by Alexey Stepin , Yaroslav Lyssenko, Anton Shilov
01/25/2008 | 10:31 AM
It’s no secret now that the first generation of mainstream graphics processors with DirectX 10 support and of graphics cards based on such GPUs was far from a success, especially in terms of performance. Notwithstanding the innovative architecture, ATI’s Radeon HD 2600 and Nvidia’s GeForce 8600 series didn’t do well in games, often being outperformed by previous-generation solutions, which perhaps were not so technically advanced but were free from obvious bottlenecks such as a small number of execution subunits or insufficiently optimized drivers. Another performance-limiting factor of both solutions was a 128-bit memory bus that couldn’t feed the GPU enough data in modern games, especially when the gamer enabled FSAA and high display resolutions.
This simplification of the mainstream GPUs could be explained by the necessity to keep their manufacturing cost within reasonable limits with the tech process available at the time. The unified graphics architecture, even cut down as it was in the Nvidia G84 and ATI RV630, was still very complex to implement, the chips incorporating 289 and 390 million transistors, respectively, which was about the same amount of transistors as in the flagship DirectX 9 GPUs with ordinary, non-unified architecture. To remind you, the Nvidia G71, the heart of the GeForce 7900 family, consisted of 278 million transistors while the ATI R580, employed in the Radeon X1900/X1950 series, was made out of 384 million transistors. A memory bus broader than 128 bits could not be used due to the thin tech process, which reduced the cost of the GPUs.
Of course, we remember that GeForce FX and Radeon 9600 could not offer high performance in DirectX 9, either, yet we hadn’t expected the GeForce 8600 and ATI Radeon HD 2600 to be so poor in comparison with the previous-generation graphics cards (Radeon X1950 Pro). The developers obviously overdid it with the cost-reducing measurers.
Nvidia was the first to find a way out of the situation, announcing a new G92 graphics core on October 29, 2007. That GPU was installed on the GeForce 8800 GT series. Manufactured on 65nm tech process, the new chip proved to be a success. Despite the awesome complexity of the core (754 million transistors), its cost, area, heat dissipation and power consumption could be kept within reasonable limits. The G92 was just slightly inferior to the G80 in terms of functional subunits, but made up for that with increased clock rates and a 256-bit memory bus. The new card priced at only $259 would be just slightly slower than the $599 flagship GeForce 8800 GTX across most of our tests (see our article called From Extreme to Mainstream: Nvidia GeForce 8800 GT from Leadtek for details). The best representative of the other camp, ATI Radeon HD 2900 XT, was beaten mercilessly as it delivered lower performance at a much higher level of power consumption. So what did AMD do?
AMD, which owns ATI Technologies now, had to hurry up with an answer considering the precarious position of the company in the sector of consumer 3D graphics. Under AMD's management, ATI graphics product group's revenues dropped 40% in 2007 from a similar period in 2006 (check out this news story for details). ATI’s Radeon market share had shrunk from 27.6% at the moment of the announcement of the merger of AMD with ATI. So, AMD had to do something to get back on track as soon as possible. A quick announcement of the new RV670 GPU and a release of mainstream graphics cards based on it would have been just what was needed, especially as that GPU had been successfully tested back in August last year, but AMD decided to launch the new ATI Radeon HD series as part of the new gaming platform codenamed Spider. But that platform was delayed repeatedly as AMD Phenom processors could not reach the desired frequencies. The announcement was postponed to November 19, and AMD’s new processors could only reach a frequency of 2.30GHz by that time while the graphics department was still losing its profits. But let’s get back to the RV670, which was released eventually.
ATI Technologies used to be criticized for the confusion in its product nomenclature and this confusion continued even when ATI became part of AMD. RV670-based graphics cards were expected to come under the name of Radeon HD 2950 since the new chip represented an evolution of the R600. Instead, the nomenclature was completely changed and the new series was called Radeon HD 3800 as if based on a new-generation GPU. We guess the DirectX 10.1 support (see below) is not a solid reason for bestowing the status of new-generation solutions on the new graphics cards, but anyway.
The new ATI Radeon cards from AMD are called as follows:
Here, the first digit means that the card belongs to the third generation. The second digit denotes the family (we’ll surely see Radeon HD 3700 or 3900 in the future) and the last two digits denote the positioning of the product within the family. Thus, 70 is equivalent to the XT suffice and 50 is equivalent to the Pro suffix, but other variants are possible in the future. This naming system is somewhat more elegant and logical than the older one. At least it helps avoid the confusion of suffixes such as XT, XTX, Pro, GT, etc. But on the other hand, the RV670 doesn’t differ from the R600 as much as to be regarded as a new generation whereas the customer may perceive the change of the first digit in the name as an indication of a new-generation product with a new level of performance, which is not true. The ATI Radeon HD 3800 X2 with two GPUs on board is the only new card that can match the performance of a couple of ATI Radeon HD 2900 XT or a single Nvidia GeForce 8800 Ultra.
Now we’ll discuss the specifications of the new Radeon and of the new graphics cards based on it.
Click to enlarge
Click to enlarge
As the table shows, AMD has gone further than Nvidia in exploring new tech processes: the new chip is the world’s first 55nm GPU, which helped achieve a frequency of 700MHz, at least with the senior model, Radeon HD 3800. The developer also reduced the complexity of the chip, probably by optimizing the internal architecture and simplifying the topology of the ring-bus memory controller because the RV670 has the same amount of functional subunits as the R600. This is the first case for a long time when a new GPU from ATI incorporates fewer transistors than a same-class GPU from Nvidia. ATI wins technical superiority again but how do the new products compare with the GeForce 8800 GT?
The ATI Radeon HD 3800 exceeds the Nvidia GeForce 8800 GT in core frequency but has fewer functional subunits, particularly TMUs. As we know, the advantage in the number of execution subunits is not very important, not only because the ALUs work at a higher frequency in Nvidia’s chips: the VLIW architecture implemented in the Radeon HD requires driver optimizations for effective paralleling of shader code. Otherwise, if not all of the ALUs are loaded with work, the efficiency of the Radeon HD’s superscalar shader processors plummets suddenly. In the worst case, only one out of five ALUs that are part of each of the 64 such processors is really working. Ironically, AMD/ATI has the most problems with software optimization as it doesn’t have access to the code of games developed within the framework of Nvidia’s The Way It’s Meant to Be Played program until their official release. This often has a negative effect on the performance of AMD/ATI’s graphics solutions in games.
If you view the flowchart of the RV670 only, you won’t find any difference from the R600. The difference can be found on a deeper level, namely support for DirectX 10.1 (Shader Model 4.1) and a full-featured hardware video-processor UVD similar to the one in the ATI Radeon HD 2600. The features and operation of the UVD were discussed in our review titled The Look of Perfect: Analysis of Graphics Accelerators Performance during Media Playback. The DirectX 10.1 support available in the new chip deserves a separate section in this article.
