Nintendo Switch Tech Speculation discussion

[Goodtwin]

Assuming Ubisoft is continuing the AC Chronicles series, then I would bet that it makes its way to Switch for sure.

So Dave over at Gaf has been continuing to run test, and it seems that the CPU can throttle the GPU, and vice versa. Basically, they cant both sustain max heavy loads without throttling. It makes sense for this to not be immediately clear when running benchmarks, seeing as how benchmarks are typically only pushing the CPU or GPU, not both at the same time. Shield TV acts more like a tablet/phone than a console. It throttles whenever it can. Once he was able to completely lock t the CPU clocks to max speed, the GPU would then throttle between 614 Mhz and 768Mhz. I would still like to see what the clocks do while gaming, but I think its likely that the clock speeds are not at maximum when playing games. Who knows, perhaps if a game doesn't push the CPU, then the GPU can stay clocked at 1Ghz, but if t the CPU is being pushed, the GPU lowers clocks. The fact that these test keep showing a 768 Mhz throttled speed under load seems to suggest that its the sweet spot, going higher isn't sustainable with such meager cooling.

Simply clocking high isn't a big deal, but clocking high under load is another story. The custom API is only going to allow utilization to go up. So at a given frequency, the Switch is going to actually be doing more work than the Shield TV can. Android is a thick API that doesn't allow developers access to the metal. I think in a few days its going to be pretty obvious that the Tegra chip powering switch is the best Nvidia's got within the design parameters, and not a cut down TX1 like some thought after hear clock speeds.

Dave also ran some test for bandwidth, and his test showed good performance at 1080p 8X MSAA. The scene looked pretty simplistic, but results didn't point to an obvious bottleneck with memory bandwidth.

Just a few more days......

 

[pc999]

I still think there is a good chance of seeing some costum work, even for the New 3DS they doubled the cores and added VRAM to the SOC without any fuss about it or hundreds of years of work. So added memory on SOC, a few extra low powered CPU cores (A 53 ? or even A57) or DSPs to handle OS or audio should be quite easy, no?

 

[BRiT]

Nintendo doesn't know the definition of "quite easy".

 

[BRiT]

Here's some more valid thoughts from GAF to temper expectations of Nintendo Switch based on Shield TV thermal throttling.

The Nintendo Switch is around 14 mm thick.
The Nvidia Shield TV is around 26 mm thick.

The Nintendo Switch has to recharge the battery while docked.
The Nvidia Shield TV does not have to recharge any batteries.

Everyone should level set expectations accordingly.

 

[orangpelupa]

Here's some more valid thoughts from GAF to temper expectations of Nintendo Switch based on Shield TV thermal throttling.

The Nintendo Switch is around 14 mm thick.
The Nvidia Shield TV is around 26 mm thick.

The Nintendo Switch has to recharge the battery while docked.
The Nvidia Shield TV does not have to recharge any batteries.

Everyone should level set expectations accordingly.

Maybe the will throttle the charging when docked while playing game? So less heat

 

[cheapchips]

Given we now know that Shield TV and Switch's clocks are pretty comparable in practice, surely the question is how noisy it'll get when docked?

And if charging too, can you hold your head over it to dry your hair.

 

[tongue_of_colicab]

Its not like charging the battery is going to add tens of degrees of extra temperature. Plus they could just trickle charge or not charge at all while playing games when docked.

Nintendo doesn't know the definition of "quite easy".

Japan Inc. doesn't know the definition of easy. Everything that is easy, straight forward or makes common sense needs to be tentacle raped into something that is the exact opposite. The longer something takes, the harder and more confusing it is the better.

Its the process that matters (99% of the time that is ass backwards in Japan), not the end result.

 

[Lalaland]

Its not like charging the battery is going to add tens of degrees of extra temperature. Plus they could just trickle charge or not charge at all while playing games when docked.

The battery raises the floor for ambient temperature versus the Shield because it itself will dump heat into the chassis as it charges or discharges complicating cooling versus an external DC source (from a wall plug). Additionally the CPU can damage the battery or lead to it having a shorter service life by adding thermal stress to the battery, the spate of Galaxy Note failures is largely theorised to be down to the battery have insufficient expansion room to deal with the stresses from charging/dishcarging and the thermal stress from the phones SoC.

