I wondered if chiplet GPU could be added to the main GPU even at a lower or different memory performance. The idea being if the Azure cloud wants something other than a gaming GPU, add tensor cores or some other variation, could that be added to the die as a chipletc connecting with the memory controller and used when not in game mode.
Console chips don't have it.
Azure chips do and when working outside game mode different portions of the GPU are active and using memory. Azure may use faster memory chips to give an extra overhead for streaming operations.
It's all going to be exposed via a HV anyhow so I am sure ms knows how to carve up chip.
Make the Azure side performant for what they want, keep gaming pure, and reduce silicon costs for consumer chips. Azure chips will cost more but be dual function and be good at both.
TLDR the Azure customisation is a bolt on to not affect gaming outside the way the chip is designed for memory.
Console chips don't have it.
Azure chips do and when working outside game mode different portions of the GPU are active and using memory. Azure may use faster memory chips to give an extra overhead for streaming operations.
It's all going to be exposed via a HV anyhow so I am sure ms knows how to carve up chip.
Make the Azure side performant for what they want, keep gaming pure, and reduce silicon costs for consumer chips. Azure chips will cost more but be dual function and be good at both.
TLDR the Azure customisation is a bolt on to not affect gaming outside the way the chip is designed for memory.