Steve
Koopa
- Pronouns
- he/him
We are so back NintendobrosWake up TSMC 4N believers! Nintendo can now use TSMC 4N because it's an outdated node now
(also is that from the current Nvidia livestream)
-
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| (New Staff Post, Please read)
- Thread starter Dakhil
- Start date
(also is that from the current Nvidia livestream)
Kise Ryota
Then I'll say "I told ya"
Interestingly, Blackwell has a decompression engine. Developed at the same time as Drake's or were they developed separately?
I saw that, and then I was thinking if it was the first decompression engine they made for these type of systems. So I googled it and the NVIDIA BLUEFIELD-3 also has a decompression engine
edit: actually, if I understood it correctly, the decompression engine on the blackwell system would be from the BlueField-3 (blackwell is like many technologies in one crazy system)
It was a jaw-dropping presentation for me, wow
Last edited:
BreakAtmo
Bob-omb
- Pronouns
- He/Him
Wake up TSMC 4N believers! Nintendo can now use TSMC 4N because it's an outdated node now
Edit: For context, next-gen DCAI parts from Nvidia are using TSMC 4NP, an evolution of 4N. Probably a further customized 4N with N4P enhancements.
Now I'm hoping for a 4NP Drake? See what you've done?!
More seriously, something I wanted to know - do "refinement nodes" of this kind that are essentially still just improved versions of the primary node (in this case TSMC 5nm) actually cost more money and see separate pricing for N5, N4, 4N, N4P, 4NP and N4X? Or are they just "this is what 5nm customers get now, for the same money per wafer"?
SiG
Chain Chomp
Where'd the video go?
Angel Whispers
Follow me at @angelpagecdlmhq
If you’re still using the line graphs and not posting Splatoon 3’s “We’re So Back”, you’re doing it wrong, tbqh. 
OP
OP
ItsukaKasuga
Piranha Plant
- Pronouns
- he/him
From what i understand, nope, this only give us a hint of possible technology for future proyects of Nvidia but nothing elseSo... did the Nvidia presentation tell us something, even maybe a slight hint because the only thing of value would maybe be this.
Angel Whispers
Follow me at @angelpagecdlmhq
Was that a confirmation, or was it not the case before?
I see wikipedia articles still saying Lovelace for DRIVE Thor.
OP
OP
NVIDIA Unveils DRIVE Thor — Centralized Car Computer Unifying Cluster, Infotainment, Automated Driving, and Parking in a Single, Cost-Saving System
NVIDIA today introduced NVIDIA DRIVE™ Thor, its next-generation centralized computer for safe and secure autonomous vehicles.
nvidianews.nvidia.com
NVIDIA DRIVE Powers Next Generation of Transportation — From Cars and Trucks to Robotaxis and Autonomous Delivery Vehicles
NVIDIA today announced that leading companies across the transportation sector have adopted the NVIDIA DRIVE Thor™ centralized car computer to power their next-generation consumer and commercial fleets — from new energy vehicles and trucks to robotaxis, robobuses and last-mile autonomous...
nvidianews.nvidia.com
navierStokes
Shriekbat
I think it’s a trade-off. I presume there’s additional costs associated with it, as that not all their machines are validated to produce chips with the refined process. but the higher efficiency and any reduction of chip size will be benificial. Thus you weigh against any reservation costs to make the chips on that latest process.
Given that NVIDIA worked specifically on a “custom version” and they’re a key customer, I think they’ve already been aware of the refinements TSMC were making and any chip designed for that was done already, so perhaps the announcement lies in some optimal volume per quarter of chips and machines they can deliver.
