kimbo99
Spirit Detective
- Pronouns
- He/him
I actually know someone who works at Zynga and they only heard that a Nintendo Switch version and mobile version were being made. Perhaps the Switch moniker covers all versions(OG + Drake)
-
Hey everyone, staff have documented a list of banned content and subject matter that we feel are not consistent with site values, and don't make sense to host discussion of on Famiboards. This list (and the relevant reasoning per item) is viewable here.
StarTopic Future Nintendo Hardware & Technology Speculation & Discussion |ST| (Read the staff posts before commenting!)
- Thread starter Dakhil
- Start date
I actually know someone who works at Zynga and they only heard that a Nintendo Switch version and mobile version were being made. Perhaps the Switch moniker covers all versions(OG + Drake)
Dark Cloud
Warpstar Knight
Another L for Nintendo it looks like. I knew the stuff people be daydreaming about was too good to be true.
/s
/s
OP
OP
According to Korean leakers, new Tegra for Switch is Samsung 7nm (not EUV).
닌텐도 후속작에 스냅드래곤이나 심지어 A칩 같은게 달리길 원하시는데 - 미코
그럴일 없다고 하더라구요 닌텐도의 개발도구가 상당히 엔비디아의 NVN api에 의존하는 구조라고 합니다. 올해 초 기사를 보면 NVN2의 유출도 있었다고 하니 아마도 후속도 엔비디아의 머신러닝 연구용으로 나온 AP를...meeco.kr
Tegra for Samsung 7nm (non-EUV) Switch was on the roadmap...
루머)삼성 파운드리와 나름 이야기가 오고가는 기업들 - 미코
현재 삼성 파운드리와 이야기가 오고가고 있는 회사들입니다. 퀄컴,엔비디아,AMD,AWS,테슬라,IBM 유명 메이커는 요렇게입니다. 3nmGAP 공정으로 이야기 오고 가는게 퀄컴 4nmLPH 공정으로 이야기가 오고 가는게 엔비...
And in the case of Nvidia's new Tegra, it seems that it will be produced with Samsung's 7nm. It seems that it is not EUV.
Edit: I don't know if he is reliable, but the following informational website presents 흡혈귀왕 (vampire king) as an informed person.
Google Tensor G3のCPUの構成が判明、Cortex-X2を飛ばしてCortex-X3を採用予定
Pixel 8とPixel 8 Proが搭載するGoogle Tensor G3は順当にいけば来年の10月に仕様が判明する見込みですが、注目している人が多い …reameizu.com
As far as I know, Samsung's 7 nm** process node has been using EUV lithography, not DUV lithography, from the beginning (here and here).
The only way Samsung's 7 nm** (DUV) process node being used could make some sense is if the same 7 nm** process node used to fabricate the Exynos 9825 is the one being mentioned there, which uses EUV lithography, but is fundamentally not very different from Samsung's 8LPP process node.
Anyway, the only information I can find about Samsung's 7LPP process node with respect to performance and power efficiency is that Samsung's 7LPP process node is ~20% to ~30% less power efficient compared to TSMC's N7P process node (and probably TSMC's N6 process node since TSMC doesn't promise performance and power efficiency improvements compared to TSMC's N7 process node, only 18% higher logic density compared to TSMC's N7 process node thanks to EUV lithography being used, with TSMC's N6 process node).
** → a marketing nomenclature used by all foundry companies
- Pronouns
- He/Him
Them saying “Drake doesn’t exist” is the same thing as them saying “Drake is on SEC 7 DUV”
I’m pretty sure what they are referring to is the 8N node that is custom for Nvidia only. There was an article about “ORIN X” which they listed it as having 7nm from Samsung. However, they since retracted that and changed it to 8N. It wasn’t the first one to do so either with regards to ORIN. And since Drake is a spin-off-not-canon-but-actually-canon-timeline chip custom for Nintendo from the ORIN chip, it is still pretty likely to be manufactured at SEC 8N regardless of what we say here as a result as the progenitor (ORIN) exists there after all. Note: I say likely, not absolute.
But that assumes the design exists twice, not once, like Qualcomm 8Gen1 and 8+Gen1. Former is at SEC 5nm fab, the latter was on the TSMC 5nm fab and came out like less than a year or two after the former.
I’m pretty sure what they are referring to is the 8N node that is custom for Nvidia only. There was an article about “ORIN X” which they listed it as having 7nm from Samsung. However, they since retracted that and changed it to 8N. It wasn’t the first one to do so either with regards to ORIN. And since Drake is a spin-off-not-canon-but-actually-canon-timeline chip custom for Nintendo from the ORIN chip, it is still pretty likely to be manufactured at SEC 8N regardless of what we say here as a result as the progenitor (ORIN) exists there after all. Note: I say likely, not absolute.
