Samsung to make 3 nm chips for Nvidia, Qualcomm, IBM, Baidu - KED Global
Samsung Electronics Co., the world’s largest memory chipmaker, will make semiconductors with the industry’s most advanced 3-nanometer process node
www.kedglobal.com
I suppase this might be the node for T243, assuming Drake is Samsung.Samsung to make 3 nm chips for Nvidia, Qualcomm, IBM, Baidu - KED Global
Samsung Electronics Co., the world’s largest memory chipmaker, will make semiconductors with the industry’s most advanced 3-nanometer process nodewww.kedglobal.com
IncredibleI suppase this might be the node for T243, assuming Drake is Samsung.
Edit: if it’s not obvious, my reasoning is:
T210 Erista
T 214 Mariko
T239 Drake
T243 die shrunk Drake.
Another possibility is:I suppase this might be the node for T243, assuming Drake is Samsung.
Edit: if it’s not obvious, my reasoning is:
T210 Erista
T 214 Mariko
T239 Drake
T243 die shrunk Drake.
Is there any indication that Nvidia is working on 6nm TSMC, because it is easier to work with one node at a time with these mass production chips with hard deadlines, Nvidia is already using a "custom" 5nm process called 4N, customizing another node might not make sense with the low costs 4N seems to be, thanks to efficiency of the process node being much higher than expectations.Another possibility is:
- T239 is Drake fabricated using TSMC's N6 process node
- T243 is Drake die shrunk from TSMC's N6 process node to TSMC's 4N process node, especially since Thor's probably going to be fabricated using TSMC's 4N process node, like with the rest of the Ada Lovelace GPUs (AD106 and AD107 are rumoured to be fabricated using TSMC's 4N process node)
Well TSMC's shifting customers from TSMC's N7 process node to TSMC's N6 process node. And Nvidia did mention that a couple of datacentre chips that are at least planned to be available this year (e.g. BlueField-3, Quantum-2, ConnectX-7) are fabricated using TSMC's 7N process node.Is there any indication that Nvidia is working on 6nm TSMC, because it is easier to work with one node at a time with these mass production chips with hard deadlines, Nvidia is already using a "custom" 5nm process called 4N, customizing another node might not make sense with the low costs 4N seems to be, thanks to efficiency of the process node being much higher than expectations.
Incredible
Sounds better. But what would Nividia seriously use 3nm Samsung for? Nvidia is already using 4nm tsmc for the 4000 series, which I imagine the node is gonna have better performance and power draw over 3nm Samsung, or equal at best.Another possibility is:
- T239 is Drake fabricated using TSMC's N6 process node
- T243 is Drake die shrunk from TSMC's N6 process node to TSMC's 4N process node, especially since Thor's probably going to be fabricated using TSMC's 4N process node, like with the rest of the Ada Lovelace GPUs (AD106 and AD107 are rumoured to be fabricated using TSMC's 4N process node)
there's always low end Lovelace cardsThat would mean Drake would have to be using a Samsung node. No thanks!
Not that I can do anything about it..
Sounds better. But what would Nividia seriously use 3nm Samsung for? Nvidia is already using 4nm tsmc for the 4000 series, which I imagine the node is gonna have better performance and power draw over 3nm Samsung, or equal at best.
Hmm Tegra Thor actually make sense for it.
That might not be true, Samsung 3nm is going to "gate all around" a more complex, but next step in transistors to help with power leak. TSMC isn't picking this up until 2nm, and considering Samsung is probably desperate, they are likely offering a better price than TSMC's 2nm, even if it won't be as good, it should have a lot of savings to gain from GAA tech, so 3nm Samsung could potentially beat 4nm TSMC.That would mean Drake would have to be using a Samsung node. No thanks!
Not that I can do anything about it..
Sounds better. But what would Nividia seriously use 3nm Samsung for? Nvidia is already using 4nm tsmc for the 4000 series, which I imagine the node is gonna have better performance and power draw over 3nm Samsung, or equal at best.
Hmm Tegra Thor actually make sense for it.
Perhaps entry level GPUs for the next GPU architecture being announced in 2024, similar to what Nvidia did with Pascal (high-end and mid-range GPUs on TSMC's 16 nm** process node and entry-level GPUs on Samsung's 14 nm** process node)?Sounds better. But what would Nividia seriously use 3nm Samsung for? Nvidia is already using 4nm tsmc for the 4000 series, which I imagine the node is gonna have better performance and power draw over 3nm Samsung, or equal at best.
