As we approach the launch of Xbox One, more official details are coming and the Hot Chips conference, John Sell (Xbox One chip architect) himself has shed new light on what the Xbox One Processor is and what it can do. First of all, he confirmed what the Xbox One design already told us: Microsoft absolutely wants to avoid a repeat of the initial overheating, crash (ring of death) and product return problems that plagued the initial release of Xbox 360. That probably explains the ample (oversized?) inside volume, and the presence of cooling vents all over the console.
It’s a big chip
Secondly, I don’t think that the 363 square mm chip surface area of the Xbox One processor was known before, so this is an interesting detail for those geeky enough to care. To give you a comparison, Intel’s Core i7 4770K “Haswell” is 177mm big, so the Xbox chip is quite enormous. Note that the Xbox chip is built with a 28nm process while Intel is inherently smaller at 22nm – we’re not comparing apples to apples, but are merely providing a frame of reference.
It may be counter-intuitive, but a “big chip” is not always “better” because the die size will affect the initial yield during manufacturing. If there is any defect on the wafer’s surface (and there are always some), the whole chip could be in jeopardy. Smaller chips will statistically lead to better yields, it’s just math. To address this immediate concern, Microsoft has built some redundancies so that such wafer defects may be worked around by disabling the malfunctioning unit and activating a backup one. I would love to see an X-ray of the chip in order to guess what’s redundant…
Microsoft in for the long haul
Microsoft has designed its main processor (or SoC) for the long haul: over the life span of the Xbox One, the chip will be cost-reduced, either by using an optical shrink and more likely by redoing the physical design to optimize its chip layout and adapt it to new libraries required for a new semiconductor manufacturing process.
That’s how Microsoft will benefit from Moore’s law: the Xbox One chip will cost exponentially less and less to produce over time. Having the graphics processor integrated into the same chip also reduces the overall cost for the lifetime of the console (1 chip vs. 2 to build), but will also allow for lower latency when GPU, CPU and embedded memory communicate. That’s particularly true for the chip’s 47MB of internal memory/cache. (Note that “47MB” includes one block of 32MB eSRAM, the rest is in fact various memory caches around the chip…)
Bandwidth war of words: mostly meaningless
Microsoft also mentioned that data could move as fast as 200GBps within the chip. This is undoubtedly going to add some fire (and mostly more confusion) in the debate between PS4 and Xbox1 fans. The PS4 uses 8GB of fast GDDR5 RAM while the Xbox1 is designed to use common RAM and optimize bandwidth with its 32MB of fast eSRAM.
The approach worked with the Xbox 360, but it’s hard to compare it with PS4 without having real apps and use cases. The bottom-line is: actual performance will depend completely on how the games are programmed. I’m waiting to hear from cross-platform developers before speculating on the impact of those architectural decisions.
Microsoft claims 200GBps of internal memory bandwidth, while Sony claims 176GBps of external memory bandwidth. Funny enough, although those two numbers can’t be compared in any way, fans are going to fight over it for the next few years…
Xbox One designed to be “always-on”
In order to reduce power consumption (and therefore heat), and since the Xbox is supposed to be available at all time via voice command, Microsoft’s hardware team has designed the chip with “power islands”. This means that many logic blocks can be turned ON and OFF, without affecting the other blocks. While this may sound quite a normal thing to do, such designs add more complexity since one needs to lay out more internal wires to completely (electrically) isolate each block. This tends to make the chip even bigger.