would be pretty interesting if nintendo then went "fuck it, we'll just work with AMD" and emulated switch 1 by using open source emulator like what sony did with their ps1 mini.
Switch 2 Speculation
- Thread starter Goodtwin
- Start date
Because of the volume at which Nintendo buys components they can source materials at a much better cost than low volume products like the Steam Deck or even the Nvidia modules. Nvidia was probably selling Tegra X1 modules back in 2017 for the same cost that Nintendo rolled out the Switch. I have also speculated the Switch OLED was an opportunity to not only offer a higher tier Switch to the market, but also to get a head start on sourcing OLED screens. Nintendo will have sold 10-20 million OLED Switch units by the time the next Switch is launched. Compared that to a product like Steam Deck where they will be lucky to move a million units a year. Nintendo doesn't put in orders for tens of thousands of components at a time, they are looking for ten times that amount.
No chance. The partnership has been fantastic for Nintendo. The Switch has had a great software development environment compared to previous Nintendo consoles and thats largely because of Nvidia. Nvidias help on the software side of things outweighs any shortcomings that their hardware offering may have. The A78 CPU cores support the same instruction architecture that the A78 cores do. So even if the newer GPU cores break compatibility, I am confident that Nvidia can create a software layer to emulate the Maxwell architecture. Nintendo themselves did something similar with Mario Galaxy on Switch. The CPU logic runs natively on Switch and they then emulate the GPU side of things.
I no longer have the files, but I have verified them, here is snippets of the code shared in another forum about it from March 1st.
NVN1 Source:
C:
// Number of warps per SM on TX1 hardware
#define __NVN_NUM_WARPS_PER_SM_TX1 128
// Number of SMs on TX1 hardware
#define __NVN_NUM_SMS_TX1 2
NVN2 Source:
C:
// Number of warps per SM on ga10f
#define __NVN_NUM_WARPS_PER_SM_GA10F 48
// Number of SMs on on ga10f
#define __NVN_NUM_SMS_GA10F 12
(note all Ampere GPUs have 48 warps per SM, and Ada GPUs have 64)
In another part of the code, we also know that the GPU has 128bits... Also GA10F is also known as T239, which Kopite7 (the Nvidia leaker) in Spring of 2021 had leaked T239 as a custom Tegra chip made for Nintendo, here it is in NVN, which is Nvidia's API designed specifically for Switch, confirming that GA10F/T239 codenamed "Drake" is a custom Tegra chip for Nintendo.
It should also be noted that GA10F SMs have the same performance as desktop Ampere, further confirming 128 cuda cores per SM, it also calls for 48 tensor cores and 12 RT cores.
The leaked specs put it between the hardware and die size of Orin AGX (T234) at the high end and Orin NX at the low-end. The leaks indicate 12 SM of Ampere, so 12 * 128 = 1,536 cores, which is more than Orin NX. On the other hand, the 128-bit bus is half the 256-bit bus of AGX/T234. There is no version of Orin that has 1,536 cores and a 128-bit bus. Could it be that Nintendo intends to use heavily salvaged AGX Orin (T234) dies to create a spec between Nvidia's Jetson Orin NX 16GB and Jetson AGX Orin 32GB?
Whether or not T239 is truly a "custom" piece of silicon or a salvaged Orin, it seems like a lot of silicon to stuff into a handheld device running at ~15W (including screen). Maybe the Samsung 8nm+ process is so cheap and so optimized for mobile at this point hat Nintendo/Nvidia can afford to go "wide and slow" at the expense of a larger die size? That must be somewhat true, as the AGX Orin 32GB (8 core CPU, 1792 Cuda Cores, 256-bit LPDDR5) can be scaled down to 15W according to Nvidia. Cut the bus and memory pool in half to 128-bit, 16GB, remove 2 SMs, and downclock things a bit, and you might save enough to power the screen, controllers, etc. within 15W.
I'd really like to see the Switch Pro scale clocks higher in docked mode, which it probably would need to do to hit 4k DLSS, if that's a realistic target.
Whether or not T239 is truly a "custom" piece of silicon or a salvaged Orin, it seems like a lot of silicon to stuff into a handheld device running at ~15W (including screen). Maybe the Samsung 8nm+ process is so cheap and so optimized for mobile at this point hat Nintendo/Nvidia can afford to go "wide and slow" at the expense of a larger die size? That must be somewhat true, as the AGX Orin 32GB (8 core CPU, 1792 Cuda Cores, 256-bit LPDDR5) can be scaled down to 15W according to Nvidia. Cut the bus and memory pool in half to 128-bit, 16GB, remove 2 SMs, and downclock things a bit, and you might save enough to power the screen, controllers, etc. within 15W.
I'd really like to see the Switch Pro scale clocks higher in docked mode, which it probably would need to do to hit 4k DLSS, if that's a realistic target.
I got bad news for you. Samsung is just bad with regards to power efficiency. All measurements from Geekerwan on YouTube.
Nintendo style visuals at an uncompromised 60 FPS @ 1080p is a good goal for portable performance but it won't take 4 TFLOPS. Steam Deck like performance is more practical and I'd be surprised if the Switch 2 was well beyond 2 TFLOPS docked.
Last edited:
Rootax
Veteran
Dlss will help for sure but they need a beefy cpu if they want current gen port for the next few years I guess. The ones in the tegra X1 are not that bad compared to jaguar, but vs Zen2, eh... Plus pretty fast storage and/or ram amount. I hope they got 8gb at least. I fear only 6.
