AtomicPi is slower than Odroid

May 13, 2019 - 1:07 AM

  • The main selling point of Intel-based SBCs is higher performance at the cost of slightly higher power consumption. That was also the reason why I decided to buy AtomicPi. Unfortunately, computational benchmarks show that the board not only does not beat ARM-based board, but is even slower. Here is a comparison of performance of APi and Odroids using SciMark benchmark from Phoronix suite. Highlighted are APi and a comparable setup for Odroid: https://openbenchmarking.org/result/1905132-KH-ATOMICPIS21,1905090-HV-1904268HV81&obr_hgv=Odroid+Ubuntu+18.4+LTS+GCC+7+Linaro&obr_sor=y&obr_rro=y&obr_hgv=Odroid+Ubuntu+18.4+LTS+GCC+7+Linaro

    The board was powered from a lab power supply, so I could monitor current draw.
    Even for benchmarks using all 4 cores at 100% load the measured current never exceeded 2A, current limit set to 3A never activated (so even at peak the draw never exceeded the limit), and core temperatures indicated by sensors, even for the most computationally intensive benchmarks, were in the range 55C-58C. Voltage measured on the contacts of the Molex power plug (on the big breakout board) was holding at steady 5 volts. Based on that it's safe to assume that the CPU was not throttled.

    In most of the (except FFT) tests AtomicPi was either the slowest board, or very close. It's not clear which Odroid board boards were used for testing, but even for the most powerful, XU4 the results are rather surprising.
    One thing I noticed is that the processor does not run at full speed. Max CPU frequency reported by cpufreq-info was 1.68GHz, much slower than highest reported clock frequency 1.92GHz. Is the cpu speed limited in bios to reduce power draw? Is there a way to unlock full performance?

    The board is running stock Ubuntu 18.04.1 LTS with kernel 4.15.18-dli, with all system updates installed.

    0
  • I think you missed the point, it was supposed to be a small x86 board that had roughly the same performance as the XU4. It was about size of the board and it being x86 and running standard distributions, including Windows, not per se the power usage / It's going to be 8x faster. For those intending to use video I am sure the experience will be much better seeing as they don't have to mess with driver blobs and goofy stuff to make things work.

    While I am an ARM fanboy at heart, I think your wasting your breath assuming this was to replace an ARM board in regards to power usage and overall performance for CPU bound things. The GPU will win out hands down here along with the ease of use factor since it's x86. The barrier to entry is much lower than that of an ARM device also.

    Additionally, your not going to get it to hit max clock speed without some type of active cooling. The passive cooling is fine for most use cases, but if you intend to get max performance and turbo boosting out of the CPU then your going to need to add a fan to the setup. This is because it thermal throttles rather quickly, so with passive cooling, while it does a reasonable job, your still going to see thermal throttling under heavy load (meaning you won't hit max frequencies very often). Also, if you want to reduce the overall power foot print, disable turbo in the Bios (or enable it to get turbo features if you haven't done this yet), this will drop usage down a bit but you will be stuck with the base clock across the board.

    my 2 cents.

    Cheers!

    This post was edited 2019-05-13 01:25:27 -0400
    0
  • Erm... I don't think I waste my breath. Performance was one of the most important selling points listed on the kickstarter page. While boasting that the board is faster than RPi is kind of unfair (it's hard to find an SBC which will be slower than RPi), it claimed to beat even some of the desktops. I understand that passive cooling may require some compromises, but the board did not even come close to the point where the CPU would throttle, and the heatsink was barely warm. I don't mind adding a small fan to the heatsink, but I would like to see the CPU running at full speed.

    0
  • For what type of workload though? That is what I think would be key here. I believe there are some other stats out and about that show it does perform better than ARM under specific workloads, one of those being video (e.g. having a actual video driver ).

    I guess I understand your frustration some, but to have a x86 SBC at $35.00 and available with free shipping which is equivalent in some way to a cheap ARM board (exposing GPIO / UART ) is pretty impressive. Any equivalent TV box based on the same SoC on Amazon is still easily double the price. It isn't like this SoC or layout is new or even really that unusual, it has been available in many Chinese tablets and TV boxes for years now. The main attraction here is they have lowered the cost of entry and provided GPIO / UART where there usually wouldn't be any such available on a commercial product. That and included a 'decent' pre-packed Linux experience, where there is known to be some bugs in the video drivers in certain Linux versions (18.04) .

    Additionally, I would highly suggest you upgrade to the 18.10 release and test again if you haven't yet as well, the 18.04 image if for the dogs and the video driver in it has bugs which cause hard reboot when trying to use full screen buffer in some cases.

    my 2 cents.

    Cheers!

    This post was edited 2019-05-13 10:19:22 -0400
    0
  • I haven't tested the Atomic pi yet because I haven't received all my parts and don't want to cobble it together but! I have another small x86 box based on the exact same processor that beats my XU4 bay 2.5X. So I would expect if not exactly the same at least very close performance. The machine I used in this test was a Z83 tv box. Once I get this ticking I'll certainly report back my results.

    0
  • 1.44GHz is base frequency, 1.92Ghz is burst frequency.

    I would be nice if they made an honest statement as to the real frequency of the processor.

    And we are comparing a generic x64 distro with one specific for the arm processor.

    And honestly if you need a x86 platform, the UP has more less the same specs but for more $$$$

    0
  • Well sysbench says it's 2.5 times faster than my XU4 and also faster than my Z83 with the same cpu, and while that's not anything approaching a complete test it's also one that's not relying on the os and therefore not prone to issues with it either.

    0
  • It has been noticed that different versions of a realtime kernel, on the same distro & machine (not particular to the Atomic Pi), can lead to different latency results.

    1