The top solution may have more than 7500.
This year we should get several new top-end video cards at once. Firstly, AMD will finally release an adapter based on the “big Navi”, and secondly, Nvidia should launch a new generation of Ampere 3D cards on the market.
We still know little reliably about all these devices, but today in the Geekbench database two different Nvidia adapters, which clearly belong to the new generation, were lit up right away.
We do not know their names, but the first has 108 CU computing units with a frequency of 1 GHz and 48 GB of memory, and the second has 118 CU units with a frequency of 1.1 GHz and 24 GB of memory. You can’t look at frequencies yet, and the gigantic amount of memory, at least in the first case, hints at a non-gaming video card. Perhaps in both cases we are talking about adapters of the Quadro family. Although the model with a large number of CU units has half the amount of memory and may well be the flagship of GeForce.
As for the number of blocks, if the new architecture retains the basic ratio of blocks to CUDA cores, then we are talking about GPUs with 6912 and 7552 CUDA cores, which seems incredible. For comparison, the GT102 GPU has 4,608 CUDA cores and even the monstrous GV100 has “only” 5,376 cores.
By the way, right there we can compare performance. The youngest of the new products in the OpenCL test scores 141 654 points, and the oldest – 184 096 points. For comparison, Titan RTX shows the result of 132 804 points, Tesla V100 – 154 606 points, and GV100 – 142 837 points.
That is, the results of the new adapters are not as impressive as the parameters of their GPU. However, the tests passed, most likely, non-serial samples, possibly with lower frequencies. For us, the first data on the configuration of the new GPUs is more important now.