Everything You Need To Know About The NVIDIA Jetson TX1 Performance

[SystemCrasher]

That's just your cluelessness that strikes again.

I've thought it's rather your arrogance. How many systems you've actually implemented using this thing? FYI, I've somehow noticed: ICs like Tegra3 and beyond weren't exactly popular... and I think there was a good reason.

1. The kind of people who are purchasing those boards aren't interested in x86.

Very strange goal on its own. Sure, x86 haves some crappy disadvantages. But just "x86" on its own is probably most silly reason to reject something. Sure, it instruction set is a crap, especially in 32 bit mode, but humans usually do not face it, unless they are up for hardcore assembly. And so, can you imagine better reasons? I can name like a 5 of far better reasons. Because I'm clueless noob, etc.

2. That "oversized board" is merely a reference design. Hardware designers can and will re-implement it in a smaller form factor.

Would not work due to thermal issues. Size of heatsink clearly gives idea of its thermal properties. Beats early Pentium CPUs any day.

They sell the CPU modules separately for a reason, after all.

I wish them good luck in sales. They would need it with their module prices and portfolio...

3. <10W is hardly "power hungry".

Of course, if you compare it to Xeon. But nobody sells Xeon as CPU module. OTOH for mobile designs, 10W is quite a lot of power and puts some demand on battery, power circuitry design and so on.

So this thing is, ahem, very specific and unlikely to see widespread adoption. And it would not allow to cut IC & modules down either. So it can easily end like with rest of NV ARMs.

[SidR]

I can see quite a few issues with this article, starting with benchmark selection. At least FFTW, John the Ripper and libvpx have architecture-specific implementations, using SIMD and maybe hand written assembly routines. JTR and FFTW are definitely using different implementations between ARM and x86 in this test. If someone is interested in using JTR or FFTW on these platforms, then I suppose the results are relevant for them. However, these results can't be used to compare hardware performance between ARM and x86 or any other two architectures.

Then some results themselves seem fishy on TK1. John the Ripper running only 10% faster on TK1 compared to Raspberry Pi 2? Either the system is misconfigured, JTR isn't built correctly, or that benchmark isn't really measuring CPU performance. On my own Jetson TK1 I get 12 fps for the VP8 libvpx benchmark, which is 2.5 times higher than the result you're reporting and actually 1.4 times higher than the result reported for Jetson TX1. Smallpt on my TK1 finished in 724 seconds, 4 times faster than reported.

[bmerry]

The CPU numbers for the Jetson TK1 are far too low. Perhaps it was running on the low-power single-core cluster. I've seen some reports of issues where SMP code sees only the one CPU, launches one thread, and then even if cluster migration happens there is still only one thread. My experience generally with the TK1 was also generally that power management could be too conservative to spin up to full speed and benchmark results were flaky until one overrode power management to force full speed all the time.

I've posted some CPU results comparing TX1 to TK1 here. The TK1's CPU actually outperforms TX1 on a number of benchmarks. It loses on C-Ray and smallpt, but ray-tracing is often memory-bound and the TX1 has much higher memory bandwidth.

[speculatrix]

It's February 2016. Pretty much a year since Nvidia first talked about the x1. All we've seen so far is a simple games console/media streamer, and an expensive unpopular Android tablet, if you ignore the expensive rare Jetson units.
Why isn't the x1 in more things?
I should be able to buy an x1 Chromebook, an x1 Chromebox, more tablets etc.
What's gone wrong with Nvidia's rollout?

[AlexSt]

Originally posted by robclark View Post
"btw Michael, where do you get gfxbench for arm linux?"

Hey Rob, no idea.... I never found any non-Android Linux ARM binaries for it, but NVIDIA sent it to me on an SD card last week. I can try to find out or if they don't mind if I send it over to you.

As of March 2017 there is still no Linux/ARM64 version publically available at the home pages of GFXBench whilst NVidia is shipping its Tegra series "Parker" SoC in the two board versions of the Drive PX2 and is just going to ship it as well with the card