NVIDIA Tegra Architecture

[Ailuros]

Tegra4 is just the "fastest" tablet SoC around when actively cooled like Shield (so far); in any other case/implementation it's 10% less frequency for both CPU and GPU, where it's quite a bit below a S800 smartphone SoC performance and roughly on last year's iPad4 GPU performance level.

Let's call T4 the "hottest" tablet SoC around and I'll actually agree.

 

[ams]

Yes, it is. 3dmark won't heat up the SoC anymore than playing NOVA3 or MC4 at 1080p.

You do realize that looping 3dmark nearly a dozen times will heavily stress the CPU and GPU? It would be exceedingly rare for most Android games to do that (let alone for most Android games to run at an average of 30fps!). By now, it is a well known fact that most very high performance quad-core CPU ultra mobile SoC's will throttle frequencies when pushed beyond certain power and/or thermal limits. What is inexcusable is issues such as blurry camera and buggy software and flickering screen that have been reported on the Toshiba tablet.

 

[ams]

Tegra4 is just the "fastest" tablet SoC around when actively cooled like Shield (so far); in any other case/implementation it's 10% less frequency for both CPU and GPU, where it's quite a bit below a S800 smartphone SoC performance and roughly on last year's iPad4 GPU performance level.

IIRC, Shield ends up having ~ 20-30% higher application performance than other T4 devices. Part of that is surely due to higher clock operating frequencies, but part of that could be due to newer drivers as well (as Shield is ~ 15% faster now compared to the Shield benchmarks that popped up a few months ago). One of the things that I really like about Shield is that, per the NVIDIA forum on Shield, NVIDIA seems very committed to providing regular software updates for this device.

 

[ams]

but if this Toshiba tablet is pulling 13W I don't see how Shield with it's faster chip is only pulling 5W, which is what the battery life tests suggest.

13w is the peak power consumption for the entire system (including screen and all the internals). 5w is the TDP of the SoC itself. Don't confuse the two. The reality is that, since the ipad has a peak power consumption of nearly 13w, then that value is acceptable (and a peak power consumption level will not be maintained for any meaningful amount of time anyway). What is not acceptable is things like blurry camera, buggy software, and flickering screen.

 

[Ailuros]

IIRC, Shield ends up having ~ 20-30% higher application performance than other T4 devices. Part of that is surely due to higher clock operating frequencies, but part of that could be due to newer drivers as well (as Shield is ~ 15% faster now compared to the Shield benchmarks that popped up a few months ago). One of the things that I really like about Shield is that, per the NVIDIA forum on Shield, NVIDIA seems very committed to providing regular software updates for this device.

So what? On estimate a 1800/605MHz T4 should with the new drivers deliver under GLB2.7 around 20 fps at best for the moment in 1080p offscreen. iPad4 is now at 19 fps whereby it started out when GLB2.7 first appeared with by 28% lower performance. The S800 in the GalaxyS4 reaches 26 fps at the moment. And again who cares about the latter anyway when it'll take only a couple of weeks for Apple to introduce and ship their new smartphones and tablets, for which I'd be very surprised if the new iPhone won't be damn close to the iPad4 in GLB.

Else there are more than frequent compiler/driver updates (some of them with significant performance increases) from the majority of players out there.

 

[Exophase]

You do realize that looping 3dmark nearly a dozen times will heavily stress the CPU and GPU?

Not really. You can see from HardKernel's power consumption video (http://www.youtube.com/watch?v=LPOUekzdDAo&feature=player_embedded) that 3DMark barely utilizes the CPU during its GPU tests.

5w is the TDP of the SoC itself.

What does this mean exactly? That nVidia specified to Toshiba to provide cooling for 5W and some level of ambient temperature, and that it'll throttle to keep the temperature reasonable for this cooling?

Or do you think the SoC will only use 5W at the tablet's advertised specs of 4x1.8GHz + GPU at some level?

 

[Nebuchadnezzar]

You do realize that looping 3dmark nearly a dozen times will heavily stress the CPU and GPU? It would be exceedingly rare for most Android games to do that (let alone for most Android games to run at an average of 30fps!). By now, it is a well known fact that most very high performance quad-core CPU ultra mobile SoC's will throttle frequencies when pushed beyond certain power and/or thermal limits. What is inexcusable is issues such as blurry camera and buggy software and flickering screen that have been reported on the Toshiba tablet.

The only CPU intensive thing about 3DMark is the physics test, for any of the rest of the tests are basically CPU meaningless as any other 3D oriented benchmark. Actual games like Nova will have constantly high CPU load.

Shield having better performance is probably solely linked to it having a fan and nothing else.

Not really. You can see from HardKernel's power consumption video (http://www.youtube.com/watch?v=LPOUekzdDAo&feature=player_embedded) that 3DMark barely utilizes the CPU during its GPU tests.

Just FYI: The CPU being at 600MHz (A7 cores) is only because of a QoS design in the graphics driver who locks the minimum at that speed. Some stupid design from Samsung which I never understood as you can just remove it without any performance decrease and actually gain power efficiency.

 

[Ailuros]

Not really. You can see from HardKernel's power consumption video (http://www.youtube.com/watch?v=LPOUekzdDAo&feature=player_embedded) that 3DMark barely utilizes the CPU during its GPU tests.

ROFL and the GPU geek in me typically concentrates more on the GPU side of things than anything else. Muchas gracias for the excellent find. The Hardkernel clocks the 544MP3 actually at 640MHz it seems. The system measures actually a peak power consumption for the GPU at 640MHz of 1.9W while the average should rather be on estimate =/>1.6W. All we'd need now is someone use the Hardkernel XU in GLB2.7 to see how much the GPU burns in that one.

For the record according to Nebu the GPU of the 5410 runs in GLB2.7 at only 480MHz and gets a peak of 14.2 fps; scaled up to 640MHz assuming perfect scaling that's 18.9 fps even a tad more than NV measured on Logan on that infamous GLB2.7/Logan run. Oh and yeah sorry for the OT...

 

[ams]

3DMark barely utilizes the CPU during its GPU tests

The 3dmark Extreme Physics test does make heavy use of the CPU. The entire test suite (physics + graphics) was looped nearly a dozen times. Don't get me wrong, I dislike throttling just as much as the next guy, but it is a fact of life for these high performance quad-core CPU ultra mobile SoC's when pushed beyond a certain power/thermal threshold. For the vast majority of Android applications and useage scenarios on a tablet, I cannot imagine that a T4-powered tablet device would suffer from severe throttling (and part of the problem is that we just don't have a very big sample size of T4-powered tablets available right now).

Or do you think the SoC will only use 5W at the tablet's advertised specs of 4x1.8GHz + GPU at some level?

Of course not. 5w TDP surely does not refer to maximum peak power consumption, and it is also not even remotely comparable to system peak power consumption either.

 

[Nebuchadnezzar]

For the record according to Nebu the GPU of the 5410 runs in GLB2.7 at only 480MHz and gets a peak of 14.2 fps; scaled up to 640MHz assuming perfect scaling that's 18.9 fps even a tad more than NV measured on Logan on that infamous GLB2.7/Logan run. Oh and yeah sorry for the OT...

I already posted the scores for 700MHz quite some time ago and we concluded that the GPU is bandwidth starved in the Exynos thread; it gets 781 frames / 14fps at 640MHz.

The 3dmark Extreme Physics test does make heavy use of the CPU. The entire test suite (physics + graphics) was looped nearly a dozen times.

And it's irrelevant; the CPU load during the GPU tests is so low that the only thing the Physics test does is just raise the average over the duration of the whole loop. The GPU is barely loaded on the physics test. It isn't a situation where thermal constraints should matter. It doesn't reflect actual gaming where both are loaded at the same time.

 

[ams]

And it's irrelevant

I disagree. The 3dmark Extreme graphics and physics tests were continuously looped nearly a dozen times, which causes the GPU and CPU to heat up without giving them adequate time to cool down (and hence inducing power/thermal overload protection). And don't forget that on the Extreme setting, even when the SoC is cool, the average fps barely reaches 30fps, so this is a pretty stressful and processor-intensive test (and far more stressful than the vast majority of Android games, including some Tegra enhanced games too like Riptide GP2).

 

[Ailuros]

I already posted the scores for 700MHz quite some time ago and we concluded that the GPU is bandwidth starved in the Exynos thread; it gets 781 frames / 14fps at 640MHz.

The ODROID in question is actively cooled; if there's any thermal throttling going on on your device it's not necesarily comparable to that case.

 

[Nebuchadnezzar]

I disagree. The 3dmark Extreme graphics and physics tests were continuously looped nearly a dozen times, which causes the GPU and CPU to heat up without giving them adequate time to cool down (and hence inducing power/thermal overload protection).

Thermal dissipation happens instantaneously, thermal mechanisms work in the 100ms range. The duration and load times in-between the tests alone make the temperature drop 25°C before the GPU even begins to load up on the next test.

Your scenarios simply are not corresponding to reality.

The ODROID in question is actively cooled; if there's any thermal throttling going on on your device it's not necesarily comparable to that case.

There is no GPU throttling mechanism. I'll redo the tests when removing the memory throttling and update you on that at a later date. That being said, the temperatures speak for themselves in the ODROID video and that's my experience as well, heat isn't a problem for 3D benchmarks.

 

[ams]

Your scenarios simply are not corresponding to reality.

The 3dmark Extreme fps results speak for themselves. If you look carefully at the review of the Toshiba tablet, you will see that the CPU throttles much more than the GPU after continuously looping the Extreme test (ie. there is much more fps performance degradation in the physics test than in any of the graphics tests)!

 

[mboeller]

Shield having better performance is probably solely linked to it having a fan and nothing else.

could be the RAM too. The shield LP-DDR3 is specified to run at up to 1866 "MHz"

iFanatiker @ 3dcenter reported that the Toshiba tablets gets really hot and the Tegra4 starts to throttle when running Chrome only (surfing a few, more elaborate websites). ouch!

 

[Turbotab]

I already posted the scores for 700MHz quite some time ago and we concluded that the GPU is bandwidth starved in the Exynos thread; it gets 781 frames / 14fps at 640MHz.
And it's irrelevant; the CPU load during the GPU tests is so low that the only thing the Physics test does is just raise the average over the duration of the whole loop. The GPU is barely loaded on the physics test. It isn't a situation where thermal constraints should matter. It doesn't reflect actual gaming where both are loaded at the same time.

More evidence that the A15s are not the real culprit for T4 heat issues, if the A15 were really that bad then Nvidia could have shipped SoC with the CPU clocked at 1.5 to 1.6 GHz and easily matched whatever Qualcomm was shiping in early 2013 in terms of CPU.

So that leaves the GPU. Nvidia were adamant that T4's non-unified architecture was more power efficient at its launch than a unified solution, but now we are to believe that Tegra 5 unified solution is not only more performant, on the same process size, but with benefit of HPM, but will sip power and won't have T4's heat issues. I want to believe, but until I see in black and white in a shipping SoC, I just can't trust them.

 

[Exophase]

The 3dmark Extreme Physics test does make heavy use of the CPU. The entire test suite (physics + graphics) was looped nearly a dozen times.

Yeah, the physics test uses the CPU heavily while barely using the GPU. And the graphics tests uses the GPU heavily while barely using the CPU.

Saying that the test will heavily stress both just seems misleading since they'll never both be heavily stressed at the same time. Meaning at any given time it should be running a "reasonable" load.

 

[ams]

Well did you actually look at the throttled fps results in the review after 10 continuous loops running 3dmark Extreme? Both the CPU and the GPU are throttling, but the CPU throttling (reflected in the Physics test) is quite a bit more severe than the GPU throttling (reflected in the Graphics tests).

 

[ams]

More evidence that the A15s are not the real culprit for T4 heat issues

For an ultra mobile SoC, the Cortex A15 core is unquestionably both very performant but also very power hungry. In fact, the 3dmark Extreme Physics test was by far the most heavily throttled test (much more so than the two graphics tests) after running 10 continuous loops, which throws your theory into the water.

So that leaves the GPU. Nvidia were adamant that T4's non-unified architecture was more power efficient at its launch than a unified solution, but now we are to believe that Tegra 5 unified solution is not only more performant, on the same process size, but with benefit of HPM, but will sip power and won't have T4's heat issues. I want to believe, but until I see in black and white in a shipping SoC, I just can't trust them.

We already know that the Tegra 5 "Logan" Kepler.M GPU will be very power efficient (if initial results are anything to go by). According to NVIDIA's CEO, there is much more to Logan than just a good GPU. I would expect some improvements to CPU perf-per-watt too.

And again, the peak power consumption and peak temperatures are quite similar between the Toshiba tablet and the ipad 4, so it is pretty one-sided to say that T4 has "heat issues". The "heat issue" is that Toshiba placed the SoC right at or near the spot where the left hand holds the tablet, which was a poor design choice for this horizontally oriented tablet.

 

[Exophase]

Well did you actually look at the throttled fps results in the review after 10 continuous loops running 3dmark Extreme? Both the CPU and the GPU are throttling, but the CPU throttling (reflected in the Physics test) is quite a bit more severe than the GPU throttling (reflected in the Graphics tests).

Yeah, that just means both heavy GPU with light CPU and light GPU with heavy CPU scenarios are causing thermal overload. What's your point? You made it sound like it was just throttling because it was under unusual stress that's nothing like what could happen in the real world and I don't think this is the case. The CPU throttling could be more severe but the impact to the GPU score is really bad too.

But no, it doesn't surprise me at all that 4x1.8GHz Cortex-A15s would use more power than even a really power hungry mobile GPU. I bet that uses at least 6W on its own, at full tilt.

But if the GPU under heavy load while the CPU is under light load throttles then drops to 50% performance after being used for a while then what's the point of even having this level of GPU capability in this tablet? 10 runs of 3dmark is nothing compared to an extended gaming session. And if games can't and won't even utilize that level of performance to begin with then I still ask, what's the point?