Hmm Tegra Thor actually make sense for it.
it's not a deal breaker. being on Samsung's nodes means it would consume more power than if it was on TSMC's node. if it was on 8nm, it's a toss up, the clocks might be on the low end, but it'd still hit close to PS4, you'd be sacrificing battery life though.I’ve gotten the impression that people like TSMC nodes better than Samsung. Is this a deal breaker when it comes to Drake’s TFLOPs? I mean, worst case scenario it’s done on the same Samsung 8nm process that Orin is, but is that really so bad? It’ll still be pretty powerful, right?
Ofcourse it will change performance.it's not a deal breaker. being on Samsung's nodes means it would consume more power than if it was on TSMC's node. if it was on 8nm, it's a toss up, the clocks might be on the low end, but it'd still hit close to PS4, you'd be sacrificing battery life though.
the performance would be the first target to hit, regardless of node, so whatever node the SoC is on, it won't change the performance because that was already decided
Team launch with TOTK.Are you guys Team 2023 or Team 2024 when it comes to new Switch hardware?
Yep, same here.Team launch with TOTK.
Team 23Are you guys Team 2023 or Team 2024 when it comes to new Switch hardware?
It wouldn’t really be close to the PS4 on the 8nm, unless you mean docked.it's not a deal breaker. being on Samsung's nodes means it would consume more power than if it was on TSMC's node. if it was on 8nm, it's a toss up, the clocks might be on the low end, but it'd still hit close to PS4, you'd be sacrificing battery life though.
the performance would be the first target to hit, regardless of node, so whatever node the SoC is on, it won't change the performance because that was already decided
Same. I want it to launch next year or I will turn into Terminator Model T-239 "Yuh".I'm part of the "for the love of god can we get more leaks I want to believe in early 2023 so bad and I don't want to be hurt again" team
I think ILikeFeet is correct here. "Sweet spot" for Nintendo won't be "a good balance of Performance and battery life" it will be "the maximum battery life that gets them over their performance line."Ofcourse it will change performance.
They will target a performance/ battery life sweet spot, and where that spot is depends on the node. On 8nm, it might be lower than OG Switch, which is why nobody saw 12sm coming while we were assuming 8nm.
Samsung 8nm gets a lot of crap, but it's not a bad node. DUV is an older technology for making chips, EUV is a newer one. Chip designers need to rebuild designs from scratch to use EUV. There is only one company in the world that makes EUV machines, and TSMC bought the first two, rapidly creating nodes that combine EUV and DUV tech.I’ve gotten the impression that people like TSMC nodes better than Samsung. Is this a deal breaker when it comes to Drake’s TFLOPs? I mean, worst case scenario it’s done on the same Samsung 8nm process that Orin is, but is that really so bad? It’ll still be pretty powerful, right?
So my very tentative analysis suggests that, when matched at the same clocks and the same number of SMs/CUs, Ampere is ~1.8x more powerful than GCN4.It wouldn’t really be close to the PS4 on the 8nm, unless you mean docked.
The power consumption gets really high on that node.
Team end of 2023 unless there is absolutely no whispers from anyone reputable(not Natedrake) in the first three months hearing it launch in 2023Are you guys Team 2023 or Team 2024 when it comes to new Switch hardware?
According to the Nvidia power estimator for Orin, 1.228tflops + 1.5GHz CPU would consume about 7 watts on average, means they could get the whole system under 10 watts in portable. Original Switch drew 9 watts, and I'd suggest a 5000mah battery minimum, which brings battery life to 2.5 hours. Modern architecture, 720p resolution and DLSS would easily push the device past PS4 when portable.It wouldn’t really be close to the PS4 on the 8nm, unless you mean docked.
The power consumption gets really high on that node.
STILLYep, same here.
Based on what they did with Mariko?I think ILikeFeet is correct here. "Sweet spot" for Nintendo won't be "a good balance of Performance and battery life" it will be "the maximum battery life that gets them over their performance line."
A "better" process node would not result in a more powerful console, it would almost definitely result in a longer playing one.
To add to this, it's more likely that they would move to a better node than redesign the SoC.
Ohhh so this is why people are saying that Samsung 8nm is a “dead end” node? Look at that. I’m learning things!Because Samsung 8nm is the most advanced DUV node on the market, if you have a DUV design, and you want to make it smaller in the future (a node shrink, which can give you big power/cost savings) you don't have a place to go. This is one of the reasons several folk here want Drake to be on a TSMC node, not on Samsung 8nm..