Like in the R600, each shader processor of the RV670 consists of six subunits: five ALUs and one flow control subunit (branching, comparison, loops, subroutine calls). It also contains a set of general-purpose registers.
Four out of the five ALUs are simple, capable of executing one FP MAD instruction per clock cycle, and the fifth ALU can execute complex instructions like SIN, COS, LOG, EXP, etc. This architecture is highly flexible and scalable, but depends heavily on software optimizations. Although each Radeon HD core contains a special task dispatcher, its efficiency depends directly on the efficiency of the shader code compiler, which is part of the driver. The superscalar architecture offers its highest performance when all the ALUs are busy computing independent operations but it is hard to achieve that because in 3D applications many operations depend on the results of previous operations. That’s why Radeon HD GPUs require application-specific optimizations in the driver. Unfortunately, AMD has problems with that as it doesn’t have access to the innards of games participating in Nvidia’s The Way It’s Meant to Be Played program. They have to optimize the driver after the final release of the game. It is more difficult and, sometimes, not as successful as we might wish.
Perhaps ATI’s vision of the future of GPUs is indeed more progressive than Nvidia’s, but game developers are always oriented at the most popular architecture, not the most innovative one, and this architecture is currently represented by the scalar cores of the GeForce 8 series.
AMD has to cooperate more with game developers in order to overcome this problem. It also has to offer new high-speed graphics solutions in due time and reduce prices for its products when necessary. Such measures would make Radeon HD cards interesting for end-users and popular among gamers – the company’s reputation on the consumer 3D market would be restored then.
The texture and raster processors of the Radeon HD 3800 do not differ from those of the Radeon HD 2900 architecturally. They are complex devices we can only roughly compare with traditional TMUs and ROPs. The RV670 incorporates four large texture processors each of which contains:
As you see, the RV670 should have no problems with texture sampling but each processor cannot filter more than four texture samples per clock cycle, 16 samples in total. Clearly, it is the filtering of textures that is the bottleneck in the texture processors of the Radeon HD 2000/3000. There is only one filter unit per each two address units, so there is no talking about free tri-linear or anisotropic filtering. It is here that the RV670, like the R600, is inferior to Nvidia’s G92 and, especially, G80 GPU with its two filter units per each address unit. There are some improvements, though, but they are on a deeper level – we’ll discuss them in the DirectX 10.1 related section below.
Each raster processor of the RV670 contains:
So, it is roughly equivalent to 4 classic ROPs and can process four pixels per clock cycle for a total of 16 as there are four such processors in the chip. When working with the Z-buffer, i.e. with pixels that don’t contain color information, the performance is doubled for a total of 32 pixels per clock cycle. Like the texture modules, the RV670’s raster processors have been redesigned to support DirectX 10.1 capabilities.
AMD says the memory controller of the RV670 GPU has been improved to make more efficient use of the available memory bandwidth. The new chip is supposed to be no worse than the R600 in this respect. However, the width of the external memory bus has been reduced from 512 to 256 bits, and the width of the internal ring buses of the controller has been reduced from 1024 to 512 bits. More subtle changes, if any, are inaccessible for us and we can’t say if the RV670 really has an improved memory controller or the memory subsystem of the R600 is just utilized inefficiently. We are prone to believe the latter idea, however, because the Radeon HD 2900 XT doesn’t have any advantages in real applications from its 512-bit external memory bus with a bandwidth of over 100GB/s.
Anyway, the senior Radeon HD 3800 model is superior to the GeForce 8800 GT in terms of memory bandwidth due to its GDDR4 memory. We’ll see in our tests if this advantage is really called for in real-life applications but the new card has a certain reserve of performance here. The ATI Radeon HD 3850 is also better than the Nvidia GeForce 8800 GT 256MB in this parameter: Nvidia’s card has a memory frequency of 1400MHz, which means a peak bandwidth of 44.8GB/s only. For these cards this parameter is not critical, though, because it is the amount of memory, 256MB that is going to be the main bottleneck in games.
So, the ATI Radeon HD 3870 and 3850 seem to surpass the Nvidia GeForce 8800 GT from a technical standpoint, but Nvidia’s card is not so sensitive to application-specific driver optimizations and is backed up by Nvidia’s The Way It’s Meant to Be Played program. It is also obviously faster at texturing. AMD responds with pricing: the official price of the Radeon HD 3870 is set at $219, of the Radeon HD 3850, at $179, which is $40 and $20 lower than the official prices of the corresponding solutions from Nvidia. This makes the new cards appealing if they are available and provide high enough performance in games.
AMD’s graphics department is also rumored to be planning to cut the price of the Radeon HD 3870 from $219 to $199 and of the ATI Radeon HD 3850 from $199 to $159 in near future.
It’s not the first time in ATI Technologies’ history that it supports a standard that hasn’t yet become really widespread. For example, the ATI Radeon 8500 supported the Pixel Shader 1.4 specification, part of DirectX 8.1. That specification was more flexible than Pixel Shader 1.0/1.1/1.3 supported by Nvidia’s chips and helped achieve better visual effects, yet it didn’t really took off during the lifecycle of the Radeon 8000 series due to the limited support on the GPU part. The same story happened with Shader Model 2.0b supported by the ATI Radeon X700/X800/X850 series as well as with Shader Model 2.0a Nvidia promoted with its GeForce FX. DirectX 9 Shader Model 3.0 only became more or less popular after games began to be ported from the gaming consoles Microsoft Xbox 360 and Sony PlayStation 3, i.e. after the life term of GeForce 6 and early GeForce 7.
It’s all normal since the game developer always works with the most widespread standards. He isn’t interested in using capabilities provided by GPUs from one GPU maker, be it ATI or Nvidia. A new standard only becomes popular and widely used when it is supported by a large number of contemporary GPUs and when these GPUs deliver high enough performance for using the new features.
When Nvidia released the first DirectX 9.0c supporting GPU, NV40, the features provided by the new standard (Shader Model 3.0) were long unused, not only due to the lack of support from ATI Technologies but also due to the relatively low performance of the GeForce 6800 series. It was only with the third generation of SM3.0-compatible GPUs from Nvidia (GeForce 7600/7900) and with the first generation of such products from ATI (Radeon X1000) that the standard came to us for real: DirectX 9 Shader Model 3 was supported by the two leading GPUs developers whose products delivered enough performance for using it and also by the two leading gaming consoles.
Currently, there is a transition from DirectX 9.0c to DirectX 10 as Windows Vista is being promoted as a gaming platform. It’s going to take some time for DirectX 10 to become the dominating standard in the PC game industry, but AMD/ATI is already offering DirectX 10.1 in its new graphics core. Let’s see what perspectives this may have.
