fewer elements than in the GeForce 6600 GT. For example, there are only two big electrolytic capacitors here, while the power circuit of the GeForce 6600 GT has as many as eight.
The card offers the standard set of connectors, but a version with two DVI-I outputs is possible – this option may be demanded, considering the ever-rising popularity of LCD panels. By the way, it also means that the chip now has an additional TMDS transceiver that earlier GPUs from ATI had not. Like with the GeForce 6600 GT, we see a place left for an additional power connector here, although the card doesn’t actually need it. We will hardly see such a connector on off-the-shelf RADEON X700 XT cards, since they should be quite satisfied with the 75 watts of electricity the PCI Express bus can cater to it.
The new cooling system is a whole new story. It can’t boast any exotic features like a blower, but looks well-designed anyway: the copper heatsink under the casing should be efficient enough. The heatsink covers the GPU as well as the GDDR3 memory chips, which is rather a rare solution nowadays. You may remember that the GeForce 6600 GT and the RADEON X800 XT/PRO have no cooling of the memory chips at all.
As a matter of fact, GDDR3 memory chips generate less heat than GDDR2, for example, but they are anyway hot, working at frequencies of about 1 gigahertz. That’s why ATI’s design looks right, and it was only the tiny fan that raised our concerns. With a short diameter and small blades, it must rotate at a very high speed, which just can’t be well for the user’s audio comfort. We will discuss the noise factor below; now let’s view the other side of the card:
The back side of the PCB is very densely populated, too. Besides a crowd of small elements, there are empty places for four more memory chips (as said above, the RADEON X700 will come with 256MB of memory). As you see, the path from the GPU to the memory chips is not as straightforward as on the PCB of the GeForce 6600 GT.
The black metal bracket is an element of the cooling system and serves to hold the heatsink and prevent the PCB from any twists. Two screw-nuts are pressed into it. The spring screws go into these nuts and hold the heatsink. There’s nothing interesting left on this side of the PCB, save for the memory voltage regulators in the upper left corner and the sticker that says we deal with a sample of the RADEON X700 XT.
Unfortunately, we couldn’t make snapshots of the graphics core or memory chips – running this test session in almost “field conditions” we had no thermal paste and tools to remove the cooler to do that.
We have reasons to think, though, that the card carries 2ns K4J55323QF-GC20 chips from Samsung rated for 500 (1000DDR) MHz frequency. On the reference card, however, the memory was clocked at 525 (1050DDR) MHz, and the graphics processor at 475MHz.
Returning to the issue of noise, we should confess our comprehensions were well grounded – the small and fast fan with a not-very-optimal design of the blades was rather loud even in the 2D mode when it was working at a reduced speed. When speeding up to its full in 3D, it became really annoying, at least in an open system case. Subjectively, this fan was louder than the noise from the cooler of the GeForce 6800 Ultra working at its maximum speed, although this latter has to cool a much more complex and hot chip than the X700. By the way, the sample of the X700 we dealt with was rather hot at work, in spite of the 0.11-micron tech process.
Still, you should remember that it was only an engineering sample of the card, only a few days old. That’s why we can’t say that off-the-shelf RADEON X700 cards are going to be as hot and noisy. By the way, it is rather strange to see such a regress in cooling systems on ATI graphics cards. Until now, all of the company’s cooling solutions have been most caring about the user’s ears – the coolers on the RADEON 9800 XT and the RADEON X800 XT/PRO were practically silent. Again, we hope they will improve the cooling of the X700 before the card reaches shops.
There were no problems with 2D quality – the card was outputting a sharp picture in all the display resolutions, like any decent modern graphics card should do.
Unfortunately, we couldn’t examine the overclocking properties of the new graphics processor due to time constraints, but we anyway carried out a few theoretical tests you are invited to look at right now.
We used two platforms with PCI Express and AGP graphics buses, respectively.
The PCI Express platform:
The AGP platform:
To avoid confusing you, we mark the graphics cards tested on the AMD64 platform (RADEON X800 PRO and RADEON 9800 XT) with the asterix sign (*).
Marko Dolenc’s Fillrate Tester opens the cycle of our synthetic tests.
The RADEON X700 XT quite naturally does multi-texturing slower than the X800 PRO that has 12 pixel pipelines against the X700 XT’s 8 pipes. More discouraging is how the new GPU compares to the NVIDIA GeForce 6600 GT: outperforming the rival with no textures and being equal to it at single-texturing, the X700 XT turns in much worse performance when rendering more textures. These two graphics cards – ATI RADEON X700 XT and NVIDIA GeForce 6600 GT – have the same number of pixel pipelines and similar GPU clock rates, but the former is two times less efficient than the latter at rendering 4 textures!
The next table shows you the efficiency coefficients of the graphics cards, i.e. the ratios of the real performance to the maximum theoretical speed at simple texturing:
| ATI RADEON X700 XT | 0 textures | 1 texture | 2 textures | 3 textures | 4 textures |
| Real Fill Rate (Mpixels/Sec) | 2578.4 | 1824.2 | 1045.7 | 721.1 | 476.8 |
| Theoretical Fill Rate (Mpixels/Sec) | 3800 | 3800 | 1900 | 1267 | 950 |
| Effectiveness (%) | 67.9 | 48.0 | 55.0 | 56.9 | 50.2 |
| NVIDIA GeForce 6600 GT | 0 textures | 1 texture | 2 textures | 3 textures | 4 textures |
| Real Fill Rate (Mpixels/Sec) | 1926 | 1863 | 1589 | 1159 | 868.5 |
| Theoretical Fill Rate (Mpixels/Sec) | 4000 | 4000 | 2000 | 1333 | 1000 |
| Effectiveness (%) | 48.2 | 46.6 | 79.5 | 86.9 | 86.9 |
The cards behave quite differently: the efficiency of the NVIDIA GeForce 6600 GT is steadily increasing to reach as high as almost 87%, while the ATI RADEON X700 XT is less successful – only 48-57% when rendering from 1 to 4 textures. Such a low texturing efficiency may be the result of several factors working together, among which we think the most crucial are the insufficient bandwidth of the 128-bit memory bus or a too severe reduction of the size of the texture caches when the X800 was converted and truncated all around to become the X700.
