General Next Generation Rumors and Discussions [Post GDC 2020]

[iroboto]

There was some mentions here and elsewhere that sony compensated "poor yields" by pushing more power to the chips to achieve the clock speeds. i.e. bad yields being that using sensible voltage the chips would not reach their specified clock speed. In my books this is FUD. Sony would have known the characteristics of their chips for long time now and it wouldn't be a surprise towards end of the project.

edit. PS5 soc is not that big of a chip presumably. It would be unlikely it would have a lot of defects and bad yields in general. AMD has enough experience with 7nm to make that a non issue. Or at least it should not be yielding any worse than any other amd designed 7nm product. The open item in the fud is achieving the 2.3GHz clock speed and chips not yielding that high clock speed.

Fan noise has no correlation to the chip.
Fan noise is a result of the quality of the fan and the quality of the cooling.

Case in point, you would agree that PS4 slim runs lower TDP and cooler overall than Xbox One X.
And most people would agree that X1X runs quieter than PS4 slim. Both use blower designs.
The quality of cooling is what makes X1X quiet. And the higher power and bigger chip means more performance. All of those resulted in a higher price device.

The two aren't even related.

As noted by Mark Cerny, the clock speeds chosen for PS5 in terms of boost were decided by what they believe would give them the best possible parametric yield on the device. That doesn't mean it's the best possible yield, just what they were willing to trade off to obtain that speed.

The XSX is 363mm2, some quick math will put PS5 at 300mm2
PS5 into a 12" wafer (300mm) is 235 chips in a wafer.

As per this report in 2019:
https://www.overclock3d.net/news/cp...uVrfanB-mziRJSJIldcTrB7PwAXx0u2oXyxo9K8BhtICY
AMD was obtaining 70% yield on 7nm process for their Ryzen 3000 line. Which was considered decent yield.

For PS5
At 70% yield that nets you 164 chips. Or a cost of 61 USD per chip. at $10K per wafer.
@ 75% you get 176 chips, or $56 USD per chip.
@ 80% you get 188 chips, $53 USD
@ 85% you get 199 chips, $50.25

If you look at just moving up from 70 to 75%, you're seeing 5USD less per chip! 10M chips that's 50M!

You look at XSX for a second here max at 363mm2 you get a maximum of 194 chips
70% 135 chips, $73
75% 145 chips, $68.72
80% 153, 64.43
85% 164, 60.64

According to these calculations here, If XSX manages to obtain 85% yield, and PS5 only 70% they would be the same price.
Whether you choose to believe that is the yield differential that is entirely up to each person.

Yield will only equate to costs. They don't sell nearly enough of these devices to warrant any type of shortage. Nor will is result in a poor thermal designs etc. Those all can be accounted for through price.

You are right that speculation is FUD in general, but we're having a technical discussion here, we're supposed to look at the full spectrum from best case to worst case and ideally if we're right the product lands somewhere in the middle. Having a technical discussion is looking at the whole, there's nothing subjective about it, power increases TDP which requires more cooling. You skip on the cooling it gets loud. You decide to keep it quiet it costs more. All of these issues can be handled through cost.

But right now, as you can see, 15% yield difference between XSX and PS5 will marginalize the difference in their chip size. What the yields are, are unknown, but as I see it
XSX is 20% larger, it's 20% more memory chips and has 20% more storage size. By all accounts it should be 20% more expensive than PS5 to the consumer. And if they come out to the same price, well you'll have to decide if all this speculation was FUD or if some of it landed.

I think I recall a lot of people calling the GitHub leaks entirely FUD as well. But aside from the variable clocking, it hit everything.

 

[iroboto]

That's certainly the way the rumours had presented things in this case. It was presented as a significant problem in PS5's design (the high clock speeds) with notable impact on availability and/or costs.

If the original info was just, "Sony not seeing the yields they hoped for," responses and arguments may have been more conservative.

hmm indeed hot takes.

 

[Shifty Geezer]

The problem I see with the yield thing, as already mentioned, is 1) why would Sony have lower yields than MS on the same architecture and fab? 2) If due to parametric yields, what can they do to improve those over time?

If the architecture cannot deal well with the power requirements, how can that be improved on the current node? To my ill-informed mind, you'd be stuck with that problem until the next node shrink. Although I guess GPU refreshes point to some changes allowing clocking higher? But until I see a good explanation how the yields could have been poor but now be good, I find it difficult to believe there were ever significant yield problems. It was conceivable based on this idea of Sony pushing the chip limits right up until 10 M units ramping up was announced; that's a very contradictory data point, and a solidly reliable one rather than a vague suggestion.

 

[manux]

Fan noise has no correlation to the chip.
Fan noise is a result of the quality of the fan and the quality of the cooling.

I agree in perfect world. My line of thinking was following. Sony assumed the worst case power output would be let's say 200W(random number) and designed cooling accordingly. Now sony gets chips back and they don't hit 2.3GHz. Sony redesigns power supply to provide 300W to achieve the 2.3GHz on poorly yielding chips. Cooling is now way under specced and super noisy. Of course sony can fix this by redesigning cooling and case but still we would see the console pull 300W from wall.