According to unofficial sources, DirectX 10.1 is going to be the first and only subset of DirectX 10. It will be officially added into Windows Vista as part of Service Pack 1 to be released in the first half of 2008. The main innovations in DirectX 10.1 include:
That’s quite a long list of innovations. The techniques for accelerating the global lighting of the scene are especially exciting as they can help improve the quality of lighting in games dramatically. But are we going to see these new capabilities of the RV670 in real games during the lifecycle of the ATI Radeon HD 3000 series? Obviously, not. First, they can only be utilized after the release of Vista Service Pack 1. Second, the current generation of Nvidia’s GPUs, even the latest G92, do not support cubic map arrays and is unlikely to be able to calculate the global scene lighting fast enough. The rest of DirectX 10.1 features do not look like a breakthrough, being either improvements on the appropriate features of DirectX 10 or standardizations like the obligatory support of 4x MSAA or standard subpixel masks for antialiasing.
So we are quite sure of an indifferent attitude of game developers towards the innovations at least in the first half of this year. Games with support for DirectX 10.1 can be expected to arrive no sooner than the second half of 2008 or even when the market is filled with DX10.1-compatible solutions from both GPU developers. It is also possible that the Radeon HD 3000 series just won’t allow using DX10.1 features due to low performance, and game developers will have to wait for the next, faster, generation of GPUs from AMD and Nvidia. There is a piece of good news, though. ATI says DirectX 10.1 is the first and only superset for DirectX 10 planned by Microsoft, so it will be used sooner or later until DirectX 11.
Like the competitor’s offer, the new ATI Radeon chip supports the PCI Express 2.0 interface that provides two times the bandwidth of PCI Express 1.0/1.1: 8GB/s in each direction as opposed to 4GB/s with the previous versions of the standard. The higher data-transfer rate is enabled only when supported by the chipset, of course. Otherwise, it is reduced to the bandwidth of PCI Express 1.0/1.1. This won’t have a big effect on performance in games because a bandwidth of 4GB/s is quite enough even if the graphics card carries at least 512 megabytes of onboard memory. And if the card has less memory, the increased bandwidth of the new PCI Express version won’t help much. The speed of the GPU’s access to system memory will anyway be much lower than to local memory that is clocked at a higher frequency and connected to the GPU with a 256-bit or broader bus. Like with Nvidia GeForce 8800 GT, we can expect some incompatibility with certain mainboards supporting PCI Express 1.0/1.0a. The lack of such problems is guaranteed for PCI Express 1.1 mainboard only.
The ATI RV670 is also the world’s first desktop processors that features the same power-saving technologies as are implemented in ATI’s chips intended for mobile applications. Particularly, ATI PowerPlay technology allows the new chip to flexibly adjust its frequencies and voltages and even disable unused blocks when the GPU is idle. As opposed to other desktop GPUs, the power-saving features are implemented in the RV670 on the hardware level, which ensures a quicker response to any change in the GPU load and avoid mistakes when determining the load type. AMD claims the 55nm tech process together with PowerPlay helped keep the power consumption of the new cards at 105W for the ATI Radeon HD 3870 and 95W for the ATI Radeon HD 3850. That’s a real breakthrough in comparison with the ATI Radeon HD 2900. We’ll try to check this out in our own tests.
Describing the ATI Radeon X1950 Pro we noted two CrossFire connectors, supposing that they could be used to build multi-processor systems consisting of more than two graphics cards. This supposition comes true now: the ATI Radeon HD 3800 series supports CrossFireX technology that allows uniting 3 or even 4 graphics cards into a single graphics subsystem. Such configurations can hardly become widely popular as not all systems can accommodate 3 or 4 graphics cards. Still, this may be an interesting feature for some users who could only build multiple-card subsystems out of GeForce 8800 GTX/Ultra so far. Theoretically, four Radeon HD 3870 cards in CrossFireX mode can deliver very high performance at a rather low, for such a complex system, power draw, about 400W. For comparison, a pair of Radeon HD 2900 XT cards consumes about 320W. The cost of four Radeon HD 3870 or of two Radeon HD 3980 X2 is going to be rather low, too, which should ensure CrossFireX systems a small but stable niche in the enthusiasts community especially as AMD promises the option of overclocking of such solutions and the support of up to 8 monitors simultaneously whereas Nvidia’s SLI can output to one monitor only in multi-GPU mode.
We’ve been describing the theoretical capabilities of the ATI Radeon HD 3800 series so far. Now it’s time to get to practical matters. We’ll tell you about the design features of the new solutions starting from the highest-performance model that has the index of 3870.
The ATI Radeon HD 3870 resembles the Radeon X1950 XTX as it uses a similar cooler and has the same dimensions. The PCB of the Radeon HD 3870 has a length of 23 centimeters (9 inches). This is almost 1 centimeter shorter than the Radeon HD 2900 (23.8cm).
Radeon HD 3870
Radeon HD 2900 XT
The wiring of the Radeon HD 3870 is much simpler than that of the Radeon HD 2900, which is not surprising considering the 256-bit memory bus and the less potent power circuit that was not designed for a power draw higher than 150W. With the dismantled elements of the cooling system, the PCB looks almost empty, at least from the front side.
Unlike the 7-phase power circuit of the Radeon HD 2900, the Radeon HD 3870 uses a two-phase circuit. By our estimates, the Radeon HD 3870 has a peak consumption of 80-85W or less, so this power circuit should be able to power the GPU although it may be an obstacle to overclocking the latter. Each phase contains four Infineon BSC 042N03LS MOSFETs rated for a current of 95A and featuring SuperSO8 packaging with advanced thermal characteristics. Like on the Radeon HD 2600 XT GDDR4, the main controller is an uPI Semiconductor uP6201AQ. The uP6101BSA chip is responsible for the memory chips. The power circuit of the Radeon HD 3870 uses one external power connector, a standard 6-pin PCI Express 1.0 plug with a load capacity of 75W. Considering the 55nm tech process, this connector can hardly be under a load higher than 35-40W.
The ATI Radeon HD 3870 carries GDDR4 memory from Samsung. These K4U52324QE-BC08 chips have a capacity of 512Mb (16Mbx32), a voltage of 1.8V, and a rated frequency of 1200 (2400) MHz. This means a good stability reserve since the default memory frequency of the Radeon HD 3870 is 1125 (2250) MHz. Theoretically, this memory should be overclockable to its rated frequency at least and perhaps even higher. The total amount of memory on board the Radeon HD 3870 is 512 megabytes.
At the default memory frequency, the bandwidth of the Radeon HD 3870’s memory subsystem is 72GB/s. This is higher than that of the Nvidia GeForce 8800 GT 512MB but lower than the memory bandwidth of the ATI Radeon HD 2900 XT 512MB (105.6GB/s). But as you know from our tests, memory bandwidth is not a bottleneck for AMD’s solutions. The Radeon HD 3870 should not feel the lack of memory bandwidth in games.
The RV670 chip strikes you with the sheer size of its core. Incorporating 666 million transistors, it is only 192 sq. mm large (13.7x14mm). This is a good illustration of what the new 55nm tech process can do. Besides an ATI logo and a manufacturing date, the core doesn’t carry valuable information. Our sample of the chip dates the 39th week of 2007 (September 23-29). Although the domains of Radeon HD chips work at 26 different frequencies, only the main frequency of the chip is declared. It is 775MHz for the Radeon HD 3870. Die packaging lacks a protective frame, making you be cautious when installing or uninstalling the cooler to avoid damaging the core.