I’m curious what the exact differences are with 4NP, Anandtech published a table hehe;
NVIDIA Flagship Accelerator Specification Comparison | |||
B200 | H100 | A100 (80GB) | |
| FP32 CUDA Cores | A Whole Lot | 16896 | 6912 |
| Tensor Cores | As Many As Possible | 528 | 432 |
| Boost Clock | To The Moon | 1.98GHz | 1.41GHz |
| Memory Clock | 8Gbps HBM3E | 5.23Gbps HBM3 | 3.2Gbps HBM2e |
| Memory Bus Width | 2x 4096-bit | 5120-bit | 5120-bit |
| Memory Bandwidth | 8TB/sec | 3.35TB/sec | 2TB/sec |
| VRAM | 192GB (2x 96GB) | 80GB | 80GB |
| FP32 Vector | ? TFLOPS | 67 TFLOPS | 19.5 TFLOPS |
| FP64 Vector | ? TFLOPS | 34 TFLOPS | 9.7 TFLOPS (1/2 FP32 rate) |
| FP4 Tensor | 9 PFLOPS | N/A | N/A |
| INT8/FP8 Tensor | 4500 T(FL)OPS | 1980 TOPS | 624 TOPS |
| FP16 Tensor | 2250 TFLOPS | 990 TFLOPS | 312 TFLOPS |
| TF32 Tensor | 1100 TFLOPS | 495 TFLOPS | 156 TFLOPS |
| FP64 Tensor | 40 TFLOPS | 67 TFLOPS | 19.5 TFLOPS |
| Interconnect | NVLink 5 ? Links (1800GB/sec) | NVLink 4 18 Links (900GB/sec) | NVLink 3 12 Links (600GB/sec) |
| GPU | "Blackwell GPU" | GH100 (814mm2) | GA100 (826mm2) |
| Transistor Count | 208B (2x104B) | 80B | 54.2B |
| TDP | 1000W | 700W | 400W |
| Manufacturing Process | TSMC 4NP | TSMC 4N | TSMC 7N |
| Interface | SXM | SXM5 | SXM4 |
| Architecture | Blackwell | Hopper | Ampere |
Nontheless, for Drake,what this all means for drake. imo, that NVIDIA is shifting its big orders for this new chip and Hopper will be de-allocated. I assume that NVIDIA will not continue to make orders for Hopper.
Thus giving transfer for Nintendo to buy out the reserved booking from NVIDIA at a better cost than before. Idk if that’s how it works haha, but that’s what I think.
Last edited:
KeijiKG
Moblin
Mario games never have that much bloat
I'd be surprised if a Mario game is more than 13 gigs on the next system because they're just that good with resource management, 4K might challenge that but I have it in good faith Nintendo wouldn't mess that up.
Aren't those tflops weak? 3.6 fp8 means only 0.9 fp32.Wake up TSMC 4N believers! Nintendo can now use TSMC 4N because it's an outdated node now
Edit: For context, next-gen DCAI parts from Nvidia are using TSMC 4NP, an evolution of 4N. Probably a further customized 4N with N4P enhancements.
Unless I'm misunderstanding something? I wonder why that number with no mention of clocks. Is that max? I'm all over the place now
Kise Ryota
Then I'll say "I told ya"
Aren't those tflops weak? 3.6 fp8 means only 0.9 fp32.
Unless I'm misunderstanding something? I wonder why that number with no mention of clocks. Is that max? I'm all over the place now
AFAIK the NVLink Switch is just to connect these multiple NVLinks, which in turn are used to connect multiples GPUs.
- Pronouns
- He/Him
Assuming the rest of the Blackwell platform is on 4NP, do we think they're just gonna keep with it for the 50-series? Ampere datacenter chips were a node ahead of the 30-series, and Hopper was the same node as the 40-series, so maybe they don't see 3 nm as worth it and are just waiting for TSMC to switch to GAAFET for N2 and use that for Blackwell-Next?
ItWasMeantToBe19
Manakete
LiC
Member
Thor having a Blackwell GPU isn't surprising. There was obviously some weirdness with the cancellation/reboot of Atlan where they basically just said that Thor was replacing it despite having the same specs with higher topline perf numbers, so they reiterated all those specs again including the GPU architecture, despite clearly not having actually started working on the chip yet. Sometime after that announcement in September 2022, they started working on it, and decided that it would have a Blackwell GPU instead. Nvidia often announces these things with early PR that doesn't correspond to what they actually end up making.
And we can actually follow the model numbers. T23x = Ampere, T24x = Grace CPU (not a Tegra but they use the model number anyway), T25x = Ada (cancelled), T26x = Blackwell. Specifically, T264 = Thor.
By the way, what Kopite said last September was "T254 is based on Blackwell," which is wrong since Thor is T264, and T25x were for the cancelled Ada-based Atlan. As usual, he just can't get his Tegras straight. In fairness, though, he was responding to somebody else who was also misusing the model number T254.
And we can actually follow the model numbers. T23x = Ampere, T24x = Grace CPU (not a Tegra but they use the model number anyway), T25x = Ada (cancelled), T26x = Blackwell. Specifically, T264 = Thor.
By the way, what Kopite said last September was "T254 is based on Blackwell," which is wrong since Thor is T264, and T25x were for the cancelled Ada-based Atlan. As usual, he just can't get his Tegras straight. In fairness, though, he was responding to somebody else who was also misusing the model number T254.
NineTailSage
Bob-omb
Nintendo also uses heavy compression on their audio files, which take up a lot of space, who knows if this will change for Switch 2 though.
That's what lost me a little, normally his speeches are more put together and some elements felt phoned in. When they brought out the two Blackwell boards, it felt like that segment went on way too long and like he was expecting a bigger payoff for people's enthusiasm...
Interestingly, Blackwell has a decompression engine. Developed at the same time as Drake's or were they developed separately?
I thought this was highly interesting as well, since I don't remember them ever having a fixed function hardware with sole purpose of hardware decompression. It makes me wonder if the new desktop GPU's will also get this feature (bringing PC hardware more in line with the current consoles on the market by that time).
GrandDemand
Cappy
- Pronouns
- He/him
Yes I think the 50 series will also use N4P. My guess is that N3 yields are atrocious for large dies and thus MCM is the way to go to pack in more transistors vs. a node shrink
some new, fun technologies
RTXGI has been updated with neural radiance cache
and the new spatial hash radiance cache, for gpus that can't run NRC
developer.nvidia.com
AMD shows off DX12 Work Graphs for gpu-driven rendering. now devs don't have to wait on the cpu when the gpu can generate work itself
really cool demo video
gpuopen.com
RTXGI has been updated with neural radiance cache
and the new spatial hash radiance cache, for gpus that can't run NRC
RTX Global Illumination (RTXGI)
Scale computation of global illumination with GPU ray tracing.
developer.nvidia.com
AMD shows off DX12 Work Graphs for gpu-driven rendering. now devs don't have to wait on the cpu when the gpu can generate work itself
really cool demo video
GDC 2024: Work graphs and draw calls - a match made in heaven!
Introducing "mesh nodes", which make draw calls an integral part of the work graph, providing a higher perf alternative to ExecuteIndirect dispatches.
gpuopen.com
NineTailSage
Bob-omb
I thought this solution was pretty interesting, when he shows off the large Blackwell GPU and mentions that it's interconnect is so seamless that the 2 GPUs think and operate like they are 1. So whatever issues they were having seem solved enough to now move forward without restrictions.
ItWasMeantToBe19
Manakete
I'm pretty sure someone in that demo should have an extra finger./s
Danny
Like Like
- Pronouns
- He/Him
Because AAA developers make their games for PC, PS and Xbox, most probably won't make day and date Nintendo releases. Just later ports. They have their priority platforms, and Switch 2 will be a bonus platform for them that comes later.
Ghostsonplanets
Ghost in the Sales
- Pronouns
- He/Him
It depends. Some flavors are straight up enhacements over the original nodes and the OG node is basically phased/morphed into these nodes. Other times, they're just specialized versions of the nodes for specific applications and don't replace the OG node but rather are offered as alternatives depeding on your design target.
This is just the Switch to connect all these GPUs hahaha. Nvidia introduced compute capability so that the Switch is now capable of detecting reliability of each part of the connected GPUs and indicate if something is wrong.Aren't those tflops weak? 3.6 fp8 means only 0.9 fp32.
