NVIDIA-NVIDIA,GeForce GTX GeForce GTX 680 preview 680,GK104,Kepler preview
Following After yesterday 3DCenter.org, the BSN website has always been good at depth analysis also released today the news about NVIDIA 28nm process Kepler architecture GK104 core: part of the site yesterday reported substantially the same, but there are some nuances worth pondering. (On yesterday's rumors PCINLIFE administrator cho's uncharacteristically did not throw a \"Fake\" or \"makes up\" left, in addition to a variety of sources has been indirectly implied the majority of the details are true)
First the core frequency and Shader frequencies on NVIDIA unified this long reports. Kepler framework, a set of SM CUDA Core amount of contrast Fermi structure GF114 was promoted to double, from 48 to 96. Memory controller for the 256bit interface, these yesterday 3DCenter message.
Next, let's look at the core schematic of GK104:
Big eyes and a look the and GF110 similar, but a closer look can be found in the SP number jumped from 512 to 1536. BSN said initially the NVIDIA consider the GK104 listed early name is the GeForce GTX 660, but 99% of the officially listed the name of GeForce GTX 680, the remaining 1% for the GTX 660/670. Price and before different mouth Charlie said: GeForce GTX 680 located in the GeForce GTX 560 Ti 2GB/GeForce GTX 560 Ti 448 Core / GeForce GTX 570's replacement, the initial price of 349-399 dollars. The performance is certainly better than the previous generation flagship GeForce GTX 580 to be much higher, the most important thing is the GK104 able to not only get rid of $ 449 Radeon HD 7950 3GB, $ 549 Radeon HD 7 970 also mention the news that 3DMark 11 X-mode score may about 3000, the night blast down 1,000 yuan seems about to appear ... (Hint: 1536/1024 or 768/512)
Next is the framework for analysis of GK104, Kepler, times a set of SM 96 CUDA Core, a GPC (Graphics Processing Clusters) containing 4 SM 384 CUDA the Core. The GK104 containing 1536 (4GPC), 2304 CUDA the Core is not difficult to speculate top GK110 may contain 6GPC. The memory controller may still be the low end entry level 64bit, midrange mainstream 128bit, high-end 256bit the top 384/512bit-- should be used in HPC, or a GK110 the GeForce GTX 690. Kepler's mobile graphics core is already long been taped to complete CES, Samsung has demonstrated a laptop equipped with the most low-end (96CUDA the Core). (Driver identify frequency / specifications)
Tahiti, XT, contrast and AMD GK104 on the number of ROP and TMU - are 32 and 128, but the memory controller indicators continued the \"traditional\" as good as AMD, 2GB 256bit GDDR5 memory on on Raeon HD 7 970/7950 of 3GB. However, the advantage on price and performance I think we will also force the rival product price.
And GF110, GK104 fact, for half of the single-precision, double precision floating point performance in two versions - Production Quadro and Tesla, the theoretical production GeForce quarter.
Comparison with the the GK104 core frequency is tentatively set at 950MHz, BSN, said the sources said that NVIDIA's laboratory initially available in two versions: 772MHz for the same the GeForce GTX 580, 925MHz for the same Tahiti, XT, for comparison, the actual shipping version of the core frequency may be between 950-1000MHz. Of course, the little blue pill for like AMD, NVIDIA is also stepping up to adjust the BIOS, fight for the public version of the the GK104 core frequency will stand on a 1GHz. If NVIDIA can not accomplish this goal, it will be the first time open to non-public version / AIC overclocked version.
Memory area, a tentative frequency of 1.25GHz equivalent 5GHz, relatively GF100/110 of up to 25%. This regard, AMD, GDDR5 memory / controller to be regarded as the standard-setting, one of AMD's GPU engineers to improve the GDDR5 memory has advantages also not unusual. Calculated by the previously described frequency, GK104 with 160GB / s memory bandwidth, compared to the GTX 580's 192.4GB / s, a slight decrease compared to the expected substitution products GTX, 560 of Ti and GTX, 560 Ti the GF110 core version / GTX, 570 128.27GB / s and 152GB / s have raised.
All these specifications, the single-precision floating-point capability 2.9-3.05TFLOPS, double-precision floating-point for the 486-500GFLOPS. Version of Quadro and Tesla will have to unlock the double-precision floating-point computing power - that is, around 1.5-1.6TFLOPS and GTX 580 single-precision floating point capability is quite significantly higher than the Fermi core Quadro and Tesla.
BSN concluded by stating that users will not wait too long, NVIDIA is poised to start Kepler's pre-sale / Distribution operations (including mid-GK107 - GeForce GTX 650, etc.), used against AMD will market action February 15 release of \"Cape Verde\", CeBIT in Hanover, Germany, the Radeon HD 7700 series, and in March, posted on Pitcairn Radeon HD 7800 series. Perhaps next month on the market will be able to see Kepler entity.
