Modern VRAM (GDDDR) runs at very high frequencies and very low latency. This makes the wiring between the RAM chips and GPU very tricky (they need to be short and the same length), so slots aren't an option. There were actually VRAM slots in some 90s GPUs.

VRAM runs at high frequencies, delivers insane throughput, but the latency is not great. GPUs don’t need low memory latency. They have very high degree of parallelism. GPU cores switch to other threads instead of waiting for data from memory.

Here’s an interesting system with AMD Zen2 CPU and GDDR6 memory, salvaged from XBox: https://www.tomshardware.com/news/4800s-xbox-chip-shows-us-w... As you see, high latency of GDDR6 memory ruins the performance of CPU-running code.

Not enough bandwidth.

My current GPU has 484 GB/second memory bandwidth. It would require 7 channels of DDR5-8400 memory (the fastest one currently defined by these specs), and GPUs aren’t yet large enough to fit 7 slots of SO-DIMM.

If you look at a modern card you can see how the ram is designed - randomly around the actual GPU core - to get it as close and consistent as possible.

Some chips are moving towards having the ram on the same die package as the actual GPU as one integrated chip.

Because that reduces bandwidth and occupies more space