Inside the Xbox One S - SOC 16nm FinFET process (TSMC)

[Bacon1]

http://www.eurogamer.net/articles/digitalfoundry-2016-inside-xbox-one-s-tech-interview

 

[exar333]

Thanks for the link. Interesting read.

 

[PhonakV30]

So Polaris From GF has really Major Issue Yield.i think so.

 

[Kenmitch]

So Polaris From GF has really Major Issue Yield.i think so.

Why would you jump to that conclusion? There really isn't any data to make a logical conclusion either way.

 

[MrTeal]

XBone S isn't the radically new design though, that's Scorpio and it will be coming next year. By all indications they've just taken the existing APU, added 4K support and shrunk it. Since the original was made at TSMC, it would make sense that the refresh would also be earlier GCN and made at TSMC.

Now if Scorpio isn't new GCN4+ and made at GloFo, then I would start to really worry.

 

[Kenmitch]

XBone S isn't the radically new design though, that's Scorpio and it will be coming next year. By all indications they've just taken the existing APU, added 4K support and shrunk it. Since the original was made at TSMC, it would make sense that the refresh would also be earlier GCN and made at TSMC.

Now if Scorpio isn't new GCN4+ and made at GloFo, then I would start to really worry.

Wouldn't the shrink at TSMC be a more cost effective solution to whomever was paying for it in the end.

 

[Stuka87]

So Polaris From GF has really Major Issue Yield.i think so.

How on earth could you come up with this?

This chip isn't Polaris. This is a shrunk version of the same chip, but up clocked some thanks to the lower power consumption.

The previous chip was being built on TSMC's process. Moving it to another process entirely would have cost significantly more than simply shrinking it for the same fab.

 

[Headfoot]

Interesting that they're building in 4k onto the die drink Slim version in addition to doing a full revamp with Scorpio. I would've guessed that they'd only add 4k support to Scorpio

 

[DaveSimmons]

^ it's just for 4K video not gaming though. MS wants to be able to sell people 4K movies and TV.

It's nice to see that the S offers some minor quality bumps on a few existing titles:
http://arstechnica.com/gaming/2016/08/surprise-new-xbox-one-s-gpu-smooths-out-some-older-games/

 

[Mercennarius]

^ it's just for 4K video not gaming though. MS wants to be able to sell people 4K movies and TV.

It's nice to see that the S offers some minor quality bumps on a few existing titles:
http://arstechnica.com/gaming/2016/08/surprise-new-xbox-one-s-gpu-smooths-out-some-older-games/

Partially accurate. While it's aimed for 4K movies, Microsoft has stated that the Xbox One S DOES upscale all games to a 4K resolution. While not the same thing as a native 4K resolution this is an improvement over what the current Xbox One is capable of in games. Not to mention the added clock speed and FPS increases..

 

[Flapdrol1337]

So Polaris From GF has really Major Issue Yield.i think so.

Why? This chip is basically a shrink of a chip already made at TSMC, would make sense to keep it there.

 

[lixlax]

So Polaris From GF has really Major Issue Yield.i think so.

Either that, GF not having enough capacity or OEM's eating up tons of chips.

 

[therealnickdanger]

Partially accurate. While it's aimed for 4K movies, Microsoft has stated that the Xbox One S DOES upscale all games to a 4K resolution. While not the same thing as a native 4K resolution this is an improvement over what the current Xbox One is capable of in games. Not to mention the added clock speed and FPS increases..

Most indications are that Scorpio will likely do 1080p60 and even 1440p60 really well. I'm sure they'll resort to 3200p/4K @ 22fps in the end, however.

 

[Elixer]

This is news in that they are using TSMC, it wasn't a simple die shrink either, it has added features.
This tells us that AMD has people working on both TSMC's & GloFlo's (Sammy's) FinFets, and this does change the big picture of how are things being worked on at AMD.

It was assumed that GloFlo would be the only game in town, since they don't have the man power to have multiple teams working on multiple different processes and this turns out to be false.

If the consoles are going to keep using TSMC and their 16nm FinFets, and 4x0 line uses GloFlo, you can pretty much bet that AMD knows which process is better, and which fab AMD will be using for future products if one process is vastly superior to the other.

 

[thesmokingman]

This is news in that they are using TSMC, it wasn't a simple die shrink either, it has added features.
This tells us that AMD has people working on both TSMC's & GloFlo's (Sammy's) FinFets, and this does change the big picture of how are things being worked on at AMD.

It was assumed that GloFlo would be the only game in town, since they don't have the man power to have multiple teams working on multiple different processes and this turns out to be false.

If the consoles are going to keep using TSMC and their 16nm FinFets, and 4x0 line uses GloFlo, you can pretty much bet that AMD knows which process is better, and which fab AMD will be using for future products if one process is vastly superior to the other.

Interesting points. This could really give them some flexibility.

 

[Qwertilot]

Either that, GF not having enough capacity or OEM's eating up tons of chips.

GF would, for something like this, surely have Samsung acting as very effective insurance cf capacity? So I don't think its that so much.

 

[Elixer]

Why? This chip is basically a shrink of a chip already made at TSMC, would make sense to keep it there.

Going 16nm FinFet (or 14nm) is NOT a simple "shrink" job. It is a new design from the bottom up.
While they can use the same logic from the older chips, actually designing said chip is very complex, and just because it was made at that foundry before doesn't help much, if at all.

This could also explain why Vega didn't launch yet, they were using the design team to fulfill their SoC orders...

 

[darkswordsman17]

This is news in that they are using TSMC, it wasn't a simple die shrink either, it has added features.
This tells us that AMD has people working on both TSMC's & GloFlo's (Sammy's) FinFets, and this does change the big picture of how are things being worked on at AMD.

It was assumed that GloFlo would be the only game in town, since they don't have the man power to have multiple teams working on multiple different processes and this turns out to be false.

If the consoles are going to keep using TSMC and their 16nm FinFets, and 4x0 line uses GloFlo, you can pretty much bet that AMD knows which process is better, and which fab AMD will be using for future products if one process is vastly superior to the other.

If I remember right, AMD said that they had different teams working at each foundry. Polaris was tied to 14nm and Vega I don't know if they said but it seemed like it was hinted that it would be done at TSMC 16nm. But I'm pretty sure they've openly said that they have good working relationships with both GF and TSMC and that wouldn't be changing and that they'd be willing to fab at either if any company wanted that. I'd wager that is what is driving this more than anything is Microsoft had a production deal with TSMC (basically they would get a certain amount of production as long as Microsoft fulfilled it).

Going 16nm FinFet (or 14nm) is NOT a simple "shrink" job. It is a new design from the bottom up.
While they can use the same logic from the older chips, actually designing said chip is very complex, and just because it was made at that foundry before doesn't help much, if at all.

This could also explain why Vega didn't launch yet, they were using the design team to fulfill their SoC orders...

That's why I'm surprised that they actually did a die shrink, as I don't know that it offered that much over the 28nm design (they could add the 4K scaling ability pretty easily, and I'm sure they could've gotten the clock speed boost as well) for the cost of the rework. But its probably a good way to get experience with the new process before doing a big die.

 

[Flash831]

If I could have a guess at why they have been using GF for the GPU's and TSMC for their semicustom SoC's, I would say it has something to do with AMD's wafer supply deals.

AMD has a wafer supply agreement, take-or-pay. That means they get their wafers, no matter the quality. Because both FinFET-nodes are still pretty immature, I guess the quality (yields) are quite varying.
Therefore the GF node is a good choice for the products where AMD can bin the chips differently to make different products (GPU's).

With TSMC, AMD has a per chip agreement, so they pay for working chips. Therefore the quality of these chips will be more consistent, which therefore should be better for the semi-custom products, as AMD cannot sell the canned chips elsewhere.

 

[NTMBK]

If I could have a guess at why they have been using GF for the GPU's and TSMC for their semicustom SoC's, I would say it has something to do with AMD's wafer supply deals.

AMD has a wafer supply agreement, take-or-pay. That means they get their wafers, no matter the quality. Because both FinFET-nodes are still pretty immature, I guess the quality (yields) are quite varying.
Therefore the GF node is a good choice for the products where AMD can bin the chips differently to make different products (GPU's).

With TSMC, AMD has a per chip agreement, so they pay for working chips. Therefore the quality of these chips will be more consistent, which therefore should be better for the semi-custom products, as AMD cannot sell the canned chips elsewhere.

I thought per-wafer was the norm for TSMC these days, after they got burned with Fermi? Prevents the customer designing a nigh-impossible-to-manufacture chip, and forcing you to eat the costs.

 

[Elixer]

Yeah, that is what I have read as well, it is per-wafer at all places these days.
If it was per-chip, that would be a killer deal.