PlayStation 5 Launch Supply Reduced due to AMD CPU/GPU SoC Yield Issues

[AleksandarK]

Today we have found out that Sony has reportedly cut PlayStation 5 launch supply due to bad yields of the SoC powering the console. Previously, we reported that Sony has doubled production of the new console amid high demand, where the company expected to sell 10 million units in the fiscal year. The original plan was to have around 15 million units of the new console available by March 31st, 2021. Sony has been spending a lot of resources to get as many units out to consumers, however, the bad SoC yields have held the company back significantly.

It is reported by Bloomberg that instead of the original 15 million units Sony plans to supply, there will be only 11 million of them. That represents a massive reduction of 4 million units. And you are wondering how bad the yields of the new SoC are to have that big reduction. According to the source, TSMC and Sony are seeing only 50% yields on the production run. It is reported that the yields are gradually improving but have not yet reached the level needed to have a stable supply. This represents a big problem for the company and we don't know who is to blame. TSMC has been very good at manufacturing 7 nm silicon, however, it could be bad design from AMD and Sony that is making the production difficult. We are waiting for more information.


View at TechPowerUp Main Site

 

[Shatun_Bear]

That reduces it to 11 million still until March 2020....why are they so optimistic on selling so many? That'll be about 6 million more than the Xbox sells combined.

 

[Melvis]

That reduces it to 11 million still until March 2020....why are they so optimistic on selling so many? That'll be about 6 million more than the Xbox sells combined.

They have always sold more then Xbox

 

[fynxer]

Lower than 50% yield on 7nm, ***SURE they have***, this is nothing more than marketing ploy maskerading as "FAKE" NEWS.

Earlier this summer report came out that it was over 70% yields, which is much more plausible.

Elaborate marketing ploy to drive pre-orders. If people think there will be a shortage they are much more likely to pre-order and as pre-orders increase it could actually create a shortage at release.

Today we can not trust BIG CORP information on the internet. To many news sites just copy each other so big corps use this network of mindless drone news sites for guerilla marketing as in this case create more pre-orders for PS5 by fabricating a yields problem that nobody actually can verify and thus create a non existing CPU shortage to get more sales.

Since there are a few redundant CUs calculated from the beginning to improve yields, usable socs should be around the following:
90% for the XSX
95% for the PS5 (more since it is smaller and less powerfull than the XSX soc)

"Yields aren't really a problem, on tsmc N7 with a defect density of 0.09 defects per square cm (which was the defect rate in july), the XSX soc yields 73% perfect and the PS5 soc is smaller and will have better yields, also most of the 41 defect dies are still going to end up usable for retail too with as many redundant CUs as they're working with. So yeah... yields are fine, for the ps5? even better. I'd estimate 90% yields for the XSX soc on the low end, and 95% for the ps5 SOC. TSMC dont play around."

 

[R0H1T]

Earlier this year report came out that it was over 70% yield which is much more plausible.

That was for PS5 SoC was it, any credible source you might have?

 

[Vya Domus]

Yeah I don't really buy the 50% yields either. I think this is simply a limitation of raw volume of wafers available.

TSMC has been very good at manufacturing 7 nm silicon, however, it could be bad design from AMD and Sony that is making the production difficult. We are waiting for more information.

If you have say a 200 mm^2 piece of silicon it will have the same probability of having a defect no matter what it contains. There's isn't anything "bad" that they could have designed, the SoCs are made up of modules regardless.

 

[Bwaze]

All flagship expensive hi-tech products now have a long period of "scarcity" - so they don't reach market saturation too soon.

And people are "lucky" if they can even get the product, nobody waits for price reductions that might not even come (Nvidia RTX 2080Ti sold for absurd full price even two years after release).

It sure looks like planned scarcity to me, industry wide.

 

[Deleted member 185088]

I'm a bit sceptical, the SOC of the PS5 is more or less a smaller version of the Series X, it should have better yield then, maybe they couldn't secure enough production for themselves given the huge competition, Zen3, RDNA2, Xbox, or they just playing with their fans.

 

[Sykobee]

Something doesn't add up here.

Sony were aiming for 10m by March 2021 - already massive numbers for a mere 5 months of availability. This was a rise from 5-6m (https://gamingbolt.com/ps5-will-have-5-to-6-million-units-produced-by-march-2021-report) to 10m (https://www.businessinsider.com/playstation-5-production-doubled-report-2020-7?r=US&IR=T)

That has become 15m at some point it appears, but there was never any leak or suggestion that 15m was ever going to be the quantity produced.

