XBone eSRAM/GPU Performance Problems Rumors
By eloj 60 Comments
There's a rumor that Microsoft are having problems with the XBone. There's been rumors before that they're behind on software, but these latest rumors pertain specifically to the hardware and the combination of eSRAM and GPU. The rumor is saying that they're having yield problems, specifically related to the eSRAM.
I will now go on and speculate widely. You have been warned.
It is claimed that the XBone APU is a one-die CPU+GPU+eSRAM solution, meaning that it's very large (~410mm^2). The only thing MS really said at their reveal was "5 billion transistors". This is interesting, and some correctly pointed out that it probably meant they couldn't compete with the PS4 on specs, because if they could, they'd be more specific.
The primary risk with large dies of course are low yields, resulting in higher costs to get enough dies that meet your specification. Now, Microsoft was not building a gaming machine first. They wanted to load a familiar operating system on there. They wanted more focus on applications that historically have not been heavily optimized for one platform. A relatively fat OS, and applications from developers who are not used to the scant resources typically awarded a gaming console, leads to one design conclusion; "we need lots of memory!"
So they design for 8GB, from the start. Again, they put applications (and cost) above gaming, and go for a well known, cheap, plentiful technology in DDR3. Back then 8GB of GDDR5 would look too expensive, almost insane, and DDR4 wasn't on the immediate horizon.
To compensate the gaming side, they now need some fast RAM, hence 32MB eSRAM on-die. The XBone eSRAM is said to be a 6T type, meaning that each bit requires six transistors. 32MB gives us 32*1024*1024*8 bits times six equals 1,610,612,736 transistors. That's 1.6 billion, a huge chunk out of the total 5 billion!
They do desperately need this very fast cache-like memory to fix the fact that they're using "slow" DDR3, which they settled on because they wanted to have a lot of it, for apps -- not games.
Recap: Decision to not focus on games results in design for a lot of memory, which is "slow" for cost reasons, which is compensated with cache. Result: Large die, with large percentage dedicated to what turns out to be a possible problematic memory to fab. Because the cache use so much die, they have to scale down on GPU.
They're reportedly roughly 2 billion transistors larger than the APU used by Sony in the PS4, which means worse thermals, which means lower clocks. The rumor is saying that the yields on this APU, due to the eSRAM, is so low, that they may have to cut the clocks on the eSRAM+GPU even further than what they've previously communicated to developers under NDA. This would make the XBone even weaker on the GPU side vs the PS4, which was already believed to be quite a bit stronger.
MS probably thought for the longest time, like PS4 devs did, that the PS4 would only have 4GB of GDDR5. Microsoft could work with that, they'd have the fast cache (which no doubt can be awesome when properly used -- I'm a huge cache nerd) and more total memory. They designed their system around being the platform with More Memory.
Cerny meanwhile had bet on GDDR5, and his bet has been that availability will increase and cost decrease at such a pace, that even back when they were first thinking about the PS4 and 2GB seemed like ample amounts, as time went on 4GB became viable.. and then at the last second... tick. Out of nowhere Sony steps up on stage and announce that they'll be using 8GB of GDDR5, twice the amount insiders thought for sure.
And they probably did this to match Microsoft, which has now lost their one advantage in specification. They on their parts are left having to pay the price of designing for apps before games, and that price just went up if this rumor is true.
There are so many "#potentialscenario" implications here, which I'll write about in a later post.
60 Comments