The inevitability of the announcement of new NVIDIA graphics solutions is confirmed by many analysts, but there is still little real evidence. As the latest, we can consider the results of testing the mysterious products of this brand in Geekbench, which allow us to talk about superiority over the Tesla V100 (Volta).
The results of testing two different NVIDIA products in Geekbench 5.0.2, which were obtained back in October and November last year, make sense to consider in the context of performance in the Open CL environment. The first product with 118 compute units and 24 GB of memory scored in this test . The second product with 108 compute units and 48 GB of memory scored . In the first case, the limiting frequency of the GPU was 1,11 GHz, in the second - 1,01 GHz.
It cannot be ruled out that in both cases, NVIDIA's next-generation GPUs, which are conventionally called Ampere, were tested. The number of CUDA cores could reach 7552 for the first GPU, and 6912 for the second. The Tesla V100 accelerator of the Volta generation, we recall, is content with 16 GB of HBM2 memory and 5120 CUDA cores, although there is also a modification with 32 GB of memory of a similar type.
Other proportions in Ampere memory sizes, if these are really new generation graphics processors, can be explained by the transition from a 4096-bit memory bus to a 6144-bit one. Moreover, the representatives of the site believe that the NVIDIA GA100 GPU in its full configuration will have 8192 CUDA cores, and its truncated version has been spotted in the Geekbench database. A proportional decrease in the number of active cores for geometrically large GPUs is inevitable, as the probability of chip defects becomes higher.
The clock speeds of the GPUs in question are certainly far from final, but even in this configuration they are faster than existing NVIDIA flagships by up to 19-40%, based on Geekbench results. It is unlikely that these NVIDIA products with HBM2 memory will find their way into consumer systems, so this comparison is appropriate to make mainly from the point of view of the alignment of forces in the server segment.
Source: 3dnews.ru