> The same exact GPU model and program versions must be used for compression and decompression. This makes the whole approach really difficult to be useful in

> The same exact GPU model and program versions must be used for compression and decompression. This makes the whole approach really difficult to be useful in practice. Basically whenever you have a method that relies on 100% reproducible float numbers. What are ways around that? Why do the quantized variants actually have the same problem?