General Next Generation Rumors and Discussions [Post GDC 2020]

[Deleted member 11852]

Sort of a variable target though. If you are looking for the cream of the crop chip and not getting as many, you're bound to have poorer yields than a lower binned chip.

Sony wouldn't be aiming for cream of the crop, this would be crazy expensive because you're literally binning everything that fails to meet that high bar. What are Sony/AMD/TSMC going to do with a PS5 APU if Zen can't sustain 3.5Ghz and/or GPU can't sustain 2.33Ghz.

Now I'm assuming clock because anything impacting Zen would be impact XSX which runs it at a marginally higher clock. There are other issues other than clocks that can impact yields but you'd expect those same issues to impact other designs based on this AMD template, which includes XSX. You can bet the core of PS5 and XSX APUs are more alike than different other than PS5 being significantly smaller. Smaller = less defect likelihood from a bad wafer.

When you rule out in the impossibilities, none of the possibilities are good.

 

[AlNom]

Sony wouldn't be aiming for cream of the crop, this would be crazy expensive because you're literally binning everything that fails to meet that high bar. What are Sony/AMD/TSMC going to do with a PS5 APU if Zen can't sustain 3.5Ghz and/or GPU can't sustain 2.33Ghz.

But the paradox is that Cerny is saying that it's moistly at 2.226GHz, so it has to run pretty hot for a lot of "not-awesome" chips.

hm... so the cooling and circuitry would have to be pretty robust.

edit: not arguing, just mulling over things.

---

Then again, maybe mostly just means that it doesn't switch clocks too often as that'd be added stress on components.

 

[Deleted member 11852]

But the paradox is that Cerny is saying that it's moistly at 2.226GHz, so it has to run pretty hot for a lot of "not-awesome" chips.

I took iroboto's post to suggest that this clock speed was at the extreme end of what the yields typically produce (i.e. the cream), rather that yields of PS5 APUs would produce plenty of chips capable of running 3.8Ghz/2.5Ghz. I.e. in broad yield terms 3.5Ghz/2.33Ghz is not the cream, it's below.

 

[AlNom]

It's yields at an acceptable TDP/power consumption, although I guess we don't know anything about the retail unit yet.

(it's 2.23Ghz rounded up )

Where's the 2.5GHz figure from?

 

[iroboto]

I took iroboto's post to suggest that this clock speed was at the extreme end of what the yields typically produce (i.e. the cream), rather that yields of PS5 APUs would produce plenty of chips capable of running 3.8Ghz/2.5Ghz. I.e. in broad yield terms 3.5Ghz/2.33Ghz is not the cream, it's below.

I don't disagree with your thought process.
But simplistic data points thinking: we heard no rumour about XSX having yield or heat, in fact they started production a while back. Not much stepping according to Github. WRT PS5, lots of steppings, hearing a lot about heat and yield rumours, Sony still has not revealed a console,Tott brings up another rumour of yet another last minute revision. Simplistically without really knowing much and what is true or not, but based on this information I'd look at clockspeed differences for yield issues. A last minute revision to improve yields?

 

[eastmen]

When did you hear about bad yields?
IIRC, there were rumors of the PS5 SoC getting yet another pre-mass production revision as late as last month.

multiple people who work at sony that I know. I can ask about the revision. I was told yields are worse than they thought driving prices up. My sources in MS have said the opposite that their yields are better than what they thought they would be

 

[Tkumpathenurple]

Wouldn't it have to be quite a lot more expensive to be more expensive than the XSX though?

Going by this wafer calculator, a wafer diameter of 450mm, and the following dimensions:
XSX = 360.45mm2
√360.45 = 18.99
Dies per wafer = 378

PS5 = 300mm2 (going from an old leak that seems more viable now that we know it only has 36CU's)
√300 = 17.32
Dies per wafer = 457

That's a difference of 79 that the PS5 would have to chew through. Sure, such high clockspeeds may have a detrimental effect on yield, but the same is true of a larger chip too.

 

[iroboto]

Wouldn't it have to be quite a lot more expensive to be more expensive than the XSX though?

Going by this wafer calculator, a wafer diameter of 450mm, and the following dimensions:
XSX = 360.45mm2
√360.45 = 18.99
Dies per wafer = 378

PS5 = 300mm2 (going from an old leak that seems more viable now that we know it only has 36CU's)
√300 = 17.32
Dies per wafer = 457

That's a difference of 79 that the PS5 would have to chew through. Sure, such high clockspeeds may have a detrimental effect on yield, but the same is true of a larger chip too.

AFAIUnderstand; defects in silicon process aren't tied to how high each chip can clock. I believe standard wafer defects determine which silicon chips were made in total functionality. Meaning the chips work.
The redundant CUs are there to shut off portions of the chip that received such a defect.
But clock speed is a different type of defect, that needs to be tested after passing the basic check, and a factor for yields (wrt clocking) can be largely determinant on the design of the chip and what you want it to do.
MS chose fixed clocks across the board and an overall lower clockrate. This might work very well with RDNA2, they chose a power requirement that fit

The above 2230 GHz may not necessarily, so perhaps some design changes are required to get it to operate better at that clock speed. It is very high for a GPU. Whereas you’d expect this clockspeed for a modern CPU running AVX256/512 instructions. CPUs won’t maintain their 4Ghz speed once they leave the single threaded domain. Some will drop to 1600Mhz.

so yea a GPU running 2230 is going to be running really hot. Be really power hungry under load. You may need better yields for that.

 

[Tkumpathenurple]

I don't doubt the potential for truth with some of the rumours regarding a redesign. Yields could be lower than expected.

