HappySqurriel said:
| Squilliam said:
Well if they are using one of the two memory stacking technologies then they can have both high performance and more memory. They will either use DDR4 or GDDR5 and I suspect the former rather than the latter. They could go as wide as 512/1024 bits and get extremely good latency and power efficiency if they are willing to take a stab at it. Given the history of consoles using exotic/fast memory buses it wouldn't be unprecedented and quite justifiable as there are many algorythims which are memory bandwidth constrained which could be used to good effect.
|
Of course there are many things that could be done, but these usually result in other trade-offs being made. With more exotic memory architectures this usually results in the system being more difficult to develop for. In some cases it may not be that taxing (for example on-chip memory which tends to have a very specific and isolated use) but in other cases it can be problematic. Hypothetically speaking, a system could be built with memory on the GPU for screen buffers and texture caching, have dedicated GDDR5 memory for the GPU, some DDR4 memory for the CPU, and a lot of low performance DDR3 to hold lower priority data could give you the performance and cost you want but it would also be a pain in the ass to manage memory in a system like that.
|
GDDR5 is going to go end of life within two years and it is very expensive and uses a lot of power. It will go EOL because the alternatives are simply that much better than what you can get with the new technologies coming on board with hybrid memory cube (HMC) and high bandwidth memory (HBM). DDR3 is simply inadequate for high performance applications so that leaves either DDR4 or an exotic application of DDR4 (HMC/HBM) as the two good options for console memory. This kind of technology has already been deployed in a console environment because the PS Vita uses Wide IO which was done before the the standard was finalised and has a 1024 bit bus which says to me it is doable and I already understand why it is desirable.
Is it worth the money? Being able to put 8GB on a 512 bit bus is quite enticing isn't it? You can get the speed of GDDR5 with the capacity of DDR3 for lifetime cost of GDDR5 AND simplify the one or two lifetime process shrinks on the APU/CPU+GPU because the memory bus and I/O is off die. I would be shocked if we don't see HBM or HMC deployed on a next gen console.