Unless I'm misunderstanding something? I wonder why that number with no mention of clocks. Is that max? I'm all over the place now
ziggyrivers
thunder
- Pronouns
- He/him/his
Hey, everyone. This may sound a bit off topic, but I wanted to ask the hardware experts here:
How has technology advanced so much?
Like, almost 60 years ago computers occupied entire rooms in order to process smaller sizes of information. Nowadays, nano sized hardware pieces process exponentially more data than what we had when we first started.
To me, it almost feels alien in the sense that: how did we as humans achieve this, and so fast?
And this is my question: how has technology advanced so much?
How has technology advanced so much?
Like, almost 60 years ago computers occupied entire rooms in order to process smaller sizes of information. Nowadays, nano sized hardware pieces process exponentially more data than what we had when we first started.
To me, it almost feels alien in the sense that: how did we as humans achieve this, and so fast?
And this is my question: how has technology advanced so much?
Kise Ryota
Then I'll say "I told ya"
Z0m3le
Bob-omb
We know the target power draw of the GPU for 2TFLOPs, 3.456TFLOPs and 4.236TFLOPs, while the GPU was designed.
So we don't have to worry about power consumption or clock. The node can be called anything, we know the specs. That's my point here, worrying about fighting over the actual process node is just for people who want to prove others wrong and there is no way of doing that with anyone who believes these performance numbers are possible on 8nm. My secret point is that they lack understanding needed to argue in good faith.
oldpuck
Like Like
- Pronouns
- he/they
Aren't those tflops weak? 3.6 fp8 means only 0.9 fp32.
Unless I'm misunderstanding something? I wonder why that number with no mention of clocks. Is that max? I'm all over the place now
Nvidia SHARP is a technology that puts a small amount of AI capacity in network devices. It's pretty weak for a GPU but pretty crazy for a network device.
- Pronouns
- He/Him
Considering how much Rise sold, Capcom would be absolute fools to not make the Switch 2 a "priority platform". Hell, Rise was a Switch exclusive for 10 months, so the MH team clearly see the value in supporting – and even giving special treatment to – Nintendo platforms when the technology is there, as they have since Tri.
Last edited:
Z0m3le
Bob-omb
There is nothing conflicting that this test was for target clocks and estimated power draw. It's actually the only thing they could reasonably be assumed to be. As for the hidden text, I agree that someone shouldn't quote your content without the hidden tag, but I'd rather fight the spread of misinformation, so I'll keep my posts and content open.
Switch 2 is a 3 to 4 TFLOPs device and those clocks from the test are not in any conflict from being used on Switch 2 by us or MLID, the idea that we then have to prove it isn't 8nm, well that's obvious to anyone you could prove it to with these numbers, but I'll refrain from that argument.
Last edited:
Kumomeme
Rattata
- Pronouns
- Mr.
Gerald
Bob-omb
- Pronouns
- He/Him
So Kopite7Kimi was right, he was also right that Blackwell is still N4,N4P
Concernt
Optimism is non-negotiable
- Pronouns
- She/Her
Huh, oh, right!
Really does come back to either 4N or an 8nm so impossibly good it doesn't matter. Ultimately doesn't mean much.
Optimism, inarguable.
I will say though with those targets we can bring that back to formfactor and point out how say, a 7.91" screen device with the GPU targeting those numbers wouldn't really need to be any thicker than the current Nintendo Switch, which ticks a lot of boxes in terms of portability and accessory compatibility potential.
OP
OP
kopite7kimi also said that T254 was based on Blackwell, which is likely wrong since Thor is likely T264, not T254.
Gerald
Bob-omb
- Pronouns
- He/Him
They will use N3 but in 6000 series, if Blackwell gaming is Early 2025, then 6000 series can be Early 2028
Gerald
Bob-omb
- Pronouns
- He/Him
N3 and N2 dosent sucks at all, maybe if you mean prices then ok
LiC
Member
You can't just say they can "reasonably be assumed to be" something without justifying your interpretation.
