I just hashed all the block headers in the blockchain file (did not prune orphans because meh) and found the average "count" multiplier for the internal buffer size in lots of 1000 blocks. Here is the results:
1000: 1035, 2000: 1014, 3000: 975, 4000: 955, 5000: 1022, 6000: 1082, 7000: 929, 8000: 947, 9000: 931,
10000: 898, 11000: 845, 12000: 916, 13000: 1015, 14000: 1053, 15000: 1016, 16000: 1050, 17000: 976, 18000: 933, 19000: 969,
20000: 937, 21000: 938, 22000: 868, 23000: 774, 24000: 781, 25000: 838, 26000: 699, 27000: 887, 28000: 909, 29000: 850,
30000: 899, 31000: 845, 32000: 791, 33000: 730, 34000: 770, 35000: 685, 36000: 675, 37000: 707, 38000: 681, 39000: 648,
40000: 570, 41000: 539, 42000: 530, 43000: 524, 44000: 449, 45000: 222, 46000: 297, 47000: 665, 48000: 434, 49000: 431,
50000: 511, 51000: 587, 52000: 504, 53000: 346, 54000: 421, 55000: 505, 56000: 862, 57000: 878, 58000: 814, 59000: 496,
60000: 278, 61000: 280, 62000: 311, 63000: 326, 64000: 240, 65000: 231, 66000: 263, 67000: 369, 68000: 336, 69000: 311,
70000: 287, 71000: 274, 72000: 330, 73000: 346, 74000: 317, 75000: 338, 76000: 344, 77000: 385, 78000: 381, 79000: 344,
80000: 374, 81000: 417, 82000: 424, 83000: 341, 84000: 368, 85000: 336, 86000: 384, 87000: 395, 88000: 385, 89000: 338,
90000: 352, 91000: 404, 92000: 386, 93000: 368, 94000: 323, 95000: 369, 96000: 337, 97000: 381, 98000: 430, 99000: 331,
100000: 372, 101000: 379, 102000: 369, 103000: 305, 104000: 339, 105000: 359, 106000: 305, 107000: 362, 108000: 311, 109000: 305,
110000: 323, 111000: 327, 112000: 318, 113000: 290, 114000: 370, 115000: 302, 116000: 283, 117000: 329, 118000: 327, 119000: 361,
120000: 345, 121000: 359, 122000: 344, 123000: 331, 124000: 349, 125000: 340, 126000: 291, 127000: 259, 128000: 295, 129000: 292,
130000: 283, 131000: 272, 132000: 276, 133000: 253, 134000: 258, 135000: 292, 136000: 269, 137000: 270
Up to block 17000 the internal multiplier looks like what I what I got from sequential data (basically random hashes) but from there it steadily counts down.
I'm not sure if this is evidence of a GPU miner or someone who is optimized for low memory usage on a CPU or something else. But the most likely explanation is that someone or many people are mining hard with a software that is optimized completely in favor of processing power over memory usage. I would expect that even if there is a GPU miner the hashes from PoB blocks would make the internal hash more random though.
What is the ratio of PoB blocks to PoW blocks?