NVIDIA GeForce RTX 4090 | torch 2.12.1+cu130
27B GDN: H=16 HV=48 K=128 V=128; 128 SMs
blocks = NV * nseq * HV, so BV=32 -> NV=4 and BV=16 -> NV=8

           cell  blk32  blk16  BV=32 us  BV=16 us  null us  speedup  bit-equal       cpu us      limit
------------------------------------------------------------------------------------------------------
     decode B=1    192    384     26.86     26.55    26.86    1.012x        YES  cpu=  37.16 HOST-BOUND
     decode B=2    384    768     28.93     28.69    28.96    1.008x        YES  cpu=  36.95 HOST-BOUND
     decode B=4    768   1536     33.50     32.13    33.47    1.043x        YES  cpu=  37.61 HOST-BOUND
     verify B=1    192    384     28.73     36.15    28.74    0.795x        YES  cpu=  37.17 HOST-BOUND
     verify B=2    384    768     50.31     49.86    50.29    1.009x        YES  cpu=  36.66 gpu
     verify B=4    768   1536    274.24    274.95   273.94    0.997x        YES  cpu=  37.24 gpu

null = BV=32 entered twice; a speedup inside the null's spread is noise.
