Regarding the 8GB HBM2 rumour vs the possible 24GB GDDR6; it's worth noting it's not a straight bandwidth comparison of 512GB/s vs 672GB/s:
The HBM solution is filling 8GB @ 512GB/s which gives you bandwidth per GB of 64GB/s, whereas the GDDR solution is filling 24GB @ 672GB/s which gives a bandwidth per GB of 28GB/s. Each GB -- or whatever arbitrary unit you wish to use -- is being moved 2.3x faster.
Whether 8GB at any speed is enough capacity for bandwidth sensitive tasks is of course another question..?
PS5 | HBM2 8GB + DDR4 16GB | Combined UMA
Pros:
- Reduced CPU<>GPU Memory Contention
- HBM: 2.3x Bandwidth Per GB
- DDR: Access & Latency Better Suited to CPU
- Can default to a single memory pool using HBCC (or similar) while providing multiple layers of abstraction/management for further optimisation.
- Reduced physical footprint, power consumption and thermals which may allow for a slightly more performant CPU/GPU.
Cons:
- HBM: Reduced Overall Capacity for Bandwidth Sensitive Tasks
- DDR: Only Suited to Low Bandwidth Tasks
- Flexibility and ability to compete with a single 24GB UMA depends entirely on the effectiveness of HBCC (or similar) or optimisation by devs if bypassed.
- Increased I/O Complexity
16+16 would be my pipedream and a no brainer if possible! :p