There is no difference from the ATI Radeon HD 2900 XT in terms of configuration: the GPU contains 320 ALUs grouped into 64 superscalar execution blocks with 5 ALUs in each. Four ALUs in such a block can perform simple instructions of the MAD type (Multiply + Add) but the fifth one is more complex and can execute transcendental instructions like SIN, COS, LOG, EXP, etc. Besides that, each execution processor contains a branch unit for flow control instructions (comparisons, loops, subroutine calls, etc). The single difference from the R600 is the support for the extended DirectX 10.1 capabilities (Shader Model 4.1). As we’ve seen in tests, the efficiency of this architecture in real-life applications depends directly on the optimization of the shader code translator in the driver for the particular application with its specific game engine. Otherwise, the performance of the Radeon HD is low since most of the ALUs remain idle.
The RV670 graphics core also incorporates 4 texture processors equivalent to 16 TMUs, 4 rasterization processors (render back-ends) equivalent to 16 ROPs, and a Compositing Engine for hardware CrossFire. Like the all previous Radeon cards with the Compositing Engine integrated into the core, the ATI Radeon HD 3870 has two CrossFire connectors and can be used in a multi-GPU subsystem consisting of four such cards. So, the purpose of the second CrossFire connector is now perfectly clear.
The RV670 contains not only an integrated audio core for outputting digital audio over the HDMI interface but also a full-featured UVD processor that was missing in the R600 chip. Now the ATI Radeon HD 3800 series offers the same options of hardware decoding and processing of HD video in H.264 and VC-1 formats as the ATI Radeon HD 2600 that has been the best solution in this area. The card is claimed to support HDMI 1.3, but the audio core seems to have the same capabilities as in the Radeon HD 2900: 5.1 audio output, 16bit/48kHz for non-compressed streams in PCM format, and no support for Dolby TrueHD/DTS-HD. You still need a special DVI-I → HDMI adapter, included with the card, to output audio over HDMI. This approach is handier than if the HDMI port were installed right on the card because it provokes no problems for owners of dual-monitor configurations.
As opposed to the ATI Radeon HD 2900 with its 9-pin connector, the ATI Radeon HD 3870 has a 7-pin mini-DIN connector for analog video output. The new series doesn’t carry a VIVO chip on board and even provides no place for installing such a chip. That’s quite logical because it is indeed an anachronism to support analog video input in S-Video/Composite format in 2008. Like on all modern graphics cards, the DVI-I ports support dual-link mode for resolutions up to 2560x1600.
The cooling system that used to be installed on the Radeon X1950 XTX is deservedly regarded as one of the best reference coolers combing excellent noise characteristics with high cooling performance. The developer drew upon that old design to create the cooler for the Radeon HD 3870.
It even doesn’t seem to differ from the Radeon X1950 XTX’s cooler at all, but its design is simpler because the RV670 generates less heat than the R580+.
Particularly, the heatsink is a monolithic thing milled out of a chunk of copper and lacking any advanced features like heat pipes or something. Such a heatsink was used in the cooler of GeForce 6800 in both copper and aluminum versions as well as in the legendary coolers Thermaltake Volcano 7+ and Cooler Master Aero 7+.
The monolithic heatsink is free from the main drawback of composite heatsinks, the additional thermal resistance in the place of connection of the ribs with the sole. The heatsink of the Radeon HD 3870’s cooler is rather small, not requiring a heat pipe for a uniform distribution of heat. It is cooled with a 7.2W centrifugal fan (12V, 0.6A) with a 75mm impeller.
The developer even saw to such small things as the slits under the fan: it is there that the power circuit heatsink is located. So, the hot air is exhausted by the fan through these slits avoiding a “heat bag” that might lead to overheat of the MOSFETs. Passing through the main heatsink and taking heat off the GPU, the hot air is exhausted out of the system case to keep better thermal conditions inside the PC. The memory chips are cooled by a separate L-shaped copper heatsink with good ribbing. That’s a proper solution for chips clocked at 1125 (2250) MHz.
Traditional dark-gray thermal grease is used between the heatsink and the GPU die. There is just the necessary amount of it here, unlike on the GeForce 8800 GT. The other components of the cooling system use elastic rubber-like thermal pads. The heatsinks on the memory and load-bearing transistors are fastened to the card separately from the main heatsink. The latter uses a four-pole fastening mechanism with a metallic back-plate and spring-loaded screws. This mechanism prevents the PCB from bending and the heatsink from misaligning. The profiled casing the fan resides in is attached to the main heatsink with six small screws. It is rather light, being no burden for the PCB.
By our estimate, the cooling system of the Radeon HD 3870 should be able to cope with 80-90 watts of heat at low noise level – that’s the amount of power the RV670 should dissipate when clocked at 775MHz. This thought-through cooler design is overall better than the reference cooler of the GeForce 8800 GT. Its only drawback is the dual-slot form-factor, but it’s a normal thing for every advanced graphics card today.
The PCB of the less expensive and less advanced RV670-based graphics cards is 1.5-2 millimeters longer than the PCB of the faster model. Its overall component layout resembles the ATI Radeon HD 3870, but there are a few differences:
For some reason the developer used a different component layout for the power circuit of the junior model. This cannot be explained by the lower power consumption of the Radeon HD 3850 as it uses the same two-phase circuit with four power transistors Infineon BSC 042N03LS in each phase and with an uP6201AQ chip as the controller. The uP6101BSA chip is responsible for memory. It’s clear the power circuit of the ATI Radeon HD 3850 can yield as much power as the one of the senior model although its load can hardly be more than 60 watts. Anyway, the card carries an external 6-pin PCI Express 1.0 connector, too. The other differences in the PCB wiring are due to the use of a different memory type: GDDR3 instead of GDDR4.
The reference sample of Radeon HD 3850 carries eight memory chips (Samsung K4J55323QI-BJ11, 256Mb, 8Mbx32, 1.9V). These chips are rated for a frequency of 900 (1800) MHz but they are clocked by the card at 830 (1660) MHz. There is some room for overclocking, but you can hardly achieve a frequency of 1000 (2000) with regular overclocking methods. Each chip has a capacity of 256Mb and a 32-bit bus, so the Radeon HD 3850 has a total of 256MB of memory accessed across a 256-bit bus. The memory bandwidth at the reference frequency is 53.1GB/s, which is lower than that of the senior model but higher than that of the GeForce 8800 GT 256MB. This should be quite enough considering that the Radeon HD 3850 is obviously not meant for high display resolutions.
The marking of the graphics core is just as vague as that of the Radeon HD 3870 core. We could only learn the manufacturing date from it: it is the 43rd week of 2007 (October 21-27), Taiwan. All the subunits of the GPU of the Radeon HD 3850 card are clocked at 670MHz. Die packaging has no protective frame.
There are no differences in the GPU configuration between the Radeon HD 3870 and HD 3850 except for memory controller settings for the use of GDDR4 and GDDR3 chips, respectively. The core incorporates 320 active ALUs grouped into 64 superscalar execution blocks with 5 ALUs and one branch unit in each. It also contains four texture processors equivalent to 16 TMUs, 4 rasterization processors (ATI calls them render back-ends) roughly equivalent to 16 ROPs, and a Compositing Engine for hardware CrossFire. Like the senior model, the ATI Radeon HD 3850 has two CrossFire connectors and can be used in a multi-GPU subsystem consisting of four such cards.
The card offers a standard selection of connectors: two dual-link DVI-I ports with support for DHCP and a display resolution of 2560x1600 pixels, and a universal 7-pin mini-DIN port responsible for analog video output in S-Video, Composite or YPbPr format. The card lacks VIVO functionality, like the Radeon HD 3870, offering no place for an appropriate chip. The more required feature for today is Audio-over-HDMI and it is supported fully by means of a DVI-I → HDMI adapter. Combining the best video-processor UVD, Audio-over-HDMI, the single-slot cooler form-factor and low power consumption, the ATI Radeon HD 3850 is perhaps the best graphics card for a modern multimedia system for processing HD content in VC-1 and H.264 formats.
The cooler of the ATI Radeon HD 3850 has a single-slot form-factor. That’s why it is more sophisticated than the cooler of the Radeon HD 3870.
A dual-slot form-factor doesn’t impose strict limitations on the type of the fan or the size of the heatsink and the developer can vary both to achieve the optimal combination of thermal and noise characteristics. It is not so with single-slot coolers: the developer has to resort to various tricks and compromises to remain within the limits of the form-factor while achieving the desired cooling efficiency. Sometimes silence has to be sacrificed even. An example of such a compromise is the reference cooler installed on the GeForce 8800 GT. The new cooler of the Radeon HD 3850 resembles it with its design and component layout.
Don’t be confused with the color of the metal the base of the cooler is molded from: it is not copper but aluminum, which we checked out by scratching the base. A flat heat pipe is pressed into the groove at the top to distribute the heat in the heatsink, and there is a copper piece where the cooler touches the GPU. Consisting of aluminum plates, the heatsink has a much larger total area than the heatsink of the GeForce 8800 GT’s cooler and is installed upright rather than at an angle. The air flow is driven away from the mounting bracket by means of special grooves in the cooler casing.
We don’t think it’s an ideal component layout, though. We guess the fan should have been instead placed near the DVI-I connectors, changing the direction of air flow to the opposite. In this case the hot air from the main heatsink could additionally cool the load-bearing components of the power circuit because the base is shaped like an additional needle-like heatsink there. Well, at an expected power draw of 50-60W, this cooler should cope with cooling the Radeon HD 3850 just fine. We can recall Nvidia who dared to use a similar component layout with a smaller heatsink to cool a graphics card with a power draw of 80W – the company and the users are still suffering from that attempt.
The cooler is equipped with a flat centrifugal fan of a rather small diameter with a 4-pin connection, which indicates PWM-based regulation of speed. Its noise parameters are questionable, but we expect it to rotate at low speed most of them due to the lower heat dissipation of the RV670 chip clocked at 670MHz and to the rather large heatsink.
There are the same thermal interfaces where the cooler contacts the GPU, memory chips and load transistors of the power circuit as in the cooling system of the Radeon HD 3870: dark-gray thermal grease for the GPU and elastic yellow thermal pads for the rest of the chips. Besides the four poles with spring-loaded screws and a metallic back-plate the cooler is fastened to the PCB with six additional screws, four near the memory chips and two near the MOSFETs. This prevents the cooler from misaligning and damaging the graphics core.
The cooler installed on the Radeon HD 3850 looks good overall. Perhaps it is not ideal (we do think that the position of the fan and the direction of its air flow is not quite correct), but this shouldn’t be a problem considering the lower heat dissipation of the chip in comparison with the Radeon HD 3870 especially as this cooler employs a more advanced heatsink that in the GeForce 8800 GT’s cooler.
We measured the level of noise produced by the reference coolers of the ATI Radeon HD 3870 and HD 3850 cards with a Velleman DVM1326 digital sound-level meter using A-curve weighing. The level of ambient noise in our lab was 36dBA and the level of noise at a distance of 1 meter from the working testbed with a passively cooled graphics card inside was 43dBA. We got the following results:
New ATI Radeon HD 3800 reference cooling systems showed their real best. The top model remained very quite no matter what mode it was operating in, and the younger model stood out just a little bit against the background of other system components only in 3D mode. Maximum graphics processor temperature hit 90ºC and 70-75ºC respectively. As you can see, ATI Radeon HD 3870 warms up quite significantly, however, there is no cause for concern here: the card never lost stability. Although, if you are not comfortable with this number you can use the latest RivaTuner version and improve the cooling efficiency by increasing the fan rotation speed. However, in this case you will have to put up with higher level of generated noise. Be careful, because extreme lowering of the fan rotation speed may cause the card to overheat and die.
As far as overclocking is concerned, our experiments turned out pretty successful. The maximum frequencies our ATI Radeon HD 3870 was working stably at equaled 850MHz GPU and 1200 (2400) MHz memory.
Despite our expectations, both cards from the new Radeon HD series had no compatibility issues with mainboards that support an earlier version of PCI Express than 1.1. This made it possible to test their power consumption using a special testbed with a modified Intel Desktop Board D925XCV. The measurement was performed with a digital multimeter Velleman DVM850BL (0.5% accuracy).
In 3D mode the cards were loaded by the first SM3.0/HDR test from the 3DMark06 suite running in a loop at 1600x1200 with 4x FSAA and 16x AF. The Peak 2D mode was emulated by means of the 2D Transparent Windows test from PCMark05.
Click to enlarge
The results of our measurements in 3D mode almost ideally coincide with our predictions abut the peak power consumption of both Radeon HD 3800 models. The numbers are normal considering the traditionally hot temper of ATI’s GPUs. It is only with the help of 55nm tech process that AMD managed to keep the power draw of its new GPUs within reasonable limits. By the way, the Nvidia G92 has similar results while having more transistors and 65nm tech process. Note also the economical behavior of the new cards in idle and 2D modes – ATI’s PowerPlay technology seems to do its job well.
Unlike Nvidia’s solutions, ATI’s Radeon HD 3870/3850 load the internal +12V line more than then external one. The load is not higher than 45W even in the hardest mode, which is far below the limit set in the PCI Express 1.0/1.1 specification: 75W for the slot and 75W for each external connector. The +3.3V line has a low load in every mode – 1.6-1.7W.