But that assumes the design exists twice, not once, like Qualcomm 8Gen1 and 8+Gen1. Former is at SEC 5nm fab, the latter was on the TSMC 5nm fab and came out like less than a year or two after the former.
OP
OP
There was an article from the Korea Herald back in 2 July 2019 about Nvidia Korea's chief, Yoo Eung-joon, mentioning that Nvidia's planning on using Samsung's 7 nm** process node using EUV lithography for consumer Ampere GPUs, which ultimately never materalised.
But otherwise, I totally agree.
** → a marketing nomenclature used by all foundry companies
AxkilAvenger
Like Like
- Pronouns
- He/him
Rumour was Zynga had a dev kit and this is the only switch game they are known to be working on
SiG
Chain Chomp
I know you put an "/s" here, but you can never be too sure nowdays.
"So 8nm Samsung > 7nm Samsung? As in better performance?"
- Pronouns
- He/Him
Think I picked out something from these tweets.
"TSMC believes N7/N6 underutilization is a temp issue. Demand will pick up again in 2H23."
What could be happening that these will become better utilized in the latter half of 2023?
... or it could mean nothing relevant to this topic.
IMO, performance isn't too important if the performance characteristics of an 8nm Drake and 7nm Drake are the same. 7nm would yield them lower power draw and area, which is what Nintendo wants. And unlike the Exynos, it's not held back by Samsung's bad chip design
- Pronouns
- He/Him
I meant performance of the node itself, which leads to slightly better battery life, cost or game performance of the console. Likely more the former.
OP
OP
I still think TSMC's N6 process node being used to fabricate Drake is still a possibility.Think I picked out something from these tweets.
"TSMC believes N7/N6 underutilization is a temp issue. Demand will pick up again in 2H23."
What could be happening that these will become better utilized in the latter half of 2023?
... or it could mean nothing relevant to this topic.
But are the yields good? Considering that the Exynos 990 bins were bad, and nobody outside of IBM (here and here) is currently using Samsung's 7 nm** process node, I presume yields for Samsung's 7 nm** process node are bad.
** → a marketing nomenclature used by all foundry companies
- Pronouns
- He/Him
I think the worry over the node in general is getting kinda silly. We know the specs more or less, the node will basically determine how high on the +-30% performance/battery range it'll go. It doesn't make a huge difference.
They abandoned it as soon as possible which is all we need to know. If Nvidia settled on Sammy 7nm in 2019 but Samsung abandoned it shortly after, Samsung is probably gonna eat that cost I think. Another "pay per working chip" which will hurt their bottom line. If I was Samsung, I'd move them to 5nm and eat some of that costs in the short term
- Pronouns
- He/Him
30% is a huge difference
It’s not insignificant I should say.
5-10? Sure that can be dealt with, 20-30? That’s a noticeable difference.
Last edited:
- Pronouns
- He/Him
Noticeable sure, but I wouldn't say huge. Especially since it'll likely be more on the battery life side.
- Pronouns
- He/Him
I’ll put it this way, 20 to 30% can be a difference between 2H and 15 minutes and a full three hours of gameplay. In the more extreme scenario.
OP
OP
No, the difference is much larger than 30%.
But Samsung's 7LPP process can be ~30% more performant than Samsung's 8LPP process node. (I presume for certain applications and in the best case scenario.)
Makes sense.
kimbo99
Spirit Detective
- Pronouns
- He/him
Yup, I'm aware. And my contact isn't exactly a developer so take with a grain of salt of course.
D
Deleted member 887
Guest
I'm not worried about the process node. Hell, I'm not worried about anything related to video games, my concerns are mostly the rise of fascism in my native country.
But I think there are two things to consider with regard to the process. As @ReddDreadtheLead likes to remind us, Samsung 8nm is an architectural dead end. It would be nice if a die shrink was a clear and easy future win, rather than a significant reengineering cost.
And, secondly, I have some lingering concerns about clocks. I know we've all poured over the power numbers for Orin, but even the fairly optimistic analyses make Erista/Mariko clocks seem like a tight fit. If Drake is on literally anything better than 8nm, then I would expect clocks in handheld mode to be roughly the same between devices. If it is 8nm, then I think that running lower clocks outside of BC mode is a possibility. It's not just about the marginal performance gains, it's about what set of trade offs are on the table between BC, performance, and battery life.
But I think there are two things to consider with regard to the process. As @ReddDreadtheLead likes to remind us, Samsung 8nm is an architectural dead end. It would be nice if a die shrink was a clear and easy future win, rather than a significant reengineering cost.
