NVIDIA Tegra K1 Compared To AMD AM1 APUs

[deppman]

About that "warm" power supply

Assuming you know the design power, you can actually make a surprisingly accurate estimate about its load based on temp. Power supplies are designed to be most efficient at ROUGHLY half their design power, and worst at very high/low loads. IOW, temp is a pretty good measure of draw for switching power supplies. It would be better if he knew the exact temp, of course.

Sure, if he knew the power supply manufacturer and the model and the ambient temperature and he handled these bricks all the time, then he could tell what kind of power draw the device was pulling. This would be similar to how a Chef can judge how well a steak of a certain cut is cooked by how soft if feels when pressed by his finger. But this skill only comes after cooking many steaks and many cuts. Maybe you are a power-brick-chef. I highly doubt that Charlie is.

In any event, see my measurements I did in post #50. These are direct off the DC line, so the brick doesn't even come into play. And Charlie is so very, very wrong.

[dungeon]

TEST1 : glmark2 -s 1920x1080 --off-screen
  Score : 282 (Intel Celeron [email protected] = 151)
  ...

Why offscreen? Offscreen and onscreen have differents results, J1900 have 151 but onscreen .

Also glmark2 as part of phoronix-test-suite, give differents results then glmark2 from ubuntu repo .

For example with Athlon 5350 i have 236 score when i run phoronix version , but with ubuntu version score is 256, further if i run that off-screen i have score 405 .

[deppman]

Why offscreen? Because of video sync.

Why offscreen? Offscreen and onscreen have differents results, J1900 have 151 but onscreen .

Also glmark2 as part of phoronix-test-suite, give differents results then glmark2 from ubuntu repo .

Running glmark2 with the NV driver on-screen results in a 60fps cap due to video sync. Maybe that isn't an issue with J1900 , although usually benchmarks like this are run off-screen to avoid these problems.

And yes, I know the phoronix-test-suite is more consistent and would give more comparable results. But I can't get it to work. So I did the next best thing. As it is, the numbers certainly look in line with what we expect. We know the TK1 beats HD4400 graphics at least in some tests, and that the J1900 uses cut-down HD4xxx graphics.

An aside: This video driver is Beta. I expect performance to improve with future releases.

[tuke81]

Tests are in!
*snip*

FULL SYSTEM JETSON TK1 DC POWER ANALYSIS
  Includes
    * Audio in/out Active
    * Gigabit Ethernet
    * USB 3 driving Keyboard, Mouse, Logitech C210 Webcam
    * HDMI out at 1920x1080
    * Cooling Fan (no heat sink)
    * Installed 64GB SD card

BASE MEASUREMENTS
  Power-in Voltage : 12.15V
  Amperage and wattage
    Idle KDE Desktop : 0.22A ( 2.67W)
    Fan Amp          : 0.07A ( 0.85W)
    System Amp*      : 0.17A ( 2.06W) *NV measurement

    Idle Less Fan    : 0.15A ( 1.82W)
    Idle Less System : 0.05A ( 0.61W)

TEST1 : glmark2 -s 1920x1080 --off-screen
  Score : 282 (Intel Celeron [email protected] = 151)
  Power Measurements:
    Base             : 0.22A ( 2.67W)
    Peak             : 0.62A ( 7.53W)
    Observed Avg.    : 0.35A ( 4.25W)

    Avg. Less Fan    : 0.28A ( 3.40W)
    Avg. Less Sys    : 0.18A ( 2.19W)

TEST2 : CUDA Smoke particle demo
  Power Measurements:
    Base             : 0.62A ( 7.53W)
    Peak             : 0.91A (11.06W)
    Observed Avg.    : 0.88A (10.69W)

    Avg. Less Fan    : 0.81A ( 9.85W)     
    Avg. Less Sys    : 0.71A ( 8.26W)

TEST3 : VLC streaming 720p video from NAS GbE
  Power Measurements:
    Base             : 0.29A ( 3.52W)
    Peak             : 0.41A ( 4.98W)
    Observed Avg.    : 0.34A ( 4.13W)

    Avg. Less Fan    : 0.27A ( 3.28W)     
    Avg. Less Sys    : 0.17A ( 2.01W)

Nicely done, thanks! Peak power usage is kind of high, die shrink with 20nm process should cut that approx 30% to 8W:s still a bit high but low enough for tablet usage with even these clock(nvidia mocha tablet should be k1 A15 with 2.1GHz).

Love to see how denver k1 will be, and if it's 20nm or even 16nm finfets the better.

[deppman]

Nicely done, thanks! Peak power usage is kind of high, die shrink with 20nm process should cut that approx 30% to 8W:s still a bit high but low enough for tablet usage with even these clock(nvidia mocha tablet should be k1 A15 with 2.1GHz).

Love to see how denver k1 will be, and if it's 20nm or even 16nm finfets the better.