The comparative theoretical and gaming performance analysis of the ATI Radeon HD 3870 and 3850 was carried out on the following standard test platform:
According to our traditional testing methodology, the drivers were set up to provide the highest possible quality texture filtering and to minimize the influence of software optimizations that both – AMD/ATI and Nvidia – enabled by default. Also, to ensure maximum image quality, we enabled transparent texture filtering options: Adaptive Anti-Aliasing/Multi-sampling for ATI Catalyst and Antialiasing – Transparency: Multisampling for Nvidia ForceWare. As a result, our ATI and Nvidia driver settings looked as follows:
For our tests we used the following games and benchmarks:
First-Person 3D Shooters
Third-Person 3D Shooters
We selected the highest possible level of detail in each game using standard tools provided by the game itself from the gaming menu to each user. The games configuration files weren’t modified in any way. The only exception was Enemy Territory: Quake Wars game where we disabled the built-in fps rate limitation locked at 30fps.
Games supporting DirectX 10 were tested in this particular mode. With a few exceptions, the tests were performed in the following standard resolutions: 1280x1024/960, 1600x1200 and 1920x1200. If the game didn’t support 16:10 display format, we set the last resolution to 1920x1440. We used “eye candy” mode everywhere, where it was possible without disabling the HDR/Shader Model 3.0/Shader Model 4.0. Namely, we ran the tests with enabled anisotropic filtering 16x as well as MSAA 4x antialiasing. We enabled them from the game’s menu. If this was not possible, we forced them using the appropriate driver settings of ATI Catalyst and Nvidia ForceWare drivers
Performance was measured with the games’ own tools and the original demos were recorded if possible. Otherwise, the performance was measured manually with Fraps utility version 2.9.1; some of the results were averaged down to ensure better precision and smaller differences. We also measured the minimum speed of the cards where possible.
ATI Radeon HD 3870 and HD 3850 will be competing against the following solutions:
Since the new ATI Radeon HD 3800 family uses a new GPU, we have also performed a number of theoretical tests to determine RV670 performance difference from R600. We used the following benchmarks:
We are going to start the performance analysis with the discussion of results obtained in these particular tests.
The ATI Radeon HD 3870 has about the same results as the Radeon HD 2900 XT in the fill rate test. They only differ much when processing one texture, which is easily explained by the difference in their GPU clock rates. The addition of a second and more textures makes the gap between them much shorter, which may be an indication of some fundamental bottleneck in the architecture. The junior model is predictably slower than the senior one, the gap being the largest when they are processing a single texture again.
Interestingly, the Nvidia GeForce 8800 GTX has the biggest advantage over the GeForce 8800 GT when processing the Z-buffer as well as with one or two textures. When there are more textures, the TMU architecture of the G92 chip proves to be more efficient that that of the G80.
The ATI Radeon HD 3870 and ATI Radeon HD 2900 XT behave identically in this test. The RV670-based solution enjoys but a small advantage which is proportional to the difference in the clock rates of the two cards. The Radeon HD 3850 is slower than the two mentioned cards but faster than the Nvidia GeForce 8800 GT except in the PS 2.0 Per Pixel test.
The results of the latter card are interesting, too, as they do not exactly copy the results of the GeForce 8800 GTX when it comes to simple shaders. We know that the G80 and G92 have identical shader processors, so the difference must be due to texture lookups because it is the TMU architecture that differentiates Nvidia’s new high-performance GPU the most from the older one.
The ATI Radeon HD 3870 behaves like a copy of the Radeon HD 2900 XT, taking the difference in their clock rates into account, but there are serious discrepancies as well: the new card is much slower than its predecessor when executing dynamic branching shaders. Such shaders make use of complex math1ematics with loops and conditional jumps but have few texture lookups. Anyway, we can only explain such a big difference by a lower performance of the memory subsystem of the newer card. Perhaps the explanation lies somewhere deeper but we cannot find it basing on the available information about the architecture of the new GPU.
The pixel shader performance test from 3DMark06 shows nothing extraordinary. The ATI Radeon HD 3870 is ahead of the Radeon HD 2900 XT just as expected. Interestingly, the Radeon HD 3850 is almost as fast as the R600-based cards despite its lower GPU frequency. Since this test mostly measures the speed of texture lookups and the efficiency of the memory controller and caches, we can suppose that the RV670’s memory controller has indeed been seriously redesigned and the size of the caches has been increased.
ATI Radeon HD GPUs have always been strong at geometry, and the RV670 is no exception, becoming the new leader in the field. The reason of this victory is simple: the GPU can use all of its 320 ALUs for processing geometry whereas Nvidia’s chips just do not have such resources. Well, tests of this type have little to do with real applications in which shader processors are mostly busy with computations of quite a different kind.
This test is rather simple except that the total amount of vertexes in it is over 4 million. There is only one light source. The outcome is predictable: the ATI Radeon HD family enjoys a confident victory, not differing much among themselves.
The Complex Vertex Shader test is far more interesting: the performance of both Radeon HD 3800 models is declining as the display resolution grows up. We can’t explain it by a lack of memory bandwidth because the Nvidia GeForce 8800 GT doesn’t behave that way but has a lower memory bandwidth than the Radeon HD 3870. The speed of the TMUs is not an answer, either, or the Radeon HD 2900 XT would slow down, too. So, this must be the result of some deeper changes in the RV670 architecture we don’t know anything about.
The Shader Particles test shows how efficiently the physical model of a particle system is processed using pixel shaders. The model is visualized using vertex texturing.
The ATI Radeon HD 3800 series is surprisingly slow. Perhaps they feel the lack of memory bandwidth because the Nvidia GeForce 8800 GT 512MB, which is inferior to the GeForce 8800 GTX in this respect, behaves that way, too.
The Perlin Noise is a complex test using a shader with 447 math1ematics instructions and 48 texture lookups. Thus, it is a good test of the overall efficiency of the given GPU.
The ATI Radeon HD 3870 as about as good as the ATI Radeon HD 2900 XT at math1ematical computations and texturing. These cards have identical results. The ATI Radeon HD 3850 has a lower core frequency, yet keeps on the same level with the Nvidia GeForce 8800 GTX.
The results of the ATI Radeon HD 3870 series in synthetic tests provoke no questions except for complex vertex shaders, which may be due to architectural differences between the RV670 and R600 or to the specifics of the Complex Vertex Shader test from 3DMark06. We expect the ATI Radeon HD 3870 and Radeon HD 2900 XT to be roughly equal in real-life applications, the former having an advantage where a really high memory bandwidth is required and the performance is not limited by other factors. Let’s see if our supposition is true now.
This game doesn’t support display resolutions of 16:10 format, so we use a resolution of 1920x1440 pixels (4:3 format) instead of 1920x1200 for it.
The ATI Radeon HD 3870 performs up to our expectations that it would deliver about the same performance as the Radeon HD 2900 XT 512MB. It is even a little bit faster due to the higher core frequencies. As you can see, the narrower, 256-bit, memory bus has no effect on the new card’s performance even at 1920x1440.