We know TSMC 7nm yields well, and we know that AMD knows how to design a APU on 7nm. This seems unlikely as a source of the problem, unless the only yield issue is hitting the desired clock speed of 2.2+ GHz.

Now if it was being built on Samsung 5nm, then I would understand the yield angle! But it's not. It's not even TSMC 5nm, although that would be more believable.

 

[fynxer]

That was for PS5 SoC was it, any credible source you might have?

TSMC N7 had a defect density of 0.09 defects per square cm (which was the defect rate in july)

SOURCE TSMC them self.

XSX soc had 73% yields in july and XSX soc is bigger so PS5 soc will have better yields than XSX soc.

 

[Sykobee]

You also have to consider that the PS5 APU is mostly GPU, and it will be overprovisioned (40CU instead of 36). Most defects can be bypassed by disabling the CU it happens in. Some will occur in other logic and may not be fixable, of course.

Even ignoring the spare CUs, playing around with a die yield calculator, the defect density would have to be over twice that TSMC have publicly reported to reduce a 300mm^2 die yield to 50%. http://isine.com/resources/die-yield-calculator (I've assumed the PS5 die is 300mm^2, fabbed on a 300mm wafer, use 0.24 defect density instead of reported 0.09 to get around 50% yield). For 11m consoles until March assuming that's six months of wafer manufacturing, let's call that 2m dies a month, that's 15,000 wafers a month (135 fully working dies/wafer at 0.09 defects/cm^2, out of 180 die candidates), not a problem for TSMC.

If the issue is hitting the 2.23GHz clock speed in the GPU, rather than defects, and if the APU has 40CU instead of 36, then there is a potential solution for Sony, use all 40CUs at 2GHz. Take the defect yield hit in order to gain even more dies that can run at the high clocks. I don't expect them to take this option, it depends on their contract with TSMC&AMD, they might be paying for working dies that meet their specification, not wafers, for example.

[another solution is stick the slower ones in the digital and reduce it to 9 TFLOPS - still over twice the XSS performance in theory. But it would make game dev a real pain, supporting these two close configurations, some would just aim for the lower figure for both systems.]

And they will have been aware of the yield issues for a couple of months by now, if it is true - maybe enough time to allow a change like this to be tested and verified as okay. Some people say that 36 was chosen because it matches PS4 Pro (and is double PS4) and that's required for backward compatibility - I don't buy that myself, but consoles can be a bit odd.

 

[Vya Domus]

If the issue is hitting the 2.23GHz clock speed in the GPU, rather than defects

I don't believe that to be the case, it has been hinted by Sony that the GPU rarely hits that clock speed in real world use. RDNA 1 can hit 2 Ghz no problem, I can't see why 2.2 would be an issue.

 

[Sykobee]

Rumours were that RDNA2 could clock really well anyway, 2.2 shouldn't be an issue.
But it makes more sense than defective dies for other reasons, because we know the process is good, and we know AMD can make an APU.

Personally, I think the Bloomberg reporter got trolled by this 'anonymous source'.

 

[laszlo]

rumors say that workers at TSMC saw a green man inside the factory without credentials...

 

[Vayra86]

That reduces it to 11 million still until March 2020....why are they so optimistic on selling so many? That'll be about 6 million more than the Xbox sells combined.

Perhaps they knew this was coming. Shoot at 11 and you might get 6. Its no secret 7nm is crowded.

 

[john_]

Maybe they where not expecting a new XBOX at $299, so they have to take this in consideration. Meaning less consoles sold in the same time frame.

 

[Vayra86]

Something doesn't add up here.

Sony were aiming for 10m by March 2021 - already massive numbers for a mere 5 months of availability. This was a rise from 5-6m (https://gamingbolt.com/ps5-will-have-5-to-6-million-units-produced-by-march-2021-report) to 10m (https://www.businessinsider.com/playstation-5-production-doubled-report-2020-7?r=US&IR=T)

That has become 15m at some point it appears, but there was never any leak or suggestion that 15m was ever going to be the quantity produced.

We know TSMC 7nm yields well, and we know that AMD knows how to design a APU on 7nm. This seems unlikely as a source of the problem, unless the only yield issue is hitting the desired clock speed of 2.2+ GHz.

Now if it was being built on Samsung 5nm, then I would understand the yield angle! But it's not. It's not even TSMC 5nm, although that would be more believable.