In the tests, this line shows estimated SOC + DRAM power use under the line [b]Avg Less Sys[b/]. The most taxing demo is the Smoke particle demo which appears to engage all CUDA cores. So let's look at peak average draw using these numbers:

TEST2 : CUDA Smoke particle demo
  Power Measurements:   
    Avg. Less Sys    : 0.71A ( 8.26W)
    Peak Less Sys    : 0.74A ( 8.99W)

Remember, peak is the highest observed number I saw on the meter. So we are looking at 8.26W for SOC + DRAM average in the most demanding demo. But this is the most extreme load. Just like a video card with a nominal 120W TDP can spike well above that, so too a SOC. For the most typical tablet uses - even playing most games - the SOC + DRAM numbers I saw ranged from 0.6W - 4.5W. This seems perfectly acceptable in a tablet, and certainly in Tegra 4 category, but with perhaps 30% bump in CPU perf, and up to 3x the GPU perf.

If 20nm shrink results in 30% savings, that should get us down to ~5.8W on the most demanding apps, and ~3.0W on average use, perhaps ~2.5W for web browsing. Those are good numbers.

Also, before anyone asks, yes, I expect Mullins to be competitive but not dominating. I think their 4.5W TDP claim is "nominal,' especially since AMD have been known to fudge their TDP claims as of late. We will only know when they ship actual parts.

I will probably do some more "typical gaming" power tests tonight.

[curaga]

While I agree with your numbers countering Charlie's earlier post, 5W is absolutely too much for a tablet. I have some 6W led lights with passive heatsinks, and they get hot enough to burn flesh.

The cooling ability of a tablet is likely less than that of those led lights. So if it ran at 5W for more than a few minutes, you'd be sterilized :P

[deppman]

A few more tests results

While I agree with your numbers countering Charlie's earlier post, 5W is absolutely too much for a tablet. I have some 6W led lights with passive heatsinks, and they get hot enough to burn flesh.

The cooling ability of a tablet is likely less than that of those led lights. So if it ran at 5W for more than a few minutes, you'd be sterilized :P

If 5W is too much for a tablet, then Mullins probably won't make it to a tablet either, judging by their "4.5W TDP" claim, which is probably like their "250 TDP" claim on the 290x, which actually is closer to 300 TDP (per Anandtech).

Anyway, I complied Xonotic for the Jetson, and also played with Web browsing using the multimeter patched into the DC supply (as before). The game was definitely playable; I would estimate an average 35FPS.

TEST 4 : Xonotic v0.7.0 normal effects GLX 1920x1080
  Frame rate in-game (single player 1) : 20-50fps (35FPS avg.)
  Power Measurements:
    Observed Avg.    : 0.56A ( 6.56W)
    Avg. Less Fan    : 0.49A ( 5.95W)
    Avg. Less Sys    : 0.39A ( 4.74W)

TEST 5 : Web browsing (Slashdot.org)
    Observed Avg.    : 0.35A ( 4.25W)
    Avg. Less Fan    : 0.28A ( 3.40W)
    Avg. Less Sys    : 0.18A ( 2.19W)

Again, in tablet form with lower power memory and I/O, I think a fair estimate is 2.0-4.5W draw for the SOC + RAM. This looks to be on par with the Tegra 4 judging by published numbers on the Shield, the Tegra Note 7, and the ASUS TF701T. Yet the K1 has 2-3x the graphics performance and ~30% better CPU.

I think proper thermal engineering is required to your point about heat. All three T4 devices named above don't have problems with heat management, and two of them are passively cooled. I know because I own 6 Android devices including those three. Interestingly, the worst device I own is the Nexus 4, which throttles within a few minutes of running a demanding game. The best is the TN7 which rarely even feels warm to the touch (my guess is it dumps heat out of its bass resonance port) . OTOH, the T4-powered Toshiba Excite does have a problem throttling and shutdowns, so how much care one takes in routing the heat must be the deciding factor.

[dungeon]

If 5W is too much for a tablet, then Mullins probably won't make it to a tablet either, judging by their "4.5W TDP" claim, which is probably like their "250 TDP" claim on the 290x, which actually is closer to 300 TDP (per Anandtech).

Seems like anandtech does not read wikipedia max TDP column for Radeon R9 290X says 290W .

http://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units#Volcanic_Islands_.28Rx_200.29_Series

TEST 4 : Xonotic v0.7.0 normal effects GLX 1920x1080
Frame rate in-game (single player 1) : 20-50fps (35FPS avg.)
Power Measurements:
Observed Avg. : 0.56A ( 6.56W)
Avg. Less Fan : 0.49A ( 5.95W)
Avg. Less Sys : 0.39A ( 4.74W)

What fps you have in Xonotic benchmark? Let set it to normal and then run the benchmark... you have here how to run that benchmark and some others

http://dri.freedesktop.org/wiki/Benchmarking/

[deppman]

Xonotic "the big benchmark" results

Seems like anandtech does not read wikipedia max TDP column for Radeon R9 290X says 290W .

http://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units#Volcanic_Islands_.28Rx_200.29_Series

What fps you have in Xonotic benchmark? Let set it to normal and then run the benchmark... you have here how to run that benchmark and some others

http://dri.freedesktop.org/wiki/Benchmarking/

AMD representatives first claimed that the R9 290x card was 250W "in typical gaming scenarios" as reported by Anandtech. After shipping, they may have changed their story:

However at the same time how power consumption is being defined is getting far murkier: AMD doesn?t list the power consumption of the 290X in any of their documentation or specifications, and after asking them directly we?re only being told that the ?average gaming scenario power? is 250W. We?ll dive into this more when we do a breakdown of the changes to PowerTune on 290X, but in short AMD is likely underreporting the 290X?s power consumption (Emphasis added). Based on our test results we?re seeing 290X draw more power than any other ?250W? card in our collection, and in reality the TDP of the card is almost certainly closer to 300W. There are limits to how long the card can sustain that level of power draw due to cooling requirements, but given sufficient cooling the power limit of the card appears to be around 300W, and for the moment we?re labeling it as such.

Now for the Benchmarks. I was already running "the big benchmark" when you last posted because I was uncomfortable with my test as I was estimating the average FPS. When I ran it, the entire application recompiled, and it reported nearly twice the FPS I was seeing. I couldn't believe it, so I opened up the app and tested it myself - it really is running now at twice the speed, and very smooth at that. Apparently, the compile flag I gave previously ("glx") was not the best choice!

So without further ado, here are the unaltered results:

Benchmarking on omg
MED: 10510 frames 121.4832903 seconds 86.5139557 fps, one-second fps min/avg/max: 48 89 135 (336 seconds)
Benchmarking on low
MED: 10510 frames 132.8224751 seconds 79.1281746 fps, one-second fps min/avg/max: 43 83 140 (336 seconds)
Benchmarking on med
MED: 10510 frames 150.7966910 seconds 69.6964896 fps, one-second fps min/avg/max: 35 75 131 (336 seconds)
Benchmarking on normal
MED: 10510 frames 159.5820060 seconds 65.8595556 fps, one-second fps min/avg/max: 34 71 120 (336 seconds)

That's right: 71 AVERAGE FPS @ 1920x1080 with normal detail. It wouldn't run "high" detail or above; this is almost certainly due to the constrained shared memory. Maybe I can squeeze it to run in high - and get better FPS - if I shut down Chromium before running next time. I had meant to this time, but Xonotic grabbed the screen before I had the chance.

On "low" detail settings, the TK1 runs 83FPS average compared to 63FPS on the J1900 @ 2.9GHz, or about 32% faster. While the J1900 claims 10W TDP @ 2.0 GHz, I'm guessing we are seeing more than that for a 33% boost in clock (13.3W?). The TK1 shows < 5W for SOC + RAM. It also beats all the recent AMD AM1 25W TDP chips while pushing 2.5x the number of pixels (1920x1080 vs 1024x768)! Assuming these tests are comparable to the Phoronix setup, that's very impressive!

[dungeon]

Yep Xonotic benchmark from phoronix suite is quite comparabile if you run it that way . But keep in mind you compared nvidia blob vs opensource radeon driver there . Also you run 32bit ARM, and all results Michael do is 64bit x86... and there can be small difference if you test on 32bit or 64bit OS and also which xonotic executable you run or if you compiled it (as you are) there also can be some difference, etc. .

So without further ado, here are the unaltered results:

Benchmarking on omg
MED: 10510 frames 121.4832903 seconds 86.5139557 fps, one-second fps min/avg/max: 48 89 135 (336 seconds)
Benchmarking on low
MED: 10510 frames 132.8224751 seconds 79.1281746 fps, one-second fps min/avg/max: 43 83 140 (336 seconds)
Benchmarking on med
MED: 10510 frames 150.7966910 seconds 69.6964896 fps, one-second fps min/avg/max: 35 75 131 (336 seconds)
Benchmarking on normal
MED: 10510 frames 159.5820060 seconds 65.8595556 fps, one-second fps min/avg/max: 34 71 120 (336 seconds)

I have Athlon 5350 and those results looks to me "the same" as one can get with Athlon 5350 (at default 2050MHz clock) runing fglrx actually on 32bit - so blob per blob results are quite the same... but 64bit x86 results are slighly better here . With radeon driver we are somewhere at 80ish percentage level of fglrx (i mean with current stable/released graphics stack ;D).

So Tegra K1 in this case with Xonotic smells more like Intel HD3000 on Windows . Must say i expected more .