Like with the GeForce GTX 580 and 750 cards, the GPU structure has not changed. The GF114 is in fact an all-active GF104 optimized for high clock rates at reasonable power consumption.
The two processing clusters include four multiprocessors each. Each multiprocessor consists of 48 stream cores for a total of 384 and is serviced by eight texture-processing units. The total of TMUs is 64 in the new chip. The TMU architecture has remained intact since the GF104. They can perform full-speed filtering of FP16 as well as INT8 textures. Textures in FP32 format are processed at only one fourth of the full speed. Like in the GF104, the L2 cache is 512 kilobytes large.
Each multiprocessor also incorporates a PolyMorph engine, making the GF114 superior to any AMD solution in terms of geometry processing and tessellation performance. Even the two third-generation tessellation units in AMD’s Cayman processor can hardly match the eight PolyMorph engines of the GF114. The new chip’s rasterization subsystem has remained the same with 32 raster back-ends. It is directly connected to the memory subsystem, so the latter still includes four 64-bit controllers. The memory bus connected the GPU to the local graphics memory is 256 bits wide.
There are no notable innovations in the multimedia department, but they are hardly necessary. The GF104 could already do everything modern users might want, offering hardware support for HD video in H.264 and VC-1 formats and Protected Audio Path for outputting multichannel audio in Dolby TrueHD and DTS-HD Master Audio formats. The single advantage AMD can boast in this field is that their solutions support hardware decoding of DivX.
Overall, the GF114 looks a well-balanced solution with reasonable functionality. It supports all modern visual and multimedia technologies but lacks rarely required capabilities such as an opportunity to connect six monitors simultaneously, for example. The new mainstream GPU from Nvidia carries on the ideas found in its predecessor GF104. Let’s now see how the GF114-based graphics card is positioned among its predecessors, opponents and elder cousins.
Just like the GF110 is the “corrected” version of the GF100, the new GF114 is what the GF104 should have been from the beginning. Nvidia’s developers have managed to make the new GPU stable in its full configuration with eight active multiprocessors for a total of 384 ALUs and at higher clock rates than those of the GF104. Let’s compare:
Thus, the new GeForce GTX 560 Ti is to the GeForce GTX 460 1GB as the GeForce GTX 580/570 are to the GeForce GTX 480/470. There is nothing wrong in having a new version of an old GPU, though. The Fermi architecture is quite a good one, except for its texture-mapping subsystem which is inferior to that of the AMD solutions.
So, we can see that the GPU clock rates have grown up: over 800 MHz for the main domain and over 1.6 GHz for the shader domain. This is an achievement for Nvidia, considering the architectural features of their solutions. Such frequencies could only be conquered by individual GeForce cards with factory overclocking. AMD has something to worry about now because even the main domain of the GF114 works at a higher clock rate than the whole Radeon HD 6950 GPU. Besides, the latter has only 352 VLIW4 processors as opposed to the GF114’s 384 scalar ALUs.
AMD’s graphics architecture is also known to deliver its maximum performance not in all cases. Its new Cayman incarnation is also devoid of a dedicated ALU for complex instructions, so such instructions have to be performed by four simple ALUs, which is not efficient. Thus, in some cases the Radeon HD 6950 is going to be much slower than the GeForce GTX 569 Ti with the latter’s 384 stream processors clocked at over 1600 MHz. We’ll see in our practical tests whether this scenario is likely or not. As for the Radeon HD 6870, it is inferior to the GeForce GTX 560 Ti in every parameter, including texture-mapping performance.
The memory specs of the GeForce GTX 560 Ti have improved over its predecessor’s, yet the peak bandwidth of 128 GBps is far from impressive compared to that of AMD’s Cayman and even Barts solutions. It is unclear why Nvidia could not have increased the memory frequency, at least up to 1125 (4500) MHz. The GF110-based solutions have the excuse of their wider bus. It is not easy to ensure stable operation of memory chips at high frequencies when the memory bus is 320 or 384 bits wide. The GF114, on the contrary, has a 256-bit bus, so the GeForce GTX 560 Ti might have been equipped with faster GDDR5 chips. So, the memory bandwidth is not a strong point of the new card. It is inferior in this respect even to the Radeon HD 6870, let alone the Radeon HD 6950.
The TMU subsystem has ceased to be a bottleneck typical of the Fermi architecture with the unlocking of the eighth multiprocessor and the increase of the GPU clock rate to 820 MHz. With 64 active TMUs, the peak texture-mapping performance is 52.6 gigatexels per second, which is even higher than that of the senior GeForce 500 series products. The AMD Cayman with 96 TMUs (88 in the Radeon HD 6950 model) can do even more, yet that’s hardly called for in today’s applications. Tests suggest that the GeForce GTX 580 and GTX 570 do not feel a lack of texture-mapping speed whereas the new GeForce GTX 560 Ti is even better in this respect.
The rasterization performance of the new card is also high because of the increased GPU frequencies. The peak fill rate of the new GeForce GTX 560 Ti is slightly lower than that of the Radeon HD 6870 which has a core clock rate of 900 MHz but higher compared to the architecturally more advanced Radeon HD 6950. The new card from Nvidia shouldn't feel a lack of rasterization speed even at high resolutions with enabled full-screen antialiasing.
The rest of the features have remained the same as in the new card's predecessor. Thanks to optimizations in the GF114 chip, its TDP is only 10 watts higher than that of the GF104. Again, the new card is actually an improved version of the GeForce GTX 460 1GB with eight stream multiprocessors and higher GPU and memory frequencies. The predecessor still remains a good product and an excellent choice for gamers who don't want to spend more than $250 for their graphics card. The successor has inherited its best features and offers them at only $20 more.
Coming at a recommended price of $249, the new GeForce GTX 560 Ti fits in between the Radeon HD 6870 and Radeon HD 6950 and we guess it has a chance to outperform both these opponents. Let's now take a look at the new card. We've got a reference GeForce GTX 560 Ti from Nvidia.