NVIDIA Tegra Architecture

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

Battery lifetime and Long term performance test results added.

 

Anyone know what the battery size in this is, in comparison to ipad Air ?

 

Google says it's 6700 mAh. The iPad Mini 2 has around 6250 mAh and has roughly 1.5 times the battery life of the MiPad. Not bad for the MiPad I'd say, considering the performance.

 

I failed to note the screen size of the Mipad. Agree that comparing to the mini2 is much more logical.

 

There are only three data points so far for long term performance and battery lifetime:

6/29/2014: 26.4fps, 156 minutes battery life
7/05/2014: 29.9fps, 208 minutes battery life
7/08/2014: 35.9fps, 208 minutes battery life

The perf. per watt of the first and third data point are similar to each other, but the second data point is an anomaly. Hopefully we will see many more data points soon to compare and contrast. The nature of the test is imprecise due to variables such as brightness and anything else that may contribute to platform-level power consumption. As far as I can tell, these two tests use the GFXBench 3.0 T-Rex HD Onscreen demo.

Anyway, in general, based on preliminary results, the performance and perf. per watt of MiPad looks really good relatively speaking.

 

Performance/watt is lowest for the first data point and highest for the third.

 

Oops, that is correct, I misread the data. The third data point has ~ 36% more performance and ~ 33% more battery life than the first data point, so close to 2x higher perf. per watt in comparison! Thanks.

 

It's strange to see GLBenchmark 2.7 test as battery life test in GFXBench 3.0
Kishonti uses low precision FP16 shaders for T-Rex by default, that has quite noticeable impact on picture quality, just look at MP(medium precision) PSNR, hence the perfect MP PSNR for K1, looks like K1 doesn't support FP16. This combined with small amount of geometry, very simple shaders, lots of blending make this test perfect fit for tiled architectures. I don't believe lifetime results will stay the same once they move to Manhattan as power test

 

That is a good point. The multi precision render quality is much higher on Mi Pad than on iPad Mini Retina:

Mi Pad: 4478 mB PSNR
iPad Mini Retina: 2392 mB PSNR

Even the high precision render quality is significantly higher on Mi Pad than on Ipad Mini Retina:

Mi Pad: 4478 mB PSNR
iPad Mini Retina: 3529 mB PSNR

Note that the multi precision and high precision render quality of Adreno 330-powered devices is virtually the same as the iPad Mini Retina numbers above.

 

Kishonti always used the older test in their benchmark suites. As for the rest there are indications of throttling (which are unproven yet) and the iPad mini Retina is estimated to last twice as much with a smaller battery. If and when Manhattan will become a battery lifetime test for Kishonti's benchmark suite there will be most likely a lot of new contenders from all sides.

Finally a synthetic test that gives the GK20A over 60fps in T-Rex which is a healthy portion over 2x times difference to the iPad mini retina doesn't sound like the GK20A is choking in that test or is in any way at a disadvantage.

I'd suggest we'll wait for independent reviews of any piece of hardware and I'll make sure NOT to "influence" those in any way (and yes that's obviously a joke).

 

I see iPad mini Retina last 1.5x longer and I still think the comparison is tricky since we are comparing devices with different performance levels(K1 is 1.5x faster), image quality aside, there is not way T-Rex last times are indicator of battery life for tons of games capped at 30 or 60 fps, it's not an indicator for Tegra specific games like Trine 2, SS3, HL2, Portal, Rochard, etc either (some of them are deferred like Manhattan and will be more power efficient on K1). The whole comparison is pretty pointless, GFXBench numbers suggest me that I can get playable ~35 fps framerate for 3 hours long vs unplayable 21 fps framerate for 5 hours long on iPad Mini Retina in this specific test, so much usefull information

 

No one ever said that synthetic tests are anything more than an indication; exactly the reason why I suggest to wait for real independent reviews.

 

I think you meant unplayable 21 fps framerate for 4.5 hours long on iPad Mini Retina Anyway, imagine how pointless GFXBench 3.0 Manhattan battery life numbers would be, where the iPad Mini Retina achieves ~ 9/13 fps in Onscreen/Offscreen performance, respectively.

The only precise way to compare and contrast GPU perf. per watt is to measure GPU + mem. power consumption at the rails while presenting the same workload to each GPU.

In my opinion, NVIDIA made a sensible and practical choice in not aiming for a mixed precision GPU architecture with TK1. Modern-day desktop and console GPU architectures do not use a mixed precision GPU architecture, and modern-day games are not designed with mixed precision in mind either as a result.

 

Sounds to me like these tests would be somewhat more useful if there were various runs with the frame rate locked at different levels.

say @1,5,10,20 FPS etc and then also do a max frame rate run.

 

This would be data overload for most people and for most reviewers, and would add a lot of complexity to the benchmark. A 30 fps framerate lock might make sense, but TK1 is the only ultra mobile GPU to date that can achieve anywhere close to 30 fps in GFXBench 3.0 Manhattan Offscreen test (and the Onscreen test result is even lower than that, in the 20's).

 

In my opinion a frame rate lock below something like 24 fps, or make that 30 fps, seems somewhat useless. I understand some of the benefits, sure, but still....

 

I don't think it's all that easy to create a synthetic benchmark to measure battery life without there being objections or loopholes in the application. At the end of the day if anyone can play a demanding mobile game (demanding being always relative to what is available) for at least 3 hours as the specific synthetic indicates, then it's perfectly acceptable in my book.

 

Sorry, this is nonsense. I can't check right now but unless I'm very much mistaken battery lifetime and long term performance results aren't even related (or they certainly shouldn't be).

 

Here is what Tom's Hardware says about the battery lifetime and long term performance testing:

"Needless to say, GFXBench's battery test is something we're evaluating more closely. It measures battery life and performance stability by logging frame and battery discharge rate as the T-Rex test loops at least 30 times. The results are given in two scores: estimated battery life in minutes, and the number of frames rendered on the slowest test run (to gauge if a device is throttling). Both tests are run at the device’s 50% brightness level in the free Community edition, while the paid Corporate edition can be set to 0%, 25%, 75%, 100%, or whatever the device’s native slider is set to."

So if Tom's Hardware is correct, what this means is that the long term performance is simply the absolute lowest fps achieved during the 30 loop run, while the actual average fps achieved during the 30 loop battery lifetime estimate run will be equal to or greater than that same long term performance number.

In other words, the long term performance fps data point is the worst case scenario single data point and is not necessarily indicative of the average fps achieved during the 30 loop run that is used to estimate battery lifetime.

To make a long story short, the long term performance fps data point accurately represents the iPad Mini Retina's average fps during the 30 loop run, but slightly underestimates the Mi Pad's average fps during the 30 loop run. Most likely the Mi Pad is close to ~ 2x faster on average than the iPad Mini Retina using the T-Rex HD Onscreen test in this loop, which is a good result relatively speaking because it achieves "only" ~ 33% less estimated battery life, while using significantly higher precision render quality too in comparison.

 

Ah well, that doesn't strike me as a particularly good way of running a battery life test... but the options are relatively limited.

Not knowing the exact profile of the fps curve/throttling over time, it's not clear how much average performance matters. For some use cases I might only care about short burst performance, while for others the worst case performance after thermal throttling is relevant.