You're on to something. Nice to see some other solid perspectives.

 

[ODOGG26]

Maybe they where not expecting a new XBOX at $299, so they have to take this in consideration. Meaning less consoles sold in the same time frame.

This is an excellent point.

 

[AnarchoPrimitiv]

Something doesn't add up here.

Sony were aiming for 10m by March 2021 - already massive numbers for a mere 5 months of availability. This was a rise from 5-6m (https://gamingbolt.com/ps5-will-have-5-to-6-million-units-produced-by-march-2021-report) to 10m (https://www.businessinsider.com/playstation-5-production-doubled-report-2020-7?r=US&IR=T)

That has become 15m at some point it appears, but there was never any leak or suggestion that 15m was ever going to be the quantity produced.

We know TSMC 7nm yields well, and we know that AMD knows how to design a APU on 7nm. This seems unlikely as a source of the problem, unless the only yield issue is hitting the desired clock speed of 2.2+ GHz.

Now if it was being built on Samsung 5nm, then I would understand the yield angle! But it's not. It's not even TSMC 5nm, although that would be more believable.

Agreed, I don't see how AMD is the source of this issue... I have no idea what the contractual agreements are, but wouldn't Sony be more responsible for most of the custom design choices and creating arrangements with TSMC?

 

[Max(IT)]

Lower than 50% yield on 7nm, ***SURE they have***, this is nothing more than marketing ploy maskerading as "FAKE" NEWS.

Earlier this summer report came out that it was over 70% yields, which is much more plausible.

Elaborate marketing ploy to drive pre-orders. If people think there will be a shortage they are much more likely to pre-order and as pre-orders increase it could actually create a shortage at release.

Today we can not trust BIG CORP information on the internet. To many news sites just copy each other so big corps use this network of mindless drone news sites for guerilla marketing as in this case create more pre-orders for PS5 by fabricating a yields problem that nobody actually can verify and thus create a non existing CPU shortage to get more sales.

Since there are a few redundant CUs calculated from the beginning to improve yields, usable socs should be around the following:
90% for the XSX
95% for the PS5 (more since it is smaller and less powerfull than the XSX soc)

View attachment 168762

"Yields aren't really a problem, on tsmc N7 with a defect density of 0.09 defects per square cm (which was the defect rate in july), the XSX soc yields 73% perfect and the PS5 soc is smaller and will have better yields, also most of the 41 defect dies are still going to end up usable for retail too with as many redundant CUs as they're working with. So yeah... yields are fine, for the ps5? even better. I'd estimate 90% yields for the XSX soc on the low end, and 95% for the ps5 SOC. TSMC dont play around."

so basically a reliable source is saying yields are 50% (improving) but YOU are saying they are lying for a not-so-clear commercial plot

Interesting theory, but for the time being I will trust the reliable source unless you don't provide for a better source

 

[Imouto]

The yield problem is likely to be overclocking the GPU to the moon and not meeting the validation requirements.

 

[Deleted member 185088]

Some people are talking about the clocks, the PS5 doesn't lock its clocks, just like modern GPUs when even in a laptop you can see GPUs clocking really high for brief moments, so I don't see the high clocks as an issue given how variable clocks work, probably Sony focused on this high figure to cut down on the deficit they have compared to the Xbox like Microsoft did with the One when they focused on transistor count.

 

[Max(IT)]

Some people are talking about the clocks, the PS5 doesn't lock its clocks, just like modern GPUs when even in a laptop you can see GPUs clocking really high for brief moments, so I don't see the high clocks as an issue given how variable clocks work, probably Sony focused on this high figure to cut down on the deficit they have compared to the Xbox like Microsoft did with the One when they focused on transistor count.

You don’t have to keep the clock for a long time, but you still have to reach it for the chip to be “certified “, so I don’t think your point is valid.

 

[Deleted member 185088]

You don’t have to keep the clock for a long time, but you still have to reach it for the chip to be “certified “, so I don’t think your point is valid.

Indeed, my point is it the requirements are way lower then having a fixed clock speed, plus given the form factor I don't believe they'll push the SoC beyond its limits.

 

[kings]

Sony probably got screwed with the "last minute" OC they had to make, so the difference in TFLOPs wouldn't be so big on paper compared to the Series X.

Even though the GPU clocks are variable, all chips that equip the console must reach 2.23Ghz. Probably many do not achieve this in a stable way, even for brief moments.