Neat! Too late for Drake, but possible for a Drake Pro? Sounds like a Drake Pro could be thinner and still clocked higher without even a die shrink if it used this, though they would probably still do a shrink for battery life.
I don't think this is too late for [redacted], as they've been pitching this technology for a while and a product was just announced with it. A handheld gaming device is
one of their example use-cases, so I have no doubt they would have been pitching to Nintendo the first chance they got.
That said, I would be very surprised if they ended up using this. Cost being the primary reason, as I'm sure they're more expensive than a standard heatsink & fan assembly, but also cooling capacity and power consumption. The larger of the two units has a cooling capacity of 8.75W with a power consumption of 1.75W. That would just put it in range of cooling in docked mode if we were looking at Mariko power consumption, but would be insufficient if the new device is closer to the original Switch's power consumption. You'd need two (or three of the small ones) in that case, and although they're thin, they'd still take up a lot of space in the device.
The power consumption of 1.75W, or 20% of the cooling capacity, is also a bit concerning for a battery powered device. I don't know what the Switch fans consume, but I'd be surprised if they even hit 1W docked in the original Switch. Let's say we're talking Mariko-level power consumption and they use one of these in the new model. In portable mode, you're probably looking at around 1W power consumption for this, vs maybe 300-400mW for a traditional heatsink and fan assembly. That's several CPU cores worth of power consumption in the difference. Whatever space savings you make will probably need to go into a larger battery just to cover the additional power consumption from the fancy thin cooling unit.
Also, as someone with a launch day Switch, I've never had an issue with fan noise. It's barely noticeable the majority of the time, and if they do a similar job with [redacted] I wouldn't have any complaints.
This will be a bit verbose and long, so I'm sorry in advance.
So,
I was looking at Andrei testing of the Exynos 7420, an SoC manufactured on Samsung 14LPE that uses 4x A57(!), and stumbled upon this:
This is a graph that shows the Freq x Power of the Exynos A57. The most interesting thing about this is that Andrei actually showed us the power consumption for when 1,2,3 and 4 cores are loaded. We can see that 4 A57@1GHz (Same frequency of Switch TX1 A57 CCX) uses 1.16W. Disregarding some small amount of power which isn't being actually used by the A57 block (As Andrei points out in the article, he wasn't able to fully separate the A57 power consumption from others IP blocks or components like screen, so a small amount of energy is representative of those), we can round up these results to +- ~1W. Thanks to Apple A9 dual sourcing manufacturing from both Samsung 14LPP and TSMC 16FF, we know that Samsung node is denser while TSMC one is slightly more efficient(!).
Now take a look at this:
That's Mariko Switch Docked, per DF testing. They found that Mariko Switch draws from 5 - 7W while docked. And while I don't have portable results here, I can guess Switch portable should draw from 3 - 5W.
So,
@Thraktor (And others), considering Mariko TX1 has a budget of 1W( Per Andrei findings) for the A57 block, and the Maxwell GPU IP should draw frow 1 - 5W (Portable lowest < - > Docked Max and I'm also assuming 1 - 2W should be for the uncore), do you think Drake being manufactured on TSMC 4N would be able to hit the same power consumption figures TX1 Mariko does? And what power budget and possible clocks a 8x A78 on 4N you think it would require? I know you did a power curve of the 8nm Orin GPU a while back, but are you able to do something similar for the Orin A78?
There was also
a review of the Exynos 5433 on Samsung's 20nm process with a quad-core A57 block, which had the following power consumption:
Obviously these are Samsung processes rather than TSMC, and Samsung's implementation may be different from Nvidia's, but going from 1.83W for the CPU on Logan to 1.16W on Mariko sounds about right.
Personally my expectations are a battery life somewhere in the middle of Logan and Mariko units, with power consumption a bit closer to the OG model, but improved battery density balancing that out. I'd personally be surprised if they managed to get power consumption down to Mariko levels even on 4N, given the size of the GPU, but can't really rule anything out. There's also always the possibility that the form-factor is changed in a way which means a larger (or smaller) battery, which would obviously impact the power consumption they can get away with.
For the CPU, the numbers I got from the Jetson Orin power estimator are as follows. These are for 8 cores, although they're A78AE rather than A78C, and split across multiple clusters, so obviously caveats apply.
Code:
Clock(MHz) Power(W)
1113.6 2.2
1267.2 2.5
1497.6 3.1
1651.2 3.8
1728.0 4.1
1881.6 4.9
2035.2 5.8
2188.8 7.1
If TSMC 4N consumes half as much power as Samsung 8N (rough estimate based on Ampere to Ada comparison), then with a power consumption similar to Logan's CPU you'd be looking at around 1.6GHz and with power consumption around Mariko levels it would be about 1.2GHz. That said, with the difference between A78AE and A78C, moving to a single cluster, and very rough estimate on TSMC 4N improvements, there's a lot of wiggle room there.