And, secondly, I have some lingering concerns about clocks. I know we've all poured over the power numbers for Orin, but even the fairly optimistic analyses make Erista/Mariko clocks seem like a tight fit. If Drake is on literally anything better than 8nm, then I would expect clocks in handheld mode to be roughly the same between devices. If it is 8nm, then I think that running lower clocks outside of BC mode is a possibility. It's not just about the marginal performance gains, it's about what set of trade offs are on the table between BC, performance, and battery life.
RennanNT
Bob-omb
I agree. A better process node gives better performance or better efficiently or a bit of both, but it's also more expensive and potentially has lower availability.
And so, it's unknown how much performance and/or efficiency and/or costs they would repass to the end user, and also how much they would use the efficiency gains to cut down elsewhere (like smaller battery and heat sink).
At the end of the day, Drake running at what seems to be the lowest clocks (which makes sense before energy savings are negligible) are still in the same ballpark of majority of the bullish expectations for a 5nm Drake.
LiC
Member
Seriously. Don't people get tired of the same panic after every one of these (non-credible) rumors, or speculative posts about whether 8N makes sense or not? It's trade-offs all the way down, and I'm sorry but Nintendo is definitely going to make some of those trade-offs on the side of cost and manufacturing. And it will be fine. You're fine. Breathe.
ShaunSwitch
Moblin
This is where I am, it's not a worry, it's more confusion as to why such a large chip would be used and then clocked into the ground on Samsung 8nm vs just using a smaller chip and clocking higher (which would be more cost effective.)
Absolutely, but those trade offs have to make sense. Its not panic, its more just a case of trying to understand the most logical outcome from what we know. At the moment I believe that given they have gone for such a large GPU and 8 CPU Cores that in order to facilitate this on Samsung 8nm this design would be far from the optimal choice for a mobile product.
Kevin
Chain Chomp
I wouldn't be surprised if Nintendo uses a really old and power inefficient node. Nintendo has a history for a long time now with going with the cheapest options. They would choose an old node over a new one if it saved them a dollar lol.
Has anyone worked up any math as to what we should expect in terms of performance if this node information is accurate? I guess my expectations are maybe around and old Xbox One with DLSS.
Has anyone worked up any math as to what we should expect in terms of performance if this node information is accurate? I guess my expectations are maybe around and old Xbox One with DLSS.
kvetcha
hoopy frood
- Pronouns
- he/him
As has been discussed, newer nodes may in fact end up being cheaper per SOC simply because you can fit more chips on each wafer, so even if the per-wafer cost is higher you end up spending less per device.
Kevin
Chain Chomp
Fair enough. Then my guess is that the smaller nodes just may not be available to use. Maybe bigger companies have claimed all of the slots. I'm not an expert on this stuff but I did have a feeling Nintendo would use an older node. Guess we shall see.
LiC
Member
This is a decision Nintendo and Nvidia made somewhere between 1 and 2 years ago, which would have been based on Nvidia's existing Ampere GPUs and their plans for the now-current Ampere Tegra, and with only tentative forecasts about the state of chip availability and supply chains and so on. So they're not made in a vacuum or with all the knowledge we have today. It's also possible that the ongoing justifications here for a more advanced process node being optimal are flawed, or that our definition of "optimal" is not the same as Nintendo's.
I personally believe 8N is as likely as anything else, because most of the (technical) attempts here to un-explain it are based on estimations for Orin, but Orin isn't designed for gaming or for power consumption, while Drake is being explicitly designed for both. There's more than enough uncertainty to go from insufficient/borderline power draw to good enough/decent. It could turn out to be a more advanced node, sure, but 8N is neither off the table nor a disaster to be feared.
D
Deleted member 887
Guest
In defense of @LiC there often is an over the top response to node information from some of the less regular posters.
The "because Nintendo" move is baked in already - Drake is riding Orin's coattails, and Nintendo is effectively paying only a fraction of the full chip engineering cost because of it. All the tradeoffs that Drake will make should be viewed in that light. Staying on Samsung wouldn't about using the cheaper node, it's about being able to reuse Orin's design, rather than start from scratch. It's not "make an optimal 8nm device for Switch" it's "take existing 8nm device, and make it the best Switch possible."
As to "why is Drake so big" I assume it's because it can be. Nintendo had a performance minimum, a power draw maximum, and I assume the basic process was "Start with Orin, and remove cores and SMs till we get below 10W. If we can't do that while staying over our perf minimum, change strategy. If we can, start tuning - swap AE cores for C cores, keep or replace Orin's tensor cores, where exactly should we set the clocks?"
Every indication is that Drake is closely related to Orin, so I don't think the "strategy change" path ever got taken. If Hovi* decided to move off of Samsung 8nm, it has to be cost effective. I would expect the path out would be that A78C already supports TSMC 7nm, and the first Ampere GPUs were on that node, leaving at least some of the engineering work already done.
The biggest consideration this thread has given, collectively, to the node change has just been that when we saw hard numbers for SMs come out of the Lapsus$ hack, it just seemed too big for Samsung 8nm. There was the collective question of how do we fit this square peg into this round hole.Some later analysis has maybe said "actually, it might be just barely doable" but just barely doable is good enough without spending all the money to start a fresh design on TSMC. In other words the peg might just be round after all.
Why is Drake big? Because Orin is big. Why Samsung? Because Orin is on Samsung. And my guess is that for all the questions about "what about after Drake" the answer is "Thor is coming in 2025"
* Hovi: Horizon Nvidia, used internally by Nvidia to refer to Nintendo, or the Nvida/Nintendo group that is working on Drake.
Kise Ryota
Bob-omb
I was reading the comparison between the 8 gen 1 (Samsung 4nm) and 8 plus gen 1 (tsmc N4) and the gains in power efficiency is 30% for the latter (and the former had a 30% gain VS the SD 888 plus (Samsung 5nm), but then it's a different chip...)
So I'm imagining that coming from 8LPP to TSMC N4 is something between 80%~100% gain in power efficiency (which doesn't translate to 1:1 in performance gains)
That would be a huge difference in battery life, which I would appreciate a lot. But, like I always say, "expect the worse and you won't be disappointed"...
So I'm imagining that coming from 8LPP to TSMC N4 is something between 80%~100% gain in power efficiency (which doesn't translate to 1:1 in performance gains)
That would be a huge difference in battery life, which I would appreciate a lot. But, like I always say, "expect the worse and you won't be disappointed"...
- Pronouns
- He/Him
There has been nothing that has or hasn't happened to suggest anything different from what we've been thinking for the past year.
In other words, if you've been expecting H1 2023 since August like I have then there's no reason to change your expectation right now.
D
Deleted member 887
Guest
I'm still thinking the same. No new info has come out to push it, and the broad assumption was always "Nintendo could announce before the Holidays for a 2023 release but probably won't" so I don't see the absence of data as a problem
Concernt
Optimism is non-negotiable
- Pronouns
- She/Her
I'm still expecting Q1 2023 and will continue to do so until April 1st.
no, there's a surplus of top end nodes. to the point where companies wanted a refund
Kise Ryota
Bob-omb
Like others said, nothing has really changed (no new info). I continue with my 2024 prediction lol
Look over there
Bob-omb
Well, late to the Samsung 7 topic, but I did want to ask/confirm something: by now, '7LPP' should've been superseded by 5LPE/5LPP, right? (Samsung's 5LPE/LPP are further refinements of the 7LPP->6LPP* line)
*not that 6LPP has appeared in recent charts...
TSMC N7/N6 demand picking back up in 2H23... my first guess would be Navi33 (RDNA3's lower end cards). Which reminds me, I hate that GPU series rollouts have the lower end coming out way the hell later.
N3E ramp being pulled in by 2-3 months is a good sign. The importance here is that N3E is basically the 'real' long term foundation for the N3 family. There is no direct migration path from N3 to N3E, so N3 is a dead end. And all the other N3 variants are planned to be offshoots of N3E.
*not that 6LPP has appeared in recent charts...
TSMC N7/N6 demand picking back up in 2H23... my first guess would be Navi33 (RDNA3's lower end cards). Which reminds me, I hate that GPU series rollouts have the lower end coming out way the hell later.
N3E ramp being pulled in by 2-3 months is a good sign. The importance here is that N3E is basically the 'real' long term foundation for the N3 family. There is no direct migration path from N3 to N3E, so N3 is a dead end. And all the other N3 variants are planned to be offshoots of N3E.
OP
OP
D
Deleted member 887
Guest
It uses both, yes? My understanding was that Samsung's early EUV capacity was too low to build entire chips with, and that the majority of layers were still DUV. This is not my wheelhouse, so perhaps that is common knowledge or I am mistaken.
Same could be said for good news.
Look over there
Bob-omb
Random tangent: The densities being achieved with LPDDR5X is blowing my mind. I wasn't 100% with the devblog post on the Grace superchip, but now with the whitepaper out, a maximum of 512 GB LPDDR5X @ 546 GB/s bandwidth is nutty. At 8533 MT/s, that should be 512-bit. One gigabyte of capacity per 1 bit of memory bus width is pretty high. Checking with Samsung's LPDDR5X page, the densities range from 1/8 of a gigabyte (or, 1 gigabit) to 1/2 of a gigabyte of capacity (or, 4 gigabit) per 1 bit of bus width. Which are already nice! I guess that's how 5X has legs until 6's time.
yes. all that cost is just part of developing a chip. if they chose a different node at the start then they don't have to worry about redesigning the chip in the future
Last edited:
