If you take apart a gpu and look at the die you can see they basically frankenstein 1 or 2 HBMs with the GPU die. Xilinx is doing this too in their latest line

If you take apart a gpu and look at the die you can see they basically frankenstein 1 or 2 HBMs with the GPU die. Xilinx is doing this too in their latest line of VU33 - VU38P HBM chips too (4-8GB on chip, 460 GB/s). Any density upgrades for the GPU will also reflect for FPGAs, but we're physically limited in terms of how many HBM dies you can slap on :(