Once consoles are released we can measure how much power ps5 draws and compare it against other devices. This would tell us how hard sony has to push the chips to reach the advertised performance.

 

[iroboto]

The problem I see with the yield thing, as already mentioned, is 1) why would Sony have lower yields than MS on the same architecture and fab? 2) If due to parametric yields, what can they do to improve those over time?

If the architecture cannot deal well with the power requirements, how can that be improved on the current node? To my ill-informed mind, you'd be stuck with that problem until the next node shrink. Although I guess GPU refreshes point to some changes allowing clocking higher? But until I see a good explanation how the yields could have been poor but now be good, I find it difficult to believe there were ever significant yield problems. It was conceivable based on this idea of Sony pushing the chip limits right up until 10 M units ramping up was announced; that's a very contradictory data point, and a solidly reliable one rather than a vague suggestion.

If we're speculating on how both could be true: ie yield is not ideal and they ramped up production

I believe that parametric yield can improve with the manufacturing process, ie when 16nm as a node was released its cost per transistor and yield is likely worse than at full maturity.

If money is a concern, then by releasing a smaller supply of devices, you lose less money over time than if you were to release larger quantities at the beginning.

I don't necessarily believe that yield problems would be so dramatic at the chip level that it would limit quantities. The goal for them is to measure demand. If they believe demand is there and the units can be sold, they will sell it. If you build too many you will suffer warehouse penalties, you make too few you are losing money to your competitor, a competitor who (rumour wise) is selling 2 SKUs of which they are sandwiched in price.

As DSoup said, there may be business justifications to selling more of that less yield run that Sony may be willing to act on to ensure long term success of PS5. As we know this generation will end with 200M units sold between the both of them over 7 years. But there will be 3 SKUs starting this generation and 2 or 3 price points as opposed to 2 SKUs then moving to 4. The pie will be divided smaller. Price conscious users may opt to go for the lowest price, power conscious users will opt for the one with more power. The general population will go for the games and where their friends play.

Last generation PS4 was $100 cheaper and more powerful and they had games coming off PS3. They had claimed the whole market without much trying. Getting to the first 10M may be their goal here to try to claim the generation early and ride critical mass to generation success.

 

[manux]

If we're speculating on how both could be true: ie yield is not ideal and they ramped up production

Amd has already released zen2 and navi on 7nm. This year there will be zen3, big navi and xbox on 7nm. The question would be why specifically the sony 7nm chip would have poorer yields than other amd 7nm designed products? The clock speed is reasonable guess(or fud) as that is one thing that is something unique to sony. PS5 chip should be fairly small which should help yields versus some of the bigger chips.

 

[iroboto]

Amd has already released zen2 and navi on 7nm. This year there will be zen3, big navi and xbox on 7nm. The question would be why specifically the sony 7nm chip would have poorer yields than other amd 7nm designed products? The clock speed is reasonable guess(or fud) as that is one thing that is something unique to sony. PS5 chip should be fairly small which should help yields versus some of the bigger chips.

indeed, yield should have improved.
Why would Sony have worse yield? It wouldn't. Yield is determinant on how high you set the performance bar for a specific design. The higher the performance you require the less yield at that level. The lower the performance, the more yield. Clockspeed/timing and therefore voltage stability is a pretty big part of that discussion.

 

[Deleted member 13524]

Fan noise has no correlation to the chip.
Fan noise is a result of the quality of the fan and the quality of the cooling.

The amount of air the fan needs to displace depends on the chip's heat output.
The rotation speed of the fan depends on the amount of air the fan needs to displace.
The fan noise depends on rotation speed.

I don't know how one can ever state that "fan noise has no correlation to the chip".

 

[mrcorbo]

The amount of air the fan needs to displace depends on the chip's heat output.
The rotation speed of the fan depends on the amount of air the fan needs to displace.
The fan noise depends on rotation speed.

I don't know how one can ever state that "fan noise has no correlation to the chip".

Perhaps a better way to word that would have been, "A hot chip doesn't automatically mean you need a loud fan to cool it."

The point being: you should be able to engineer around it at some additional cost.

 

[iroboto]

The amount of air the fan needs to displace depends on the chip's heat output.
The rotation speed of the fan depends on the amount of air the fan needs to displace.
The fan noise depends on rotation speed.

I don't know how one can ever state that "fan noise has no correlation to the chip".

That's only for a perfect fan. There's pitch and tone. Ball Bearing noises. All sorts of other issues that people pick up on that aggravates them other than the pure decibel of the noise. The quality of a fan can remove everything but air displacement. And the design of the heatsink and case are for reducing the sound of airflow.

 

[sir doris]

Why would poor yields or power issues result in a loud fans ?

More marginal dies need to be used, requiring more voltage, which produce more heat, thus requiring the fans to spin faster and louder?

 

[iroboto]

More marginal dies need to be used, requiring more voltage, which produce more heat, thus requiring the fans to spin faster and louder?

once again, doesn't mean that hot chips results in loud fans. If you already knew it was going to run hot the voltages around it, you would design the cooling system around it if you have an audible quality target. If you are going to cheap out on the audible portion of the requirement, then yes it would get louder.

