As the first GeForce 1000 graphics card based on the Pascal architecture, all eyes are on the GTX 1080. At the time of the official launch, Nvidia shared only a scant few high-level specifications for the card. Fortunately, someone has leaked the full specifications and slide deck so we won’t have to wait till May 18th for the NDA to lift. This answers many of the remaining questions about the card though real-world performance is still an unknown.
As know from earlier, the GTX 1080 features 2560 CUDA cores, but we now know that it will also feature 160 TMUs and 64 ROPs. This is the same numbers of ROPs as the 980 but the TMU count is closer to the 980Ti. There are 4 GPC (Graphics Processing Cluster) each with 5 Streaming Multiprocessors which have 2 blocks of 64 CUDA cores each. The total die size is 314 mm2 with 7.2 billion transistors, nearly twice as dense as Maxwell.
GP104 also brings Nvidia’s fourth-generation delta color compression. This improved compression delivers 1.2x greater effective bandwidth which is why the card can get away with 320GB/s of bandwidth. There is also a new GPU Boost 3.0 which appears to be more aggressive and better track theoretical performance. Lastly, we have HDR, something AMD has also been talking about with Polaris.
Overall, we’re getting a better picture of what the GTX 1080 looks like. I suspect GP100 may have better performance per clock than GP104 and it will be interesting to see how the card compares to Maxwell. In only 3 days, we may just find out.
