NVIDIA Tegra Architecture

[Blazkowicz]

It's funny. Now that's a "smart TV". It could be used in a bedroom or somewhere, can be used with a game console or DVD player.

The OS is dumbed down, but you don't have much choice for a commercial OS with support for mobile SoCs. Windows RT would work but Microsoft is probably not fond of the idea (and they made the OS free.. if the display is < 9 inches)

Tinkerers could try running something else. Desktop linux is obvious, but would be more workable with Tegra K1 I think. Ubuntu Touch and KDE tablet interface (Plasma Active) would be interesting to try.

Then I can see it been used for computers used to buy train tickets, and such.

 

[Exophase]

Who buys those things? What are they used for?

Well, I suppose you could alternate between using it as a desktop/laptop monitor (it has HDMI in) and a gigantic tablet that you lay on your lap or on the table. It'd at least be cheaper than getting both.

At one point I actually considered buying a monitor with a touch screen, I wonder if this could be used that way too?

 

[Ailuros]

http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=NVidia Mocha

 

[Deleted member 13524]

An 8" tablet with twice the performance of a Snapdragon 801 and Apple A7. Sounds like they're delivering on their promise.

This should be their Tegra Note update, so Shield 2 could have better results if it has a heatsink.

 

[ams]

NVIDIA Mocha Geekbench 2.4.3 benchmark:

http://browser.primatelabs.com/geekbench2/2431774

At roughly the same CPU operating frequency as [2.15GHz] Snapdragon 800, the [2.12GHz] R3 Cortex A15 scores ~ 30% higher in comparison!

 

[ams]

http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=NVidia Mocha

This link doesn't seem to work.

I found some screen captures elsewhere:

 

[Turbotab]

NVIDIA Mocha Geekbench 2.4.3 benchmark:

http://browser.primatelabs.com/geekbench2/2431774

At roughly the same CPU operating frequency as [2.15GHz] Snapdragon 800, the [2.12GHz] R3 Cortex A15 scores ~ 30% higher in comparison!

Given that it'll compete against a S801, with whatever evolutionary uarch refinements Qualcomm have extracted, running at 2.5 GHz means the IPC advantage of the A15 are largely rendered mute. I'm more interested in the overall power efficiency of both solutions, the GPU on the K1 is looking like a very solid effort.

 

[wco81]

Yeah I thought A15 wasn't a good core for mobile devices.

 

[Turbotab]

Yeah I thought A15 wasn't a good core for mobile devices.

Perhaps the combination of the ARM's R3 revision, TSMC's 28nm HPM and Nvidia's experience of the A15 in Tegra 4 may surprise us?

 

[ams]

Perhaps the combination of the ARM's R3 revision, TSMC's 28nm HPM and Nvidia's experience of the A15 in Tegra 4 may surprise us?

Yup, the R3 Cortex A15 variant used in Tegra K1 appears to have superior power efficiency compared to the Cortex A15 variant used in Tegra 4. And even though the max frequency could be as high as 2.3GHz in the former, using a max frequency of ~ 2.1GHz appears to give much lower power consumption and only slightly lower performance, so it really makes sense for a mobile device:

 

[ams]

Given that it'll compete against a S801, with whatever evolutionary uarch refinements Qualcomm have extracted, running at 2.5 GHz means the IPC advantage of the A15 are largely rendered mute. I'm more interested in the overall power efficiency of both solutions, the GPU on the K1 is looking like a very solid effort.

The issue with Snapdragon 801 is that it will need to run at a much higher CPU clock operating frequency to come even close to the performance of R3 Cortex A15, and even then I don't think that a 2.45GHz Krait core will match a 2.12GHz R3 A15 core with respect to performance or with respect to performance per watt.

 

[Exophase]

NVIDIA Mocha Geekbench 2.4.3 benchmark:

http://browser.primatelabs.com/geekbench2/2431774

At roughly the same CPU operating frequency as [2.15GHz] Snapdragon 800, the [2.12GHz] R3 Cortex A15 scores ~ 30% higher in comparison!

Do you have a link for the 2.15GHz S800 you're comparing with?

 

[wco81]

Wait, isn't Nvidia also developing a custom CPU core like Krait and that was what the K1 SOC would use?

Or this new A15 is that custom core?

 

[ams]

Do you have a link for the 2.15GHz S800 you're comparing with?

I used one of the highest scores shown for the Sony Xperia Z Ultra: http://browser.primatelabs.com/geekbench2/2210469

Some of the scores for Xperia Z Ultra are much lower than this one, but I'm not sure why that is.

 

[ams]

Wait, isn't Nvidia also developing a custom CPU core like Krait and that was what the K1 SOC would use?

Or this new A15 is that custom core?

NVIDIA's "custom" CPU core is Denver, and it is far different than both Krait 400 and Cortex A15. The R3 A15 is the latest and greatest variant of the A15 core. The efficiency gains compared to prior gen A15 core are due to architectural, process, and experiential factors.

 

[A1xLLcqAgt0qc2RyMz0y]

Wait, isn't Nvidia also developing a custom CPU core like Krait and that was what the K1 SOC would use?

Nvidia will have two different Tegra K1's.

The first one which has been released is the K1 with quad R3 A15's (and a fifth low power A15 core) and Kepler GPU.

The second will have dual 64bit custom Denver cores and Kepler GPU.

http://www.extremetech.com/gaming/1...-paired-with-64-bit-denver-cpu-coming-in-2014

 

[Wynix]

They really need to use different colours, green vs slightly different green is just idiotic.

 

[Ailuros]

GFXbench3.0 Mocha results dissappeared yesterday within minutes; I had already seen that some sites had captured screenshots so I didn't bother keeping any.

 

[Nebuchadnezzar]

Perhaps the combination of the ARM's R3 revision, TSMC's 28nm HPM and Nvidia's experience of the A15 in Tegra 4 may surprise us?

You're giving this too much credit to r3 / Nvidia is boasting it too much about. Basically what r3 brings (WFI & WFE) is the same as what Samsung has had in their chips for a year now. The HPm process is what'll bring the big benefit.

 

[Exophase]

You're giving this too much credit to r3 / Nvidia is boasting it too much about. Basically what r3 brings (WFI & WFE) is the same as what Samsung has had in their chips for a year now. The HPm process is what'll bring the big benefit.

Those are things that the TRM documents, yeah. But the TRM doesn't describe internal design beyond a few very high level details. So if a new feature or optimization isn't something that impacts visible behavior or can be configured (either at runtime or as part of the RTL) there's a good chance it'd go unmentioned.

On the other hand, the TRM does actually mention something significant outside of the ability to drop voltage during WFI/WFE:

Regional clock gating

Note

This feature is not available in revisions prior to r3p0.



In addition to extensive local clock gating to register flops, you can configure the Cortex-A15 MPCore processor to include regional clock gates that can perform additional clock gating of logic blocks such as the register banks. This can potentially reduce dynamic power dissipation.
You can set bit[31] of the ACTLR2 to 1 to enable regional clock gating for each processor. See Auxiliary Control Register 2.
You can set bit[26] of the L2ACTLR to 1 to enable regional clock gating in the L2, Interrupt Controller, and Timer. See L2 Auxiliary Control Register.
At reset, both of these bits are clear and the regional clock gates enables are tied HIGH. You must set these bits to 1 to enable additional clock gating in the regional clock gates for potentially reduced dynamic power dissipation.

And that is not something that Samsung could have already been doing. I'm not actually sure why this has to be enabled, maybe ARM is afraid of bugs in the implementation.

nVidia credits the changes in r3 as helping reduce dynamic power consumption, which is nothing that would apply while the processor is idle. But being more aggressive with internal clock gating could make a real difference.