Can you imagine how pissed off folks would be if they docked their Switch for 3 hours and had <50% charge? I think any sort of slow charging for portable devices is simply off the table in today's market

 

[Jubei]

I dont think there will be a surprise in performance. But i would be shocked if this is not on 16nm. I dont really subscribe to the idea that a company like Nvidia has a bunch of 20nm Tegra X1s laying around to use in a 5+ million selling console. Looking at their financials, their Tegra revenue has steadily been going up, not down

 

[Deleted member 13524]

Man years are the best kind of metric, who gets counted and how? Who cares? Feel those digits!

Every single competent project manager in the world that is working in a developed country.
Man-years/months/hours is one of the most important metrics when applying a project for investment funds, in parallel with milestone dates and budget-per-milestone.

I think in a few days its going to be pretty obvious that the Tegra chip powering switch is the best Nvidia's got within the design parameters, and not a cut down TX1 like some thought after hear clock speeds.

It's not even the best nvidia has because Parker would most definitely achieve better performance on the same power budget.
If anything, it's the best nintendo was willing to pay for, and/or the best nvidia was willing to part with.

I still think there is a good chance of seeing some costum work, even for the New 3DS they doubled the cores and added VRAM to the SOC without any fuss about it or hundreds of years of work. So added memory on SOC, a few extra low powered CPU cores (A 53 ? or even A57) or DSPs to handle OS or audio should be quite easy, no?

Looking at the tens of companies making ARM64 SoCs with decent GPUs right now, I'd say it shouldn't be hard at all for a company like nvidia to do that.

I dont think there will be a surprise in performance. But i would be shocked if this is not on 16nm. I dont really subscribe to the idea that a company like Nvidia has a bunch of 20nm Tegra X1s laying around to use in a 5+ million selling console. Looking at their financials, their Tegra revenue has steadily been going up, not down

Sales of the TX1 for the Switch might have been hidden throughout the last year as several small-ish shipments to Nintendo.
Truth be told, the Pixel C is a commercial failure and the Shield TV isn't exactly a kind of device that moves high volumes.
And while the TK1 was shipped to a bunch of devices (Nexus 9, Shield Tablet, Acer Chromebook, Xiaomi MiPad, a bunch of chinese consoles and tablets), the TX1 wasn't nearly as popular.
The Shield Tablet 2 being cancelled in early 2016 makes it even more suspicious.

 

[Jubei]

Sorry i should have been more clear. By Tegra revenue, i was talking about their automotive efforts (i dont believe they use any other chips for that? could be wrong)

Your point about Shield being low volume is why i think its unlikely to Switch chip is 20nm. Only companies like Apple and Qualcomm can afford to skip "just in time" production because they have a guaranteed constant demand for their products. It would be crazy for Nvidia to order millions of TX1 chips because of wafer costs and hope to sell them off later

Pixel was a failure but it was also a low volume product. It was sold in only a few select markets and mostly though Google Playstore. The original Shield tablet was also a low volume product, when it failed to sell at 299$, Nvidia repackaged it and did a firesale essentially for 199$. That experience suggests they would not be foolish enough to think a version 2 would somehow be a big seller

http://i.picpar.com/v8Wb.png

 

[Lalaland]

Every single competent project manager in the world that is working in a developed country.
Man-years/months/hours is one of the most important metrics when applying a project for investment funds, in parallel with milestone dates and budget-per-milestone.

Of course there are legitimate contexts but in this context it was a remark from JHH about the level of work gone into the design, i.e. marketing BS.Frankly even from a planning context I have an issue with this metric as it encourages a dangerous mindset that implies if 5 people do a 500 man hour project it would take 100 hours so let's double the team and have it take 50 hours!

It's not even the best nvidia has because Parker would most definitely achieve better performance on the same power budget.
If anything, it's the best nintendo was willing to pay for, and/or the best nvidia was willing to part with.

Parker hasn't shipped in anything beyond a few self driving car prototypes, we have no idea of it's power consumption and thus far it's been pitched for automotive solutions only. None of that suggests a platform one should launch a multi-million unit mobile platform on in Q1 2017.

Looking at the tens of companies making ARM64 SoCs with decent GPUs right now, I'd say it shouldn't be hard at all for a company like nvidia to do that.

I think you're grossly underestimating the difficulties here, TX1 is very close to an ARM ref design on the CPU side yet even at that they managed to mess up the Big.Little design by messing up cache coherency and produced one of the least energy efficient designs in the ARM market. CPU design is very, very hard.

 

[sebbbi]

I think you're grossly underestimating the difficulties here, TX1 is very close to an ARM ref design on the CPU side yet even at that they managed to mess up the Big.Little design by messing up cache coherency and produced one of the least energy efficient designs in the ARM market. CPU design is very, very hard.

Samsung's 5410 wasn't perfect either... 5420 fixed it.

 

[pc999]

The battery raises the floor for ambient temperature versus the Shield because it itself will dump heat into the chassis as it charges or discharges complicating cooling versus an external DC source (from a wall plug). Additionally the CPU can damage the battery or lead to it having a shorter service life by adding thermal stress to the battery, the spate of Galaxy Note failures is largely theorised to be down to the battery have insufficient expansion room to deal with the stresses from charging/dishcarging and the thermal stress from the phones SoC.

Can you imagine how pissed off folks would be if they docked their Switch for 3 hours and had <50% charge? I think any sort of slow charging for portable devices is simply off the table in today's market

Those are the real reasons to think that they would go with 16nm even if they hadn't tinkered that much with the design, that they somehow got a SoC that is working on the limit in a much coller ambient and get it in a smaller space with a lot more electronics and a screen and a charging battery and it wold work flawlessy is expecting too much IMO.

Besides Wii U, did they ever had gone with a older process than the latest, IIRC both Wii and 3DS had gone with the more recent ones?

I think you're grossly underestimating the difficulties here, TX1 is very close to an ARM ref design on the CPU side yet even at that they managed to mess up the Big.Little design by messing up cache coherency and produced one of the least energy efficient designs in the ARM market. CPU design is very, very hard.

But nvidia as a whole have been doing lots of work in CPU design, they even plan on going full Denver in Xavier by the end of the year, they wouldn't develop CPU just for Nintendo, but I belive they would allow the team working on Switch to access all the technology they could.

 

[Lalaland]

Samsung's 5410 wasn't perfect either... 5420 fixed it.

From what I can read they added Global Thread Scheduling, but there's nowhere that describes what that is beyond the conceptual level. I presume this is a combo of an improved memory controller and cache coherency but if you have any docs you can point me to I'd appreciate it (my Google-Fu has failed me).

 

[sebbbi]

From what I can read they added Global Thread Scheduling, but there's nowhere that describes what that is beyond the conceptual level. I presume this is a combo of an improved memory controller and cache coherency but if you have any docs you can point me to I'd appreciate it (my Google-Fu has failed me).

Some info (about "global task scheduling"): https://en.wikipedia.org/wiki/ARM_big.LITTLE

It basically allows using all 8 cores (big and little) at the same time in a flexible way. 5410 only had 4 cores active at time (switch all at once <- STALL).

 

[Lalaland]

I looked through the WIkipedia stuff but it's still quite high level, I guess I was hoping it would provide an idea of how much work went into adding GTS as making the 5420 from the 5410 as it's the closest analogue we have for a TX1 orig to a putative TX1 w/GTS we have. There are dieshots of the 5410(http://www.techinsights.com/teardown.com/samsung-galaxy-s4/) and 5420 (http://www.anandtech.com/show/6768/samsung-details-exynos-5-octa-architecture-power-at-isscc-13) but with them having transitioned to an entirely different GPU set up it's hard to do a straight comparison.

Of course without actual dieshots of the switch it's all speculation and I really doubt NV or Nintendo will spend too much time on tech like this, ah well roll on March and the inevitable teardowns!

Edit: Wanted to rephrase to make clear I wasn't demanding that Sebbi, or any other poster, answer my questions

 

[function]

Assuming Ubisoft is continuing the AC Chronicles series, then I would bet that it makes its way to Switch for sure.

