Nintendo Switch Tech Speculation discussion

[Deleted member 13524]

Each stack of HBM2 is 256 GB/s. "Half a stack" (e.g. going from 4-Hi to 2-Hi) will only reduce memory amount, not bandwidth.

Not that there's even a chance of seeing HBM2 in a handheld. HBM2 is very expensive and its power envelope is not for handhelds.
OTOH, 128bit LPDDR4 would be a plausible fit.



BTW, this thing is now selling for £99 (£130 out of black friday deals):
https://www.amazon.co.uk/dp/B0187U2M1O/?tag=b3d-21

One would think Nintendo could do a lot more with a smaller/cheaper screen, twice the BoM money and much bigger production scale.

 

[Goodtwin]

Yes, Nintendo usually makes a profit from their console hardware. However this time their entire business depends on a successful launch. Selling hardware at zero profit on launch is entirely possible. Zero profit at launch doesn't mean that the lifetime profits of the hardware will be zero. Consoles have long life time and parts get cheaper every year. They need to get the new ecosystem rolling and that requires lots of customers.

Only Nintendo knows their business best. Traditionally their average customer buys less games per sold console as Playstation and Xbox gamers. IIRC an average PS3 gamer bought 8 games, while an average Wii gamer bought only 2 games. But Wii was a one trick pony, unlike Nintendo handhelds. Nintendo handheld market is completely different (long product life time, lots of good games). Considerations like this make a big difference when you calculate an acceptable profit margin for the hardware.

Where are you getting this info? If you go to Nintendo's Investors Relations page, you will see the Wii hardware sales are listed as 101.63 million, and software sales are at 916.15 million. 9 to 1 is pretty similar to other consoles. Even the Wii U is at 13.36 million hardware sales with 92.35 million in software sales. A little lower ratio, but pretty decent considering it had so little support.

 

[MrFox]

Half a stack then

Each stack of HBM2 is 256 GB/s. "Half a stack" (e.g. going from 4-Hi to 2-Hi) will only reduce memory amount, not bandwidth.

Not that there's even a chance of seeing HBM2 in a handheld. HBM2 is very expensive and its power envelope is not for handhelds.
OTOH, 128bit LPDDR4 would be a plausible fit.

Yeah, it's power hungry, and obviously too expensive so it's not happening, but it's fun to speculate what if.

HBM2 is flexible with the number of channels per die, so they are free to implement only 4 channels total instead of 8. 256GB/s is all 8 channels enabled, and each die can be designed with 1,2 or 4 channels.

Normally it means any of these give 256GB/s:
2Hi, 4 channel per die
4Hi, 2 channel per die
8Hi, 1 channel per die

But 2ch dies are what is currently being produced for the 4Hi devices GPU manufacturers are ordering. It's the only thing in mass-production (publicly, at least). So if they simply halve the number of dies of this existing production, they end up with 128GB/s. And it's stupid because they also end up with 2GB total.

HBM1 doesn't exist anymore, and it doesn't have the capacity per die for Nintendo's needs anyway.

HBM "low cost" is an interesting idea but Samsung is barely in the idea phase, it doesn't even have final specs, let a lone a mass production.

Where is WideIO2 when we need it? This would have been the best application for it, much lower power consumption than lpddr4, and a nice 68GB/s. It never appeared even on the most high-end mobile devices, so I'm pretty sure it's Dead.

Nvidia Parker have a 128bit lpddr4 so that's a much better possibility. But twice the interface is twice the power. It might be too much for a 6.2" portable.

 

[mpg1]

Not sure if this was posted before but Digital Foundry did some tests of the Dolphin emulator on Nvidia Shield TV:

Surprisingly decent results....at least when it comes to potential BC.

 

[Rikimaru]

Not sure if this was posted before but Digital Foundry did some tests of the Dolphin emulator on Nvidia Shield TV:

Surprisingly decent results....at least when it comes to potential BC.

There is nothing to make BC. Everything will be ported
But seriously, no. Nintendo would not make an emulator. They even made special hardware mode for Wii on Wii U.

 

[Goodtwin]

I think any and all modifications that Nvidia would make come down to reducing cost, reducing energy consumption, and lastly increasing performance. Removing the four A53 CPU cores seems likely, because from what I have read they aren't really in use when gaming on the Tegra X1. This would free up dye space, and perhaps Nintendo will opt for 32MB of esram. They have a history of using small pools of very fast memory for their GPU's going back to the Gamecube. I would assume this would be less power hungry than a 128 bit bus to the main memory. I could also see them cutting down on the number of rops if they have no intention of offering 4K video output, and I honestly don't expect 4k video output to be a thing for a product like Switch. The Tegra X1 has 16 rops, pretty rop heavy unless your doing 4k video output. Im not sure how much dye space 32MB of esram would take up, but perhaps removing the A53 cores and 4-8 rops would free up enough space.

 

[dogen]

Not sure if this was posted before but Digital Foundry did some tests of the Dolphin emulator on Nvidia Shield TV:

Surprisingly decent results....at least when it comes to potential BC.

Pretty old version of the emulator too. Last I checked there were a few people actively working on optimizing the ARM backend.

 

[milk]

Nintendo would not make an emulator. They even made special hardware mode for Wii on Wii U.

Of course nintendo would never MAKE an emulator when they can just use an open source one avaliable on the web, just like they did for virtual console up to n64.

 

[618116]

I'm still thinking they might implement some ESRAM thing in SoC, because of their comment at 2014 in shareholders meeting...

On the other hand, it was a natural and inherent decision to aim for hardware efficiency, including a not huge capacity but low-latency memory, in designing Wii U, as we have done since the days of Nintendo GameCube. These sorts of things have been ubiquitous across the entire company; it is in our DNA.
https://www.nintendo.co.jp/ir/en/stock/meeting/140627qa/02.html
​

At least ESRAM is more realistic than HBM, isn't it?

 

[Rootax]

Does nVidia have experiences with ESRAM (or similar things) ? It would be nice, and help the bandwidth a lot, if it's not too small.

 

[mpg1]

rumors on hardware from "someone working at Foxconn": https://www.reddit.com/r/NintendoSw...r_someone_who_producing_switch_at_foxconn_is/


Dock:
- no advanced technology in dock
- seems cheap and light, feels really plastic
- no extra power, its just a output
- 1x USB3.0, 1x hdmi, on the side 2x USB2.0

Console:
- pretty sure the console screen is 1080p and multi touch
- ...but heat performance is really good, the software demo testing millions of fish running almost 8 days, theres no single frame drops
- screen is not very bright
- looking at the core it's a 10x10 core, from the test, CPU 1785mhz, GPU 921mhz, EMC 1600mhz
- specualted CPU is arm A73 pascal, much powerful than X1, when tested only shows ARM_V8 structure hence the spcualtion
- confirmed it's USB-C charging
- speculated core is made by TSMC
- saw orange and blue controller
- heating fan noise not loud
- power adapter is external
- theres 4G console version
- can be charged while playing
- battery 4310mA, not changeable
- speculated 2X ram = 4GB
- about 300g weight excluding joycon

Joycon
- 2 shoulder button on each joycon, called SL, SR
- very complex inside, apart from motherboard in console, screen
- Joycon is the most valuable in whole system
- Joycon is about 5cmx2cmx0.5cm, very light about 50g, battery 525ma

 

[mpg1]

it's just rumors but definitely some interesting things in there...

for instance 1080p display as opposed to 720p...dock just acting as a video out.....perhaps much stronger cpu? although not sure what is referred to by 10x10 core...also 4G version!?...

edit: speculation in reddit thread that 10x10 core could refer to die size ie 100mm2

 

[Shifty Geezer]

although not sure what is referred to by 10x10 core

mm is my guess. Oh, I see you notice that in edit.

 

[Goodtwin]

Does this info have any sort of credibility? I know mods over there typically try to vet rumored information like this. Assuming it is legit info, just how much info is lost in translation.

 

[fehu]

So are you really saying

No Denver?

 

[AlNom]

Isn't Denver rather power hungry?

 

[Benetanegia]

Isn't Denver rather power hungry?

From what I understand as a layman, I don't think that's the case form the CPU perspective. Denver relies much more on external memory transactions to access the micro op buffer, so its power probably depends heavily on the efficiency of the memory used. Denver K1 used LPDDR3 which would compare very unfavorably to LPDDR4 at the time.

 

[mpg1]

So right now we can speculate from newest rumors is :

4-core ARM A73 CPU
256 CUDA Pascal cores GPU
4GB LPDDR4 memory
6" 1080p display

I'LL TAKE IT!!

 

[cheapchips]

Screen not being bright would be a bit rubbish if true. I'd only be able to take it to nighttime roof top parties.

 

[Deleted member 13524]

100mm^2 die size isn't bad but it isn't really great either.
The Tegra X1 has 120mm^2 and its 20nm node isn't that far off in transistor density from the FinFet nodes. The apple A10 has 125mm^2.

100mm^2 in FinFet is probably good for around 2.5 billion transistors. That would be about half the transistors available to the Xbone and PS4.
If they go with LPDDR4 alone then there might be a decent CPU/GPU combo (EMC 1600MHz suggests LPDDR4 3200MT/s), but if they put eDRAM in there then both the CPU and GPU will have to be tiny.

I guess an area advantage the Switch's SoC has over e.g. apple A10 and TX1 is that they don't need a large high-performance ISP and video encoder. More than mediocre 4MPix photos and 720p videos are probably an overkill for a portable console to do augmented reality stuff, taking a photo for an online profile and eventually video chat.

As for the A73, the chances for that core to be in there are close to zero. They were announced in May this year so I doubt nvidia - who is still using Cortex A57 in Parker/TX2 - would get the new cores into a custom SoC up and running in time for production in November.
Plus, they're supposed to be releasing only on 10nm, so unless there's a super surprise and Nintendo got the very first 10nm chips in the market, this is just not going to happen.

EDIT: Just remembered that the Kirin 960 is shipping in the Huawei Mate 9 which has a quad Cortex A73 module using 16FFC, so it's possible the Switch's SoC to be using the Cortex A73.


At those rather high clock speeds, I'm hoping for 2* 4-core Cortex A53. They wouldn't be all that much slower than the 2* 4-core Jaguars at 1.6GHz if we look at Geekbench scores so it would be a much more balanced configuration than using big cores. And they would save loads of die-area for more GPU:

Oh but if this was using 10nm and Cortex A73 cores then 100mm^2 would rock.
It isn't, though.