GeForce GTX 580 architecture features
The new GeForce GTX 580 is not something revolutionary. The main task that the company's engineers faced was to correct previously made errors, as well as increase productivity for a more significant lead over the main competitor AMD.
Unfortunately, the first accelerator based on the new architecture, the GeForce GTX 480, never received all the computing capabilities that NVIDIA engineers had laid down in GPU GF100. The fact is that in the serial GTX 480 models one of the 16 streaming multiprocessors was disabled, as a result, despite its status, the new top model lost 32 CUDA cores, four texture units and one Polymorph Engine. Most likely, NVIDIA engineers took this step solely because of production problems that had been haunting the company for a long time in preparation for the release of Fermi-based solutions. However, now this is not so important, because the new GPU The GF110 has all 16 SMs at its disposal, meaning 512 full-fledged CUDA cores, 64 TMUs, and 16 Polymorph Engine modules. Moreover, NVIDIA engineers have made a number of modifications that allow the GeForce GTX 580 to perform full-speed texture filtering with floating point and 16-bit precision (FP16), which makes it possible to increase performance in applications that require a large amount of work with textures.
The fact is that the texturing speed on the GeForce GTX 480 drops by half when using 64-bit (fp16x4) formats with non-trivial filtering. For example, when texturing from fp16x4 with bilinear filtering, the speed drops to 30 pixels per clock (PPC), compared to 60 PPC for 32-bit formats. On the GeForce GTX 580, texturing from such formats occurs at “full speed”, i.e. Both 32-bit and 64-bit formats with bilinear filtering operate at 64 PPC.
In addition, the GeForce GTX 580 has improved the Z-cull rejection unit for invisible surfaces. In the case of the GF100 GPU, when using certain Z-buffer formats, the efficiency of the Z-cull block decreased. With the introduction of new Z-cull depth buffer representation formats in the GF110, this problem has been resolved. In general, the increase in performance of the GeForce GTX 580 relative to the previous flagship GeForce GTX 480 will largely be due to an increase in the number of active functional units and an increase in operating frequencies, and to a lesser extent due to architectural improvements associated with strengthening the Z-Cull unit, as well as an increase in the speed of FP16 texture filtering .
Here a completely justified question arises: what to do with the power consumption of the new accelerators, because it is no secret that one of the significant shortcomings of the previous flagship - GeForce GTX 480 was far from the best ratio of performance per watt of consumed energy. In the case of the new GeForce GTX 580, this problem was also solved, since NVIDIA engineers carried out a deep redesign GPU, using transistors with lower leakage currents in critical areas. This allowed not only to raise the operating frequencies of the GTX 580 relative to the values of the previous top, but also to reduce the power consumption of the new flagship in comparison with the GeForce GTX 480.
Like most other modern NVIDIA accelerators, the GeForce GTX 580 fully supports NVIDIA PhysX and CUDA technologies, which allow the use of GPU resources for various types of parallel calculations, for example, PhysX hardware acceleration is used to enable a number of physical special effects in games based on the PhysX engine. Unfortunately, at the moment there are not many games that support NVIDIA PhysX technology, although there are some very outstanding projects among them, for example, Batman Arkham Asylum, Mafia II and Metro 2033.
It is also impossible not to mention NVIDIA Surround and NVIDIA 3DVision Surround technologies. In essence, NVIDIA Surround is a somewhat limited analogue of Eyefinity technology, which AMD announced back in the fall of 2009. The essence of NVIDIA Surround is that by connecting three monitors to a computer, the user can combine them into a single workspace, which is treated by the system as one large monitor. Limitations of NVIDIA Surround are associated with the impossibility of combining more than three monitors into a single space, as well as the need to use at least a pair of NVIDIA video cards in SLI mode to activate NVIDIA Surround technology. As for NVIDIA 3DVision Surround, this technology allows you to get an additional effect of presence in virtual worlds through the joint use of NVIDIA 3D glasses and three 120 Hz monitors.
Despite the fact that the new GTX 580 is noticeably different from the cooler of the reference version of the GTX 480, the design of the cooling system of the new product shows characteristic features characteristic of many Hi-End NVIDIA products. As you have probably already noticed, the new cooling system of the GTX 580 is devoid of heat pipes that were present in the GeForce GTX 480. This time the cooling system of the top NVIDIA accelerator is based on an evaporation chamber. The graphics processor is in contact with a platform, directly above which there is a sealed compartment - it contains a liquid with a low boiling point. As the temperature rises, the liquid changes its state of aggregation and turns into gas, carrying heat with it to the top of the cooling system. Cooling slightly, the gas condenses and turns back into liquid, then the cycle repeats. In turn, the turbine drives cool air through the heated fins of the radiator and removes heat outside the housing. Video memory, as well as elements of the power system, are in contact with the metal base of the cooler through special thermal pads.
The GF110 graphics processor has six power phases, which are controlled by the CHiL Semiconductor CHL8266 controller, exactly the same one was installed on the reference versions of the GeForce GTX 480. The memory power supply is made according to a 1+1 scheme (1 Vdd phase + 1 Vddq phase), controlled by the Anpec APW7066 controller. A similar controller was once used on the GeForce 9800 GTX. In this case, one two-channel controller controls both video memory voltages (Vdd and Vddq), one channel for each voltage.
The next innovation was the mechanism of hardware monitoring of current and voltage on the main 12 V power lines of the video card (6 and 8-pin power connectors, as well as the PCI-Express slot). If the video card is supplied with excess power, exceeding the values permitted by the specifications, the driver automatically reduces the clock speed of the accelerator to prevent its failure. The manufacturer especially notes the fact that the modern NVIDIA driver automatically reduces the accelerator's operating frequencies by 50% when running stress testing utilities such as OCCT or FurMark. According to NVIDIA engineers, such programs create a critical load on the video card, while real applications never create such a load. This is why the decision was made to artificially reduce frequencies using drivers. It is also reported that end users will not be able to remove this restriction. Unfortunately, at the moment none of the utilities are able to track the decrease in frequencies caused by the activation of the protection mechanism in the drivers, but after the release of the corresponding updates, the user will be able to notice all the changes in the operation of the GeForce GTX 580 video cards.
Frequency GPU of the new product is 772 MHz (1544 MHz for the shader domain, operating at double the core frequency). This is approximately 10% higher than the operating frequency of the GF100 video card GeForce GTX 480 (700/1401 MHz). Taking into account the architectural improvements GPU The GTX 580's performance advantage may be even more significant.
| Characteristics of NVIDIA GeForce GTX 580 |
|
It is easy to see that before the release of the GeForce GTX 580, NVIDIA carried out serious and painstaking work to eliminate the obvious shortcomings of the GeForce GTX 480. Reduced power consumption compared to its predecessor, less heating GPU and a fairly quiet cooling system - all this despite the fact that the new product turned out to be noticeably faster than the GTX 480, moreover, in many applications the GTX 580 turns out to be more productive than the Radeon HD 5970.