Alas, the RV670-based cards cannot rival Nvidia’s GeForce 8800 GT, yet they ensure a high level of comfort at resolutions up to 1600x1200/1680x1050. Their minimum speed is never below a critical level even at 1920x1200. Considering the low price of both models, they seem to be quite appealing. Let’s see if it’s true for other games as well.
BioShock doesn’t support FSAA when running in Windows Vista’s DirectX 10 environment. That’s why we benchmarked the cards without FSAA.
The ATI Radeon HD 3800 series sets no records, yet doesn’t look like a loser, either. Particularly, the senior model is no worse than the GeForce 8800 GTS 320MB and even beats it at 1920x1200, the minimum speed being always high. The junior model’s average frame rate of 45fps with a minimum of 27fps at 1600x1200 seems to be quite enough for normal play. At a price of $200, this is a real gift for every gamer who cannot afford a top-end graphics card.
ATI’s Catalyst seems to manage memory more effectively than Nvidia’s ForceWare: the ATI Radeon HD 3870 is only slower than the GeForce 8800 GT 512MB at 1280x1024 and then goes ahead, being only second to the Nvidia GeForce 8800 GTX. This impressive result has little practical worth, unfortunately, due to very low performance. The GeForce 8800 GTX cannot even provide 25fps at 1280x1024 if you use the highest graphics quality settings together with FSAA.
Alas, application-specific driver optimizations are still a weak point of graphics cards from the former ATI Technologies. Although the ATI Radeon HD 3870 is no worse than its predecessor, it is not enough to provide a comfortable frame rate even at 1280x1024. You can try to play at such settings, but you are not safeguarded against slowdowns below 23-24fps. You can increase the speed of the game on an ATI Radeon HD 3870 by disabling FSAA, but this won’t do for a Radeon HD 3850.
The game being too hard at its Very High level of detail, we benchmarked the cards without FSAA to get a more playable speed.
The 256-bit memory bus doesn’t constrain the ATI Radeon HD 3870 even in such a difficult test as Crysis. Moreover, the senior model of the new family is slightly ahead of the ATI Radeon HD 2900 XT 512MB not at low but at high resolutions. This doesn’t mean the game is playable at the highest quality settings, though. The Nvidia GeForce 8800 GTX is unable to do that, even. On the other hand, the ATI Radeon HD 3870 is at least as good as the ATI Radeon HD 2900 XT 512MB here. The cheaper ATI Radeon HD 3850 is 15-20% slower than the senior model depending on display resolution.
The frame rate is fixed at 30fps in this game as this is the rate at which the physical model is being updated at the server. Thus, this 30fps speed is the required minimum for playing the game.
The ATI Radeon HD 3870 is unable to match the Nvidia GeForce 8800 GT 512MB at 1280x1024, the gap being as large as 27%, but it catches up with the opponent at 1600x1200! Perhaps this is the consequence of the improvements in the memory controller which help it cope better with MegaTexture technology. The ATI Radeon HD 2900 XT 512MB doesn’t show such outstanding results, at least. So, the senior ATI Radeon HD 3800 is just as good for this game as the more expensive and rare-to-find GeForce 8800 GT 512MB. That’s another point in favor of AMD’s graphics department.
The ATI Radeon HD 3850 looks good, too, providing the required 30+fps even at 1600x1200/1650x1050 and being almost as fast as the noisy and non-economical Radeon HD 2900 XT 512MB. The new card just doesn’t have enough of local memory at higher resolutions, but it does offer a solid performance at lower resolutions, which makes it an excellent choice for people who can’t spend much money for their graphics card.
Like Battlefield 2142, this game does not support resolutions of 16:10 format. So, we use 1920x1440 (4:3 format) instead of 1920x1200 in this test.
The ATI Radeon HD 3800 series is too slow for you to run this game on them at resolutions above 1280x1024. The latter resolution is perfectly playable, though, and it is at this resolution that there is the biggest gap between the two models: 17% in favour of the ATI Radeon HD 3870.
You can only play Episode Two at 1280x1024 on the ATI Radeon HD 3870/3850 and on the ATI Radeon HD 2900 XT 512MB unless you disable FSAA. The average frame rate is barely above the playable limit at the higher resolutions. Anyway, this is much better than what you could have with the previous generation of mainstream graphics cards such as ATI Radeon HD 2600 XT GDDR4 and Nvidia GeForce 8600 GT/GTS.
The game doesn’t support FSAA when you enable the dynamic lighting model, but loses much of its visual appeal with the static model. This is the reason why we benchmarked the cards in S.T.A.L.K.E.R. using anisotropic filtering only.
The ATI Radeon HD is still plagued with performance-related problems in this game but the minimum speed of the ATI Radeon HD 3870 is considerably higher than that of the Radeon HD 2900 XT 512MB at 1280x1024. That’s not yet enough for comfortable play, but there are no catastrophic slowdowns, either.
Every way of software-based boost of ATI Radeon HD’s performance seems to have been exhausted in this game – for classic single-processor cards. But judging by the latest test of Crossfire technology, we expect the dual-processor version of ATI Radeon HD 3800 with two RV670 chips is going to achieve acceptable results in S.T.A.L.K.E.R.
Forcing FSAA from the graphics card’s driver doesn’t produce any effect as yet. Perhaps the FSAA support will be added by means of a patch.
The ATI Radeon HD 3870 is second after the Nvidia GeForce 8800 GTX which makes it the best choice for playing Unreal Tournament 3 considering the price factor. The less advanced Radeon HD 3850 is not as successful as the rest of the participating graphics cards, but only at a resolution of 1920x1200 that is hardly important for users who buy $159 graphics cards. Even so, the average frame rate of the junior Radeon HD 3800 is over 50fps, which is quite enough for comfortable play.
This is a second test, after F.E.A.R. Perseus Mandate, when the ATI Radeon HD 3870 is slower than its R600-based predecessor. The gap is smaller in Lost Planet, though, at only 1.5fps. The ATI Radeon HD 3850 proves it once again that 256 megabytes of local graphics memory is not enough for modern games.
Generally speaking, this game has too high system requirements at its High level of detail which do not show up visually, however.
The problem with the minimum speed of ATI Radeon HD cards has been finally solved, so they can now be used for playing this game. The ATI Radeon HD 3800 is fast at resolutions up to 1600x1200 while the junior model has an acceptable minimum speed at 1280x1024 only. Here, the GeForce 8800 GT looks preferable but you shouldn’t forget about the price factor: Nvidia’s card is more expensive and not so easily available as AMD’s new solutions.
The average frame rate of the ATI Radeon HD 3870 is as high as that of the more expensive GeForce 8800 GT 512MB, but its minimum speed is much lower, probably due to its having fewer TMUs. Well, neither of these cards can deliver acceptable performance in this game as it has too high system requirements. The less advanced cards are out of play as well, having an average frame rate of less than 25fps.
The current version of the game doesn’t support FSAA, so we performed the test with anisotropic filtering only.