And I'm not sure what level of marginal dies PS5 can run. They run a fixed voltage variable clock setup that is controlled by workload activity. If the die cannot produce the frequency of the workload at a given power level, it wouldn't run in sync with the rest of the consoles.

 

[Shifty Geezer]

I believe that parametric yield can improve with the manufacturing process, ie when 16nm as a node was released its cost per transistor and yield is likely worse than at full maturity.

What constitutes 'full maturity'? Is that per chip, or per process? I thought it was the latter, in which case surely TSMC's 7nm capabilities are very mature, meaning very little room for improvement.

once again, doesn't mean that hot chips results in loud fans.

The problem with your statement is you said fan noise is out of the equation, but it's part of it. The hotter the chip, the higher the cooling requirements, the more pressure there is to run fans faster and louder relative to a cooler design.

"hot chip != loud fan", but "fan noise ∝ chip temp" as part of the temperature > size > heat sink > fan size > fan speed relationships.
I guess in exact terms, your phrase, "Fan noise has no correlation to the chip," should be, "Fan noise has weak correlation to the chip." The chip doesn't determine fan noise, but it does influence it, especially in price constrained products.

 

[iroboto]

The problem with your statement is you said fan noise is out of the equation, but it's part of it. the hotter the cheaper, the higher the cooling requirements, the more pressure there is to run fans faster and louder relative to a cooler design.

"hot chip != loud fan", but "fan noise ∝ chip temp" as part of the temperature > size > heat sink > fan size > fan speed relationships.
I guess in exact terms, your phrase, "Fan noise has no correlation to the chip," should be, "Fan noise has weak correlation to the chip." The chip doesn't determine fan noise, but it does influence it, especially in price constrained products.

Fair enough. The correct terms to use would be that weakly correlated as there are other features that can have an impact on noise and may not represent the majority fan noise.

As for parametric yield it’s outside my understanding. Functional chips that don’t make the performance cut, I do not know what makes it better. I just know that the higher the frequency you fewer chips can handle the timing and voltages required to support it and I don’t know what can be done to resolve that.

 

[Shifty Geezer]

As for parametric yield it’s outside my understanding. Functional chips that don’t make the performance cut, I do not know what makes it better. I just know that the higher the frequency you fewer chips can handle the timing and voltages required to support it and I don’t know what can be done to resolve that.

That's it. In my mind, that's tied to chip design. You can't really do anything to make it run better once you've taped out the design, so I don't know how yields can then be improved. That's different to manufacturing faults which improve at the fab and do improve over time as the fab improves its process.

I don't know if we have anyone at B3D who knows chip fabrication well enough to know if it's possible or not.

 

[Deleted member 13524]

Perhaps a better way to word that would have been, "A hot chip doesn't automatically mean you need a loud fan to cool it."

This is correct, but @iroboto 's sentence was "fan noise has no correlation to the chip", which is a demonstrably false assertion in every cooling solution that uses a fan.

That's only for a perfect fan. There's pitch and tone. Ball Bearing noises.

Pitch, tone and all other noises are influenced by rotation speed. I'm not sure what you mean with perfect fan. I never said only airflow influences noise. I said rotation speed influences noise.

The quality of a fan can remove everything but air displacement.

A fan of this quality doesn't exist. That is what I would call a perfect fan.

 

[iroboto]

This is correct, but @iroboto 's sentence was "fan noise has no correlation to the chip", which is a demonstrably false assertion in every cooling solution that uses a fan.



Pitch, tone and all other noises are influenced by rotation speed. I'm not sure what you mean with perfect fan. I never said only airflow influences noise. I said rotation speed influences noise.



A fan of this quality doesn't exist. That is what I would call a perfect fan.

Sure but you’re picking at my words after it’s been explained that I’ve used incorrect wording on it and I’ve walked back on that already.

Fan noise is subjective. Quiet is subjective. As long as it’s fairly quiet most people will be happy with it. I was writing against the assertion that just because the chip runs hot that doesn’t imply that fan noise should be intolerable or not capable of being quiet. Additional engineering should be capable of bringing the noise to a level that most people would be comfortable with.

Which i agree and understand that is not the same as saying it’s not directly correlated to chip temperature which is something else.

 

[Arwin]

I am really curious about the cooling because Cerny hinted at it being something special. Back in the PS3 days it was also pretty neat (huge fan cleverly placed). Maybe this time they have a corkscrew design like in a submarine displacement engine, watercooling or they worked together with Dyson [emoji16]

 

[Deleted member 13524]

Sure but you’re picking at my words after it’s been explained that I’ve used incorrect wording on it and I’ve walked back on that already.

Sorry, I've been writing my (short) posts over the course of hours, when I get the time to do so. When I hit post I wasn't aware the issue had been discussed already.

 

[Deleted member 13524]

I am really curious about the cooling because Cerny hinted at it being something special.

Cerny's statements and the "sandwich heatsink" patent have put the PS5's cooling system on quite the hype train.
OTOH, the console is huge so perhaps there's just an enormous amount of copper and large fans inside. Which is pretty cool by itself, IMO.