Switch 2 Speculation
- Thread starter Goodtwin
- Start date
Na, you cant compare Orin to a gaming SoC. Series S on TSMC's 7nm has 8 billion transistors with a die size of 197mm^2. Samsung's 8nm allows for ~50 million xtors per mm^2, so <=10 billion xtors at 200mm^2 is possible.
There is absolut no point in paying >2x for a wafer just to massively downclock the SoC to hit a certain power level.
There is absolut no point in paying >2x for a wafer just to massively downclock the SoC to hit a certain power level.
The video is saying the Switch 2 should have 15 billion transistors.
Subtlesnake
Regular
The issue is with Thraktor's analysis showing that 12 SMs would need to be clocked below the voltage floor on 8nm, so essentially Nintendo would be throwing away performance for nothing.
This is just speculation. The actual transistor count is unknown.
Charlietus
Regular
Are you challenging Nintendo? They will throw away performance for no reason just to spite us.
Subtlesnake
Regular
What I mean is that if the analysis is correct, Nintendo could have built a smaller chip with the same performance.
Charlietus
Regular
I'm disillusioned with Nintendo after 15 years, if it's 5nm Samsung I will be surprised, but if it's 8nm I will just think "that's Nintendo".
i expect it to have more modern features(than ps4) such as ray tracing, and if it can have some DLSS, it could end up having games looking better than what we've seen on PS4.
We should know soon enough. The main issue would be ram speed compared to PS4s, what's the average speed of LDDR5X ?
I'll just revisit my old post here.
Samsung foundry is such a weird choice no matter if it is 4nm, 5nm or even 8nm compared to TSMC 7nm.
But wasn't all NVIDIA GPUs in that generation made by Samsung as well?
Samsung's 8nm process was the last competitive with TSMC. Transistor efficiency was nearly the same, just density was only 1/3 of 7nm DUV. Going with Samsung and 8nm makes a lot of sense when you want a cheap SoC.
Switch runs at 160Mhz in handheld mode on 16nm FF. Must be the lowest clocked mobile SoC on this planet... Nintendo will just use the lowest voltage step while massively downclock the chip.
Last edited:
What? AFAIK, it runs at 307Mhz? Where are you getting 160 from?
Subtlesnake
Regular
Again, the point is that once you are at the lowest voltage, decreasing the clock further just loses you performance with no gain to efficiency. So in that case Nintendo would simply have designed a smaller chip. So either the voltage floor is different on the process they are using, or the power consumption differs from Thraktor's calculations.
Yeah. Switch has multiple GPU profiles ranging from 307Mhz to 460Mhz. The important thing being that those clocks were dictated by the Switch V1 (so X1 on 20nm). So stating that Nintendo is running A 307Mhz in handheld mode on a 16nm SoC is correct but misleading.
Anyway, it's fun to see the Nintendo craze back to kick off this new year. Sure, the confirmation that we'll get a Samsung made SoC is not a good news per say. But a lot of the analysis we get come from educated guesses and can be far from the actual thing.
About the size of the chip:
It actually boggles me how wide the die size area estimates are. I've seen values ranging from 160mm² to 218mm²... We have a high res picture of the thing, with solid known references, and still, those wild values are used to try to rule out option A or B.
My personal measurement gives me ~214mm², so, yeah, I'm with the 200-220mm² guys. And this only confirms that the chip is large.
About the transistor count:
I'm not sure how some can guestimate this. Sure, we have Orin. But Orin has a lot of logic that would be useless on T239. We also don't know how the chip will make use of cache. Small caches to keep the size and cost down? Super large caches to save on memory bandwith and boost energy efficiency? Large caches come with a lot of transistors that can explain the "large" chip on a "small" node.
Likewise, smaller caches and overall trimmed down logic (vs Orin) could explain the "too small for 8nm" take.
About the expected frequencies:
The only thing that I think is granted for now: the predicted "CPU@2.1Ghz/GPU@660Mhz" dream is likely not to happen, but not everyone expected this to be real so...
I think we'll be lucky to get CPU@1.4Ghz/GPU@440Mhz.
But I also think we may still get "high" clocks in docked mode. Like maybe a 3x increase instead of the Switch 2x.
I'm glad we got this leak, but AFAIK, we can't rule out any option yet.
Since it's a Samsung chip, I hope it will be 5LPP (5LPP is not that bad, and can be considered as quite good on the slower frenquencies range). And that's it.
The disappointment will be real if it turns out to be 8nm though.
In any case, actual reveal should be very very close now
Happy new Year Subtlesnake
Regular
Yes, and if you could get the same performance with a smaller chip at a higher clockspeed (and no loss in efficiency), why would you not build that chip instead?
Cost, assuming it's a smaller node. Heat, assuming it's the same node. This is all very reminiscent of discussions around Switch 1 Tegra X1. Many people were expecting a more customised chip or at a smaller node but it didn't happen. Tegra X1 was already old tech by the time Switch launched. Why should it be different for Switch 2. We also have rumours of Nvidia annoyed at Nintendo because T239 has been ready for ages and it was most likely on Samsung 8N, as it was designed around the same time as Ampere. Porting the chip to another node would have probably taken Nvidia more time so they wouldn't have been as annoyed, plus would also likely have a different code name by now if it was ported...
I fully believe the chip is 8N unfortunately, as I'm looking forward to upgrade my Switch..
Last edited:
Im sorry but why is Thraktors napkin math taken as fact? His last post basically said he has no idea and it could very well be 8nm
Famiboard went down in full meltdown mode because it wasnt TSMC 4nm but there was never any evidence for that. Simply amateur engineers doing napkin math while wishing for the best case scenario in terms of performance. Whether its 8nm or 5lpp or whatever node, keep in mind that its a 400 dollar console
Plus this SoC must performs at two very distinct power profiles.
Having the lower one below peak efficiency and the high one past peak efficiency could be an ok option (if the other option is handheld@peak/desktop@toohigh)
Going with a narrower design, requiring "high" frequencies in handheld mode could not be an option if the chip targeted desktop frequency becomes not possible at all.
We'll know soon enough how large the perfs gap between handheld and docked will be. But if it turns out it's 3x (or more), going with a narrow design could simply not be an option in the first place.
Subtlesnake
Regular
I am talking about on the same node. And surely thermal limits would come into play for docked, not mobile clock speeds. I guess it could limit the docked clock speeds, but then you are trading extra cost (bigger chip) for better docked performance, and no better portable performance, when the claim seems to be that Nintendo is primarily cost focussed.
I am not taking Thraktor's math as fact. I am using his reasoning as a jumping off point. If you disagree, you have to establish which of his assumptions are mistaken.