1536 Cuda cores scales fast. Steam Deck has 512 cores? The problem with going super low on the clocks is that you run into a question of why the GPU is so big, why make a GPU bigger than you find in Orin NX, if the goal is just to clock it below 2tflops when docked? You'd have to have a clock around 600MHz docked to be under 2TFLOPs, portable clocks are even more sus, I think at minimum we should expect Switch's clocks in both modes, so 460MHz for 1.41tflops in portable and 768MHz 2.36tflops in docked.
I also don't know if Nvidia stuck with Samsung for this device, they seemed to have moved back to TSMC for all future chips.
Well, the efficiency curve will be much better while still offering superior performance and longer play time (this thing still has a battery).
Wide and slow is usually better on a handheld device in exchange for increased die size.
We should hope that they have ditched Samsung as the foundry as TSMC gives much better performance / watt.
Yes, when docked the limitation of the battery is removed, so hopefully we will see a bigger clock than expected at that point. DLSS will also play a huge role in catching up to current gen's graphics.
Portably I don't know if ~300MHz for instance would make sense, I think given the efficiency of even 8nm vs 20nm, 400MHz would be reached. DLSS would also help here, if the screen is 1080p, you'd be able to render at 540p and use DLSS to hit 1080p on a 7inch screen, that should look really good and offer a lot more performance than a PS4.
Have in mind that the new chip is probably 3 to 4 full fab nodes ahead of the current 16nm shrink of the Tegra used in the Nintendo Switch 2019 model. With overclocking you can easily get the 16nm CPU up to 2.3GHz, GPU up to 1.3GHz and memory up to 2GHz at the expense of fan noise and battery life
I'm inclined to believe Nintendo will keep the 720p screen, though hopefully with an OLED as standard. 720p looks pretty good and it's probably cheaper. The OG Switch does a pretty good job hitting 720p without DLSS or the like, so the Pro shouldn't break a sweat at native. This will allow longer battery life and lower heat generation. Also, keeping the Pro at 720p may ease the burden on developers by limiting the number of different configurations they need to account for. The downside being that developers may continue programming for the lowest common handheld denominator (OG Switch). On the other hand, the Pro should ensure a steady frame rate and reduce or eliminate reliance on dynamic resolution scaling.
Yeah 720p is perfectly fine, DLSS would still offer lower power consumption with 480p or 540p rendering, and because of the screen size, it would be even harder to notice any artifacts from the technology. This is something 3rd parties pushing a current gen game to the limit on other systems, would be able to more easily fit on this Drake powered SoC.
I also think they could turn to the cloud to support the OG model, if Nintendo partnered with Nvidia for a cloud service with their own titles and maybe some select publishers, Nintendo could sell Drake games and just have them play automatically in the cloud on TX1 models, that way they can continue to sell to the current ~115M Switch owners, who have bought their software in record numbers. However I would expect something like that to happen after a soft cross gen phase of at least 2 years.
That's an overstatement based on the latest rumors. If we assume Drake/T239 is on the same process node as T234, which seems likely given the model number and Kopite's suggesting that T239 is a "customized" version of T234, then we are looking at Samsung 8nm, down from the TSMC 16nm. That's one full node jump, maybe 1.5 if you give extra points for Samsung's 8nm node being particularly mature at this point.
To get "3 to 4 full fab nodes" beyond TSMC's 16nm, you'd need to make a chip on TSMC's 5 or 3nm or Samsung's 3nm. If that's the case, we aren't going to see a new Nintendo chip for years because Nintendo isn't likely to compete and pay for cutting edge, capacity limited fab processes.
One can dream but you are right.
Orin NX is, unfortunately, currently fabbed on Samsung 8nm.
Actually Switch was built on 20nm which was a 2d node, while TX1 is 16nm 3D node, it's clocks are limited to 20nm, which was a dead end node, the performance of 28nm is the same, thanks to 20nm's leaking.
Also the model number has no barring on other model numbers, for instance TX1 was T210 but TX2 was T186.
There was a leak in spring last year about a power consumption issue with the chip, which was rumored to be "Dane" at the time, so Nvidia/Nintendo might have moved it down to 5nm Samsung (with all the companies leaving that node like Qualcomm, and not winning Nvidia Ada GPUs, there is plenty of capacity there, and Nvidia actually reportedly has a ton of 4N capacity that they don't want to use now) it's also worth noting that Nvidia uses AI to quickly shrink nodes and move to different fabs now, Samsung has also moved from DUV 8nm fab process to EUV 7nm process within 6 months last year on a SoC. Basically everything is on the table with Drake's process node.
Kise Ryota
Newcomer
That's exactly what I'm expecting. 2024 here we go lol
I don't bother at all having to wait for a batter hardware (in this case, better node for a price Nintendo is more suitable to pay).
Switch is still selling a lot, and Nintendo can't even keep up with its demand. I don't see why take the risk with a new hardware in 2023. Switch can sell +18~20kk next FY (2024), which is pretty good. World's economy is in trouble, shortage is still bad, and their current product still has steam.
Now is the time to release big games and profit with those 15kk+ sales per title. Let the new hardware to a better time, with better conditions in general (which could be 2024, I believe).
But that's just my view.
Last edited:
