Why is nobody making a desktop unified memory platform with upgradeable components? This is 100% possible and apple proved unified memory can be pretty dang good.
>>109416554And how upgradeable are thos Apple machines?
>>109416554>Why is nobody making a desktop unified memory platform with upgradeable components?Allocate it a few braincells. I'm sure you will find out why.
You need some fast channels for that, hence why they tend to solder the memory. Haven't looked into LPCAMM2, but maybe it could be your solution?
>>109416554amd apus have had unified memory since forever and no it has no benefit even apple is very stingey with their memory controllers and bandwidth
>>109416554Unified memory slows down the GPU. GPU needs memory physically very close and running at different latency/profiles per die in order to max out efficiency. CPU memory on the other hand is a single addressable pool, without granularity.That's why GPU bandwidths peak in the 1TB/s range, while CPU have 10-20x less bandwidth.In practice, if your workload is mostly bottlenecked by CPU <-> GPU transfers, then unified memory may be good (depending on how it's implemented). But anything that's GPU compute intensive will be drastically slowed down because GPUs need very fast memory.
>>109416554Socketed memory is slower. Unified really needs soldered memory as is the case for Apple machines.
>>109416554maybe because it's a HUGE FUCKING BOTTLENECK
>>109417254M3 Ultra bandwidth is 800GB/s
>>109418110That's the theoretical bandwidth maximum. In reality it's much slower, you can measure it and it's around ~90GB/s for the CPU and a bit higher for the iGPU. These are standard numbers because the CPU maximum bandwidth is simply core * clock_speed * max_read_or_write_per_cycle, it's the same for GPUs, but they often have heterogeneous cores with different memory access schemes. It's not that the ram is slow, it's just that even using all components that can access ram at 100% of their throughput capacity it won't be enough to use all the bandwidth.Real GPUs with their own memory can reach much faster practical speeds because they own the memory and control what gets written and where, instead of sharing with the CPU.So the GPU bandwidth is not impacted by the south bridge or north bridge/pcie or even the cpu.
That new 9050 would be excellent if they figured out how to stick it onto a CPU substrate and made an APU out of itInstead it's dog shit
