From my understanding of what I have read, the HBM2 specification is designed to use the same memory but at different power saving states etc., so you get the same memory running at either 800 MHz or 1000 MHz with or with out voltage increases depending on the specific memory, and the final product designer can choose what memory bin best suites their power and performance envelope. Further reading also indicated that at the time of the Vega 10 cards launch, when HBM2 was still in its early stages, Hynix were the originally preferred vendor but were unable to produce memory to the required specification in time. As a result, AMD had to use Samsung's (Flarebolt?) memory in their reference cards after it was the first to market with in early 2016, which although followed the HBM2 specification of different power state specs, were designed to run at a higher than originally intended voltage of 1.35 volts for 2 Gbps HBM2. Given the apparent fact that all Samsung HBM2 memory seems to overclock to the same general numbers and carry the same part number as Samsung Flarebolt, the general conclusion appears to be that these were all designed to run at 1 GHz. However, given the heat intensive environment of graphics cards, and perhaps the difficulties in the early stages of HBM2, AMD must have decided to run these at 945 MHz to keep yields high. This would also explain why the majority will overclock to 1050 to 1100 MHz once the voltage is increased to 1.35 V, well at least that is my take from this whole situation. Different vendor bios may also contain differences in memory timings, so the 1025 MHz HBM2 max may yet change depending on the bios.