If you are choosing a graphics card to play Gothic 3, you should strike the Radeon HD 3850 out from you list due to its very low performance even at 1280x1024. The ATI Radeon HD 3870, on the contrary, looks good: it is rather cheap, noiseless, widely available, and offers good performance at resolutions up to 1600x1200. It won’t suit for large monitors, of course. If you’ve got a large monitor, you should search among GeForce 8800 GT cards or wait for dual-processors solutions from AMD.
The game loses much of its visual appeal without HDR. Although some gamers argue that point, we think TES IV looks best with enabled FP HDR and test it in this mode.
The new Radeon HD 3800 series performs perfectly fine in outdoor and indoor scenarios of this game. The top model provides comfortable gaming performance in all resolutions, although it yields to Nvidia GeForce 8800 GT 512MB, while the younger model loses to Nvidia GeForce 8800 GTS 320MB only in outdoor scenarios.
The new add-on to Company of Heroes is tested in DirectX 10 mode only since it provides the highest quality of the visuals.
The ATI Radeon HD 3870 has a higher minimum of speed than its predecessor and almost equals the result of the Nvidia GeForce 8800 GTS 512MB. You can try to play at 1280x1024 with FSAA turned on although there will be little playing comfort then. So, we don’t recommend to enable 4x FSAA in this game on such graphics cards as Nvidia GeForce 8800 GT and ATI Radeon HD 3870.
The game having a frame rate limiter, you should consider the minimum speed of the cards in the first place.
The ATI Radeon HD 3850 is the only graphics card here that cannot provide a minimum speed higher than 20fps at resolutions of 1600x1200 and higher. The more advanced Radeon HD 3870 is no worse than the monstrous GeForce 8800 GTX in this test.
Although the ATI Radeon HD 3870 is rather far ahead of the ATI Radeon HD 2900 XT 512MB, the overall performance level of AMD/ATI’s solutions is still very low. They don’t suit for playing World in Conflict, at least with the current version of the Catalyst driver.
The ATI Radeon HD 3870 scores a mere 57 points more than the ATI Radeon HD 2900 XT 512MB but that’s enough for it to claim absolute superiority in 3DMark05. The ATI Radeon HD 3850 in its turn scores only 105 points less than the Nvidia GeForce 8800 GT 512MB, which is a brilliant performance for a $159 product.
The individual tests we ran at higher display resolutions and with enabled 4x FSAA produce a different picture, of course, yet the Radeon HD 3800 is never slower than the ATI Radeon HD 2900 XT 512MB and even leaves it far behind in the second test that puts a stress on processing vertex shaders. The junior model is a loser just as you could expect, yet it is about as fast as the Nvidia GeForce 8800 GTS 320MB in a few cases.
The ATI Radeon HD 3870 doesn’t win the more complex 3DMark06 and even cannot outperform the ATI Radeon HD 2900 XT 512MB. The ATI Radeon HD 3850 is more successful, beating the Nvidia GeForce 8800 GTS 320MB by as many as 656 points.
Both ATI Radeon HD 3800 models behaved predictably later: a rather modest result in the SM2.0 group of tests and a very good one in the SM3.0/HDR tests which are more favorable for the Radeon HD architecture. The senior model is a mere 100 points behind the Nvidia GeForce 8800 GTX in the latter case. The Radeon HD 2900 XT 512MB is somewhat faster due to its more advanced memory subsystem with a 512-bit bus.
The ATI Radeon HD 3800 series has low results in the first test as it cannot match Nvidia’s solutions in terms of fill rate. It also feels a shortage of memory bandwidth as is indicated by the higher results of the ATI Radeon HD 2900 XT 512MB. The RV670-based cards feel more confident in the second test although the senior model is inferior to its predecessor. There is no competition with Nvidia’s GeForce 8800 GT/GTX as the Radeons are over 20% behind.
The SM3.0/HDR tests are quite a different story: the ATI Radeon HD 3850 equals the performance of the Nvidia GeForce 8800 GTS 320MB, which is quite good for its price. The more advanced Radeon HD 3870 comes closer to the GeForce 8800 GT 512MB: the gap is 10% in the first test and 7-8% in the second test.
ATI’s new graphics cards Radeon HD 3870 and HD 3850 seem to be very appealing for the end-user considering the price reduction AMD can go for. However, the overall trend shows the weakening position of AMD’s graphics department.
The senior model of the new series cannot unfortunately match the Nvidia GeForce 8800 GT 512MB but we had not expected it to. These products come from different market niches, being officially priced at $219 and $259, respectively. AMD’s card has something to beat the Nvidia GeForce 8800 GT 256MB with – 512 megabytes of graphics memory. The Radeon HD 3870 is overall fast for its price, providing as much performance in popular games as the previous flagship card from AMD but being twice more economical and absolutely silent.
The ATI Radeon HD 3850 in its turn looks good against the Nvidia GeForce 8600 GTS as well as GeForce 8800 GT 256MB. It provides higher performance than the former due to a more advanced GPU and a 256-bit memory bus and costs less than the latter ($159 against $179) while being more available.
So the ATI Radeon HD 3850 is the best choice in its price range just because it has virtually no opponents. As opposed to the GeForce 8600 GTS, AMD’s solution features high enough performance for playing modern games at highest settings, even though at a resolution of 1280x1024, which is in fact the main one for owners of mainstream graphics cards.
To present the results of the ATI Radeon HD 3800 series in a more visual way we created a summary table that shows in what games AMD’ graphics cards can give you a playable frame rate and in what they cannot. We made up our opinion basing on the average and minimum speed. If these were not lower than 30-35fps and 20-25fps, respectively, the given display resolution was considered as uncomfortable for normal play at the highest graphics quality settings.
As you see, the ATI Radeon HD 3870 is indeed often good even at 1920x1200 whereas the practical application of the ATI Radeon HD 3850 is mostly limited to 1280x1024. In games that all of the resolutions are marked as insufficient performance the gaming comfort is often not provided even by the Nvidia GeForce 8800 GT 512MB. These are Call of Juarez in DirectX 10 mode, Crysis, Lost Planet: Extreme Condition and World in Conflict. Unfortunately, the old problem of the ATI Radeon HD series – the dependence of performance on application-specific driver optimizations – persists still in such games as S.T.A.L.K.E.R.: Shadow of Chernobyl, Hellgate: London and Company of Heroes: Opposing Fronts.
Besides good performance, the ATI Radeon HD 3800 series features a number of other strong points. Particularly, it is economical, especially in 2D mode, has a quiet but high-performance reference cooler, incorporates an audio core, and is compatible with PCI Express 1.0a. The practical worth of the DirectX 10.1 support is hard to determine until the new standard becomes truly widespread, yet this is kind of a reserve for the future anyway.
So, the ATI Radeon HD 3870 and HD 3850 are interesting products with low price and good performance but they won’t suit for everyone. We want to remind you that the choice of the graphics card depends only on your personal gaming preferences – we can’t choose for you. The best we can do is to offer you detailed information about the performance of the particular product in popular games so that you can make a reasoned and, hopefully, right choice.