We need videocards with lots of memory. Give me a 4080 with 192gb. I would be happy. We really need AMD to come up with new cards to wake up NVidia and start some fierce competition
It's not really feasible to scale GDDR-based designs that big. The 5090 is expected to have 32GB which probably means the workstation variant will have 64GB, but that's the limit of the conventional GPU memory architecture for now. HBM is fast and high capacity but prohibitively expensive, and LPDDR is cheap and high capacity but relatively slow, so there's no free lunch to be had.
What would it take to have a unified memory architecture to rival Apple's ? Is it theoretically possible with PC motherboards and GPUs that sit in card slots of some form?
> Is it theoretically possible with PC motherboards and GPUs that sit in card slots of some form?
As far as I'm aware, not with the speed that unified memory gets. The trace and path lengths alone put hard caps on what can be done in terms of signalling speed. But I'm not an expert, I'm recounting what I was told when I asked the same question! Perhaps the state of the art has improved in this space?