The RADEON X700 XT takes a good start here, even leaving behind the RADEON X800 PRO with no textures thanks to the higher clock rate, but rolls back to the last position at rendering two or more textures.
Everything looks normal when we disable color writes: the RADEON X700 XT makes a good use of its high clock rate to outperform the RADEON 9800 XT, but of course loses to the 12-pipelined RADEON X800 PRO and to the NVIDIA GeForce 6600 GT which can write down as many as 16 Z-values per clock.
To check out the performance of the new graphics processor with pixel shaders we ran four programs: Marko Dolenc’s Fillrate Tester, Xbitmark, 3DMark2001 SE and 3DMark03. Here are the results of their running:
The RADEON X700 XT behaves just like a GPU with its specifications should do – slower than the RADEON X800, but faster than the RADEON 9800 XT. In three cases out of four the ratio of the results of the RADEON X700 XT to the NVIDIA GeForce 6600 GT roughly corresponds to the ratio of the clock rates of the two graphics processors. Only using the half precision the NVIDIA GeForce 6600 GT gains a heftier advantage over the competitor (The RADEON X700 XT is always using the 24-bit precision and cannot get any speedup from enabling 16-bit precision in pixel shaders).
Nothing new happens as we disable Z writes: the showings of the RADEON X700 XT and the others remain practically the same.
The GeForce 6600 GT accelerates rapidly as we disable color writes. That’s quite predictable, though.
Next goes our exclusive Xbitmark:
The RADEON X700 XT is advancing confidently through this test, especially processing complex shaders like Dot Product Bump Mapping + Specular + Reflection and Wood. It lost, however, to the GeForce 6600 GT in shaders that actively use textures, and this is yet another proof that the RADEON X700 XT has a low texturing efficiency.
The RADEON X700 XT is slightly better than the GeForce 6600 GT in this test as you see in 1024x768 and 1600x1200 resolutions. Well, it is rather the memory bandwidth than the speed of executing version 1.1 pixel shaders that limits the results here.
This scene uses version 1.4 shaders, and the X700 XT is noticeably slower than the GeForce 6600 GT, and just slightly slower than the top-end card of the last generation, the RADEON 9800 XT.
The RADEON X700 XT has average results executing version 2.0 pixel shaders: faster than the RADEON 9800 XT but slower than the GeForce 6600 GT. This quite well agrees with the results we got in Xbitmark.
Here it is – the triumph of the RADEON X700 XT. The three vertex pipelines of the GeForce 6600 GT are no match to the six vertex processors of the new RADEON.
It’s even better for the RADEON X700 XT in the analogous test from 3DMark03 as it gets practically the same result as the RADEON X800 PRO! Well, there’s nothing to wonder at – the two chips have the same number of vertex pipelines as well as the same frequencies.
The lights tests of 3DMark2001 SE also testify to the advantage of the RADEON X700 XT in processing geometry. The difference between the X800 PRO and the X700 XT is probably due to the different configurations of the test platforms.
The RADEON X700 XT and the RADEON X800 PRO are racing neck and neck in the test with eight light sources – the difference between them is negligible.
The speed of the vertex processors alone cannot win this test – a high texturing speed and a fast memory subsystem are also important. That’s why the X700 XT with eight pixel pipelines and a 128-bit memory bus fell behind the RADEON X800 PRO here.
In the EMBM relief imitation test the results of the graphics cards on the AMD64 platform are limited by the performance of the CPU and the system at large in low resolutions. For a more correct comparison take a look at 1600x1200 resolution where the RADEON X700 XT is equal to the GeForce 6600 GT.
When using the Dot3 method to imitate relief, the new GPU from ATI has the poorest result among all the graphics cards, although the gap is small. The low texturing efficiency of the RADEON X700 XT affected its performance but slightly, if at all.
The Ragtroll test is a good trial of the balance of the “CPU-driver-graphics card” system: the central processor is handling the physics of the scene, while the graphics processor is busy with the geometrical data. The results of the RADEON X700 XT are between those of the GeForce 6600 GT and the RADEON 9800 XT. Well, we tested the RADEON 9800 XT on a different platform, so you shouldn’t compare its results to the others directly.
The performance of the RADEON X700 XT in synthetic tests doesn’t allow us to make a final verdict about it. It is worse than the GeForce 6600 GT in simple multi-texturing tests and at executing pixel shaders that actively use textures. On the other hand, ATI’s new GPU has six vertex processors against three in the GeForce 6600 GT and outperforms the rival in tests where the speed of processing vertexes and executing vertex shaders is paramount. The RADEON X700 XT is also better than the GeForce 6600 GT when executing “calculations-heavy” pixel shaders. We might say that the new mainstream chips expose more clearly all the advantages and weaknesses of the respective top-end architectures, NV40 and R420.
The newcomer looks excellent against the top-end product of the previous generation, the RADEON 9800 XT. Despite its narrower 128-bit memory bus, the RADEON X700 XT delivers an overall higher performance, mostly due to the high frequencies of the graphics core and memory.
Of course, we will have a fuller view of the capabilities of the RADEON X700 XT only after getting it through gaming tests. We are going to dedicate a separate review specifically to that.