A few days ago appeared information that in graphicaccelerators GeForce GTX 970 is no problem addressingvideopametta ie over 3.5 GB last 0.5 GB of 4 GB. The newly opened problem occurs in that last 0.5 GB provide much lowertransfer speed compared to other 3.5 GB, which can lead tovisible loss of productivity in very heavy duty load (4Kresolution and high visual settings). Over the weekend we sawa lot written on the internet, both from users and the NVIDIA.
Indeed, these results were achieved by the test Nai's Benchmark, in which is seen a serious loss of transfer usingthe last ~ 500 MB videobufer. After the initial assumptions thatGTX 970 can have serious problems under heavy load, it turned out that the real decline in productivity is completelynormal and is relevant to what is obtained with GTX 980 under the same load steps. The differences between the GTX 980(free from "problem"), and GTX 970 are within 2-3%, as seen from these tests:
Given the specificity of GTX 970 graphics accelerator is normal to assume that the biggest drop in performance compared to GTX 980 is something that is rather expected.GTX 970 has fewer SMM blocks and therefore has aprioritization in the work memory, so that the memory isdivided into two primary segments - 3.5 GB and 0.5 GB.Based on the serious decline in productivity in the programNai's Benchmark for the past 0.5 GB should see quite a big loss of performance modes in which they are tested shown above games. However, this does not happen.
At the time by NVIDIA do not give more explanations oradditional information on further actions of rectification indrivers or BIOS, or even replacement of graphics accelerators,but in the coming days will witness and more information fromthe manufacturer. According to the creator (ie himself Nai) program Nai's Benchmark, this software is unsuitable for checking video memory the way that many people exploit it. Further more added that for some reason the software is not properly test the entire amount of memory, in many cases after 3.5 GBreal test permeability of the PCI Express bus, not the actual video memory.After the weekend the information on this matter was quite and not what you character, finally the whole truth (as if?) Come out.The whole problem with the GTX 970 comes from the need tostop some SMM to achieve 13 active SMM and receiving 1664cores CUDA. Along this "cut" and having cut level rasterizing operators and past them is obtained and cut cache. Here's theword:
In the case shows the configuration of the graphics coreGM204 trimming it to the graphics accelerator GTX 970. The three SMM-disabled and visible, and one block of L2 cache, but no one more thing - the number 8 ROP-old, making a total of 56 ROP -a in GTX 970. Due to this specific effectivememory GTX 970 may be divided into 7/8 and 1/8 segmentsegment representing a total of 4 GB graphic buffer. NVIDIAdivide these 3.5 GB and 0.5 GB, respectively, of two parts - a quick segment and "slow" segment. The fact is that there is a256bit bus connection to memory, but the way the two segments communicate with memory actually leads to loss ofefficiency and those 256bit real underused. Technical GTX970 has a 256bit bus and offers 224 GB / s bandwidth formemory access, but the configuration of this access is notpossible to achieve the stated speeds. Here is the revisedtechnical specifications of the GTX 970:
Although the level of performance does not change, GTX 970is no longer looks like this graphics accelerator, which wasofficially announced and offered to customers. This leads toserious discontent in some users and even the creation of a petition for refund (money back) by NVIDIA for misleading advertising of graphic accelerator GTX 970. NVIDIA currentlystill at the level of recognition of the error as the main reasonhighlighted poor communication between the technical department and PR department of the company, which has led to "confusion" at-announced and advertised specifications ofGeForce GTX 970. For companies such as NVIDIA, however,such an explanation of the situation seems mildly ridiculous ...
Expect more on this topic in the next few days ...