So Dave over at Gaf has been continuing to run test, and it seems that the CPU can throttle the GPU, and vice versa. Basically, they cant both sustain max heavy loads without throttling. It makes sense for this to not be immediately clear when running benchmarks, seeing as how benchmarks are typically only pushing the CPU or GPU, not both at the same time. Shield TV acts more like a tablet/phone than a console. It throttles whenever it can. Once he was able to completely lock t the CPU clocks to max speed, the GPU would then throttle between 614 Mhz and 768Mhz. I would still like to see what the clocks do while gaming, but I think its likely that the clock speeds are not at maximum when playing games. Who knows, perhaps if a game doesn't push the CPU, then the GPU can stay clocked at 1Ghz, but if t the CPU is being pushed, the GPU lowers clocks. The fact that these test keep showing a 768 Mhz throttled speed under load seems to suggest that its the sweet spot, going higher isn't sustainable with such meager cooling.

Simply clocking high isn't a big deal, but clocking high under load is another story. The custom API is only going to allow utilization to go up. So at a given frequency, the Switch is going to actually be doing more work than the Shield TV can. Android is a thick API that doesn't allow developers access to the metal. I think in a few days its going to be pretty obvious that the Tegra chip powering switch is the best Nvidia's got within the design parameters, and not a cut down TX1 like some thought after hear clock speeds.

Dave also ran some test for bandwidth, and his test showed good performance at 1080p 8X MSAA. The scene looked pretty simplistic, but results didn't point to an obvious bottleneck with memory bandwidth.

Just a few more days......

Well, a few of us have been making the point that NX has to handle both the CPU and GPU being hammered - by code written explicitly for that platform - without throttling ever since the final clocks were revealed

Race-to-sleep throttle-me-do phone and tablet clocks were always going to be a poor stick to beat the revealed final clocks with.

On the BW front, it looks like CPU on shield should be able to pretty much saturate memory BW at full speed. Most tablet games won't do this, bit console games are likely to push harder.

 

[Goodtwin]

Given we now know that Shield TV and Switch's clocks are pretty comparable in practice, surely the question is how noisy it'll get when docked?

And if charging too, can you hold your head over it to dry your hair.

Rumors have suggested audible noise when docked, so I would guess that the fan does spin up pretty fast when docked. Where as the Nvidia Shield is said to be nearly silent even when gaming.

It's not even the best NVidia has because Parker would most definitely achieve better performance on the same power budget.
If anything, it's the best nintendo was willing to pay for, and/or the best NVidia was willing to part with.

Cost is a huge factor though. Rumors are pointing to a $250 launch price, and this is without losing money. Parker is also running a lot hotter than what is sustainable in the form factor for Switch. Listed clock speeds for the TX1 and the assumption that the Shield TV ran locked clock speeds being incorrect mislead us. So I think its fair to say that within Nintendo's budget for both price and thermal limits, the TX1, or a custom processor strikingly similar to the TX1 is the best Nvidia could offer within the scope of those limitations.

On the BW front, it looks like CPU on shield should be able to pretty much saturate memory BW at full speed. Most tablet games won't do this, bit console games are likely to push harder.

Wouldn't that be a nightmare scenario though? Meaning that the CPU is missing the L2 cache constantly. We also know that the A57 cores are clocked at 1Ghz, so bandwidth demands for Switch from the CPU can only be half the theoretical maximums for the A57. Also, just because the CPU can consume a ton of memory bandwidth doesn't mean it does so for very long. Point being that a 60fps game gives the hardware 16ms to complete a frame, the CPU may at some point consume a ton of bandwidth, but its not going to last very long. Could be for a ms or two, but not sustained for the duration of rendering the frame.

 

[Deleted member 13524]

Battery is 2.5 to 6 hours.
So maximum consumption drains the battery in 2.5 hours.
This is about the same as the 3 year-old Shield Tablet with a Tegra K1:
http://www.anandtech.com/show/8329/revisiting-shield-tablet-gaming-ux-and-battery-life

The TK1 has a single Kepler SMX at ~750MHz and is made on 28nm.
Plus, the Shield Tablet is substantially thinner than the Switch.



The games we saw.. zero multiplatform titles in the reveal. Mario Odyssey looked a bit mediocre in the "real-life" stage that was shown.