Beijing time, 2012 at 1:22 on March 2, the US West Coast time on March 6, 2012 on the 22nd, NVIDIA officially released the second generation of DirectX 11 graphics architecture products - code-named Kepler's GTX680. This is a set of small, low cost, low power consumption and high performance in one of the new products, it will be to the highest standard single-core flagship identity in the face of rival AMD released three months before the Tahiti architecture flagship product directly HD7970.
NVIDIA officially announced its latest generation of graphics products, GPU core architecture "Kepler," the GEFORCE GTX680. This is since the previous generation "Fermi" architecture of NVIDIA's GPU core architecture fully updated, also following a new generation flagship GTX580 graphics cards, NVIDIA is the combination of the most advanced technology, top products. GTX680 with a new 28-nanometer manufacturing process, GTX580 compared with the previous generation flagship NVIDIA, CUDA core number jumped to 1536, soared to the core frequency 1006MHz, be the first default core frequency exceeding 1GHz flagship graphics card. And surprisingly, the number of transistors and core area but had been reduced, much lower than the GTX580 and HD7970, so the supply of electricity need only to double 6Pin external PCI-E power, thermal design power consumption is only 195W, in recent years become the most energy-efficient flagship graphics card. Graphics memory, GTX680 uses 256Bit GDDR5, and a substantial increase in memory frequency, reached a record of 6000MHz. Furthermore, the support 4 can GEFORCE GTX680 screen display, on a display is also improved with the interface will be used for years to + MiniHDMI combination DVIx2 DVIx2 + HDMI + DP. Price, the official positioning of 3999, which generally 4000 to more cost-effective
GTX680 (9 Zhang)
than the HD7970 noun interpretation
NVIDIA and AMD GPU architecture relative term interpretation
1. NVIDIA prefer to use the name as a physicist named GPU architecture, from the beginning of the first generation of DX10 is called Tesla (Tesla), second generation Fermi (Fermi) , third generation Kepler (Kepler), the fourth generation of Maxwell (Maxwell), these scientists are outsize name, do not explain.
2. Kepler (Kepler) is NVIDIA's next generation GPU architecture code, the first one based on Kepler GPU architecture code-named GK104, GK104 based graphics card is the first core GeForce GTX 680.
3. AMD and NVIDIA previous GPU code is the same as letters + numbers, but from the beginning of the era of DX11 GPU code-named with the word, such as the HD5000 series architecture code-named EverGreen (evergreen), high-end HD5870 the core code-named Cypress (Cypress), midrange HD5770 core code-named Juniper (juniper);
HD6000 architecture code-named North Island (Northern Islands), high-end HD6970 core code-named Cayman (Cayman Islands), mid HD6870 core code-named Barts (St. Bart's, a Caribbean island);
HD7000 architecture code-named Sourth Island (Southern Islands), high-end HD7970 core code-named Tahiti (Tahiti), midrange HD7870 core code Pitcairn (a British south Pacific Islands), mainstream HD7770 core code Cape Verde (Cape Verde, Africa's southernmost island).
For graphics, the core frequency remains is to determine the key performance parameters, in the case of the same structure and the number of cores, the higher frequency means higher performance. This time, "Kepler" In addition to comprehensive changes in architecture, GTX680 the introduction of a new technology: GPU Boost, which means GPU acceleration. This dynamic overclocking technology is not new for us, Intel Turbo Boost used in the CPU (Turbo) technology, NVIDIA's GPU Boost technology this closely resembles, but is the first to use this technology in the GPU, GTX680 It is the first application of dynamic graphics acceleration technology.
GTX680 power consumption of only 195 watts of amazing, and NVIDIA graphics power that the previous high, heat is completely different, even lower than the HD7970, GTX680 done to improve performance but lower power consumption, phase than the HD7970, GTX680 only 3.54 billion transistors, while the 7970 has 4.3 billion transistors, more bits wide, GTX680 is 256Bit, 7970 is 384, memory aspect is not dominant, in order to compensate for this deficiency, NVIDIA will is the default memory GTX680 6008Mhz a high frequency (memory data transfer rate 6.0 Gbps), the final memory bandwidth still reaches 192GB / s, and with the previous generation of memory interface 384bit on theoretical performance GTX580 flat.
GPU and CPU design ideas of the art Boost Turbo Boost, Turbo Core technology is somewhat similar, but not the same. This CPU technology field is mainly used in multi-core CPU, the effect achieved is simply when the application software can not take advantage of multi-threaded operation, CPU will automatically be able to make "extra" cores are close to the closed state, and to improve "available" core frequency, thereby achieve the purpose of acceleration. TDP rule is the "reference value" CPU, the "off" core can bring down the TDP, and enhance the "available" to bring the core frequency will increase the TDP, but in the end no matter how to adjust the number of cores and enhance the work of frequency, and ultimately will not exceed the CPU TDP.
But the GTX680 and HD7970 does not win too many tests GTX680 only about 9 percent faster than the HD7970, 2012 June 22, AMD released a new flagship HD7970 GHZ, this card will not only frequency upgraded to 1G, and also joined the BOOST function, HD7970 GHZ although defeated GTX680 in performance, but it brings greater power consumption
may be due to GF110 / GF100 core power consumption control failure in the design of the NVIDIA GK104 core is very focused on performance per watt (performance per watt), which is what we usually refer to the performance / power ratio. TMSC said before the 28nm NVIDIA to achieve this goal provides a good foundation, the next step is architected to work hard.
The reason why the new generation of NVIDIA SM called SMX, really is because compared to the previous SMX SM larger changes. So SM specifically to do what?
SM is NVIDIA Unified GPU architecture of the core part inside the CUDA Core may be performed Pixel (pixel), Vertex (vertices), geometry shading (geometry shader), (physical) Physics and compute (operation) etc.; the texture unit can perform texture filtering (texture filtering) load / store and to read and save data to the work memory; Further, SFU (special function units, special function units) transcendental may process (a priori) and graphics interpolation (interpolation pattern) instructions; finally PolyMorph engine for processing vertex fetch (vertex acquisition), Tessellation (tessellation), viewport transform (viewpoint transformation), attribute setup (property), and a stream output (output stream).
Number of transistors: 3520000000
Graphics Technology: 28nm
stream The processor (sp): 1536 th
raster units: 32
core frequency: 1006MHz
memory frequency: 6008MHz
memory Bandwidth: 192GB / S
memory capacity: 2GB
memory Interface: 256bit
maximum resolution: 2560X1600
graphics power : 195W
power supply: 6Pin + 6Pin
GPU Boost: 1058MHz
DirectX version: DX11.1
cooling: cooling fan