There is no reason to use T239 target clocks on a different GPU on Windows. There's literally no value to doing that. You might as well use a random number generator for how meaningful those clocks would be in the test. Using clocks that are relative to each other similar to the (not-yet-existent) handheld and docked profiles means that you'll get execution times distinct from each other, which makes looking at your results more intuitive, and that's the reason why the clocks are different at all. Beyond that, they're arbitrary.
The tests were KPI profiling of DLSS execution under NVN2. That means they needed to keep everything the same between runs so that any change in the output metrics -- execution time, memory usage, and image tests -- could be tracked as they made changes to the code. They wanted to see improvements and guard against regressions. That code is what was being tested, not clocks, not power consumption, not T239. The resolution, quality mode, and P-State are relevant for exercising the code paths closely to what T239 would do, but the clocks are not relevant because they have zero effect on how the test executes, only how fast it executes. And how fast or slow the test runs is meaningless in absolute terms, only relative ones. But even if you hypothetically wanted the absolute time to match T239 for some reason, you would not be able to accomplish that by applying its clock speeds to a random RTX 20 series GPU.
I could go on. We could talk about how timeline-wise, T239 didn't even physically exist at the time of the hack, and the test cases could have been written months or years earlier on top of that. But I've already made enough of an argument. If you want to keep going around telling people that these are T239 clocks -- invoking my name, no less -- then you really need to offer an argument for why they would use T239 clocks in these tests in light of everything I've laid out here. My belief is that the best possible argument one could make is "they did it arbitrarily," which is pretty weak, but I guess better than having no argument at all.
oldpuck
Like Like
- Pronouns
- he/they
Moore's law, and Math.
The guy (I believe his name was Cain) who discovered Rock did more than just think of a cool way to kill people. Rock could be used to make Sharp Rock. Sharp Rock could be made into knives, knives could be attached to Stick (older technology) to make Spear, Spear could be made small to make Arrow. Sharp Rock could also be turned into Axe, Axe could fell tree, Tree plus Rock could become Trebuchet.
The early discovery of a new class of technology always leads to rapid innovation, as there are lots of obvious and untapped applications. But this era of innovation can be intensified if the new technology can be used to refine itself. Rock can be used to make better Rock.
The transistor was not just a new class of technology, but it was one that had an obvious technological improvement - make it smaller. And how would you do that? With a whole bunch of automation made possible by... the transistor.
This lead to 30 years of increasing power, simply by riding the node shrink. Which is not to be dismissive of the electrical engineers and physicists who worked hard to make this possible, but the fact that Gordon Moore was able to chart three decades of technological progress pretty accurately shows that it was an era of low hanging fruit.
Along the way came Math - Math is obviously thousands of years old, and the basis of software is Math. The neural network is 81 years old, but practical software applications didn't exist until we made powerful enough GPUs. The math was waiting there for the hardware to catch up. There have been huge advances in the field of software engineering, and of course, advances in Math - but again, Ada Lovelace wrote the first program well before the first computer. These things were ready to go the moment Electronic Math Machines were viable.
Z0m3le
Bob-omb
You found the test and shared it with the community, if you look at my posts where I use your name, I do say that my conclusion is in conflict with yours. Now these tests were done in summer 2021, when there was no card that could match the test's configuration.
The reality is that this test gives more than just random clocks, it estimates power consumption at those clocks. That is something that can't be explained away by your explanation. It's far more likely that they were just adding the DLSS function to the NVN2 API and that they chose target clocks because they tied them to estimated power consumption. If they were not sure of the clocks and power consumption they would be targeting with T239, they wouldn't have been designing the chip yet, but we know they were.
- Pronouns
- He/Him
And then once VLSI came along, suddenly, computers were used to design computers. And fucking Xerox were the ones to do it because IBM were a bunch of transphobic cunts.
Last edited:
- Pronouns
- He/Him
So the reason computers got so good so quickly is Minecraft
oldpuck
Like Like
- Pronouns
- he/they
dammiSo the reason computers got so good so quickly is Minecraft
oldpuck
Like Like
- Pronouns
- he/they
this was a typo, but I like it as a fami related swear
Last edited by a moderator:
