September 26, 2020

Run Game

Control The Game

Xbox Series X silicon: the Hot Chips 2020 breakdown

Earlier this week, Microsoft took part in the Hot Chips 2020 symposium to deliver its...

Earlier this week, Microsoft took part in the Hot Chips 2020 symposium to deliver its now traditional console silicon breakdown, this time focusing on the core make-up of the Xbox Series X ‘Project Scarlett’ processor and how it compares to its predecessors. Not only did we learn more about the properties of the chip, we also gained an insight into the market conditions surrounding and influencing its design. The conclusion is inescapable: cutting-edge consumer electronics can still deliver generational gains in performance – Moore’s Law isn’t dead – but costs are rising and these are having a fundamental influence on the way Microsoft shaped its new console.

Kicking off the presentation, Microsoft’s silicon engineers revealed the Xbox Series X processor lay-out – the die-shot – showing how all of the individual components slot into place within a single 360.4mm2 slice of silicon. Showing how much space the CPU, GPU and other key components are allocated gives us some idea of their importance to the overall design, and the balance looks similar to what we’ve seen in existing AMD-based consoles. Around 47 per cent of the entire area is gifted to the 56 AMD RDNA 2 graphics compute units (four of which are disabled in order to allow chips with minor defects to make their way into production consoles) with around 11 per cent of the space given over to what Microsoft describes as server-class Zen 2-based CPU clusters. A similar amount of area is also consumed by the GDDR6 memory controllers – there are ten of these in total – and while they address 16GB of total RAM in a retail console, the channels are also good for 40GB of memory in the Project Scarlett devkit, and we should assume that once integrated into the Azure cloud, the chips will be using some other kind of non-retail memory set-up.

Referring to the CPU clusters as server-class has caused some confusion as the basic configuration and cache set-up is remarkably close to AMD’s Renoir design (Microsoft has a ‘Hercules’ codename internally for the CPU design) used in its Ryzen 4000 notebook line, as opposed to the monstrous many-core Epyc offerings that are actually used in enterprise environments. It’s far more likely here that server-class designation simply refers to the necessary security features and memory support required to integrate the Scarlett silicon into the Azure cloud. The importance of the processor’s cloud support can’t be understated: the chip can run and stream next-gen games, obviously, but it can also virtualise four Xbox One S instances simultaneously.

Read more

Source link