Considering hardware exists since at least April this year, there is no chance that it has changed since the Nvidia hack which was only 2 weeks old when it was released to public on March 1st this year.I’m curious, let’s say late 2022 was originally a target floated for launch - how much can change given 6-9 months additional time - is that enough to make a material difference on the product that hits the market? Would moving to a better node be in the cards?
It’s not like we didn’t have some evidence that 2022 was possible. Mochizuki’s reporting on developers targeting late 2022 is often quoted. Tears of the Kingdom, a game that any sensible decision making would see launch with better hardware, was originally was slated for 2022. Recent release performance from first party titles Bayonetta and Scarlet/Violet have some in the media talking about needing better hardware. Kat Bailey on NVC posited that these games (among others) may have genuinely been targeting better hardware this year and that something in the background changed.
Based on past Nintendo behavior across their entire handheld line, and reports out of the development of the device, which have already suggested a late 2021 rethink to get more battery life out of Drake. See also: every mobile SOC ever made. Nintendo tried to hold on to that 300 Mhz clock as long as possible in handheld mode, after all. And had they launched with 16nm, I feel fairly confident they would have found similar power levels in favor of the much expanded battery life.Based on what they did with Mariko?
I think there were a lot of other reasons they didn't clock Mariko higher than base units. I think if they had launched on 16nm, they would definitely have had higher base clocks imo.
Considering hardware exists since at least April this year, there is no chance that it has changed since the Nvidia hack which was only 2 weeks old when it was released to public on March 1st this year.
What?which have already suggested a late 2021 rethink to get more battery life out of Drake.
The problem is that the delay would be before engineering samples and release candidates. Basically spring/summer of last year, not this year, a year later they are in production with at least engineering samples and right now they are ready for mass production or they wouldn't be updating the public Linux kernal.So if a delay did happen this year, production concerns seems more likely, not because they had an opportunity to make things any better. Right?
What?What?
I don't agree with you. Nintendo sells updated models of their hardware with focus on efficiency and feature updates, it's the Drake Switch v2 that would have the better battery life, not the initial launch model. They are only looking to clearly beat steam deck's low battery life, and thanks to ARM, it should do it no problem.Based on past Nintendo behavior across their entire handheld line, and reports out of the development of the device, which have already suggested a late 2021 rethink to get more battery life out of Drake. See also: every mobile SOC ever made. Nintendo tried to hold on to that 300 Mhz clock as long as possible in handheld mode, after all. And had they launched with 16nm, I feel fairly confident they would have found similar power levels in favor of the much expanded battery life.
I am pretty certain the conversation has been "how much battery life can we get for ~PS4 perf" not "how much power can we get for 2 hours of battery life"
The part about the 2021 rethink for battery life.What?
Because Samsung 8nm is the most advanced DUV node on the market, if you have a DUV design, and you want to make it smaller in the future (a node shrink, which can give you big power/cost savings) you don't have a place to go.
In November of last year, Nate said that he had heard that battery impact of RT was higher in handheld mode than originally estimated when building devkits. This report essentially matched up with several pieces of information that were dropped by a user who has since signed an NDA and deleted their old posts. I am reluctant to drive attention to that person, but I am sure you can find the conversation if you dig. Many of this person's comments on the device were confirmed by the much later Nvidia leak.The part about the 2021 rethink for battery life.
Ah I stand corrected! You are right, though the path out of Samsung 8nm is still limited by embedded Samsung IP, requiring a redesign regardless.I don’t think this is true. There is DUV TSMC 7nm. And the EUV usage on other 7nm SKUs is limited to a few layers.
The solution is to not use RT in portable mode, which is what was being said at the time by those conversations.In November of last year, Nate said that he had heard that battery impact of RT was higher in handheld mode than originally estimated when building devkits. This report essentially matched up with several pieces of information that were dropped by a user who has since signed an NDA and deleted their old posts. I am reluctant to drive attention to that person, but I am sure you can find the conversation if you dig. Many of this person's comments on the device were confirmed by the much later Nvidia leak.
They will start at 5nm and move to 3nm for a lite 3 years later.Ah I stand corrected! You are right, though the path out of Samsung 8nm is still limited by embedded Samsung IP, requiring a redesign regardless.
Team launch with TOTK.
My first and only time. Let’s enjoy this moment.You are right.
Couldn't that kinda cause compatibility issues with some games? Or does every game that uses dedicated RT hardware have a fallback option? I'm not too familiar with RT games.The solution is to not use RT in portable mode, which is what was being said at the time by those conversations.