We don’t technically need to retool to go back, HBM could be used by a normal system by having it directly built into the CPU package. Intel Xeon Max CPUs already did this. Of course there are some significant limitations to doing so, and because they were enterprise chips they were horrifically expensive, but it is possible to run them without DRAM.

Especially for something like laptops where there isn’t any significant need for replaceable CPU/ram it could be a cheaper option than retooling. This is especially true if people want to run a local LLM for inference on that laptop, since that kind of setup would be great for it.