But I find it harder to believe that yields are so poor at 2.23GHz, that it could mitigate the benefit of 79 extra dies per wafer and end up making the PS5 more expensive.

 

[MrFox]

Cerny said they chose the clock based on parametric yield.

 

[AlNom]

Cerny said they chose the clock based on parametric yield.

What parameters?

(yeah yeah, just poking )

 

[iroboto]

I don't doubt the potential for truth with some of the rumours regarding a redesign. Yields could be lower than expected.

But I find it harder to believe that yields are so poor at 2.23GHz, that it could mitigate the benefit of 79 extra dies per wafer and end up making the PS5 more expensive.

I'm not really convinced either. I'm not suggesting this is for sure what is happening. Just offering an explanation given all information we know about steppings, revisions, articles about heat challenges etc.

 

[iroboto]

Cerny said they chose the clock based on parametric yield.

pre or post tapeout?

pre-tapeout that's just an estimate.

 

[iroboto]

I don't doubt the potential for truth with some of the rumours regarding a redesign. Yields could be lower than expected.

But I find it harder to believe that yields are so poor at 2.23GHz, that it could mitigate the benefit of 79 extra dies per wafer and end up making the PS5 more expensive.

Btw: I found the correct terms for what I was describing.


https://www.inf.usi.ch/faculty/papadopoulou/publications/bookchapter08.pdf
******
Catastrophic Yield Loss. These are functional failures such as open or short circuits which cause the part to not work at all. Extra or missing material particle defects are the primary causes for such failures. A nice picture of a bridging fault is shown in Figure 1 drawn from [63]. Critical area analysis is used to predict this type of yield loss and is discussed later in this chapter.

Parametric Yield Loss. Here the chip is functionally correct but it fails to meet some power or performance criteria. Parametric failures are caused by variation in one or set of circuit parameters, such that their specific distribution in a design makes it fall out of specifications. For example, parts may function at certain VDD, but not over whole required range. Another example source of parametric yield loss is leakage in deep sub-micron technologies [28]. Parametric failures may be caused by process variations. Several kinds of integrated circuits are speed-binned (i.e. grouped by performance). A common example of such class of designs is microprocessors wherein lower performance parts are priced lower. The other class is typical ASICs which cannot be sold if the performance is below a certain threshold (for example due to compliance with standards). In the latter case, there can be significant performance limited yield loss which is why such circuits are designed with a large guardband. In the former case too, there can be significant dollar value loss even if there is little yield loss.

 

[Shifty Geezer]

So the only way PS5 could be having higher yield issues than XBSX is if the chips performance at speed is below expected. Wouldn't that result in fewer units, but no additional costs, because surely AMD are contracted to chips to spec? They'd have provided Sony with estimates that Sony agreed to, and failure of enough chips to be able to run at Sony's specified clock speeds would be a fault of AMD's estimates. I think.

 

[iroboto]

older rumours, but its the first thing that came to mind:

So the only way PS5 could be having higher yield issues than XBSX is if the chips performance at speed is below expected. Wouldn't that result in fewer units, but no additional costs, because surely AMD are contracted to chips to spec? They'd have provided Sony with estimates that Sony agreed to, and failure of enough chips to be able to run at Sony's specified clock speeds would be a fault of AMD's estimates. I think.

https://www.bloomberg.com/news/arti...an-limited-playstation-5-output-in-first-year

Sony Corp. plans to produce far fewer units of its upcoming PlayStation 5 in its first year than it had for the previous-generation console’s launch, according to people familiar with the matter.
...
The company has told assembly partners it would make 5 to 6 million units of the PS5 in the fiscal year ending March 2021, according to other people involved in the machine’s supply chain. When Sony released the PlayStation 4 in November 2013, it sold 7.5 million units in its first two quarters.

hmm.. you may be right.


https://www.forbes.com/sites/paulta...ft-ahead-of-ps5-series-x-launch/#292a6c605786

There have been rumblings on both sides of this issue. I first heard that Sony was waiting for Microsoft to go first, announcing its price so that they could undercut it like they did in 2013 with the PS4 going up against the Xbox One, $400 versus $500.
...
Now, I’m hearing a counter-narrative. That parts/manufacturing costs of the PS5 have pushed its potential price up to probably $450-500 instead, and now Microsoft may be the one waiting for Sony to announce its price first so then they can undercut instead.

 

[eastmen]

I don't doubt the potential for truth with some of the rumours regarding a redesign. Yields could be lower than expected.

But I find it harder to believe that yields are so poor at 2.23GHz, that it could mitigate the benefit of 79 extra dies per wafer and end up making the PS5 more expensive.

What does sony do with any chips that don't reach the full ps5 spec ?

What does microsoft do with any chips that don't reach full xbox series x specs ?

My understanding is that the xsx is yielding better and Microsoft has other projects that are able to use chips that don't make the grade in

also because of other projects from MS they are buying a lot more wafers

 

[iroboto]

My understanding is that the xsx is yielding better and Microsoft has other projects that are able to use chips that don't make the grade in

yup they certainly do.

 

[eastmen]

yup they certainly do.

shh don't spill the beans ! if you know what i'm laying down we gotta keep it as vague as possible. Bill we be upset if everyone finds out they are putting an xsx inside everyones skin to track them

 

[iroboto]

shh don't spill the beans ! if you know what i'm laying down we gotta keep it as vague as possible. Bill we be upset if everyone finds out they are putting an xsx inside everyones skin to track them

...
I'm referring to Xcloud, each Scarlett SoC is being put to support 4 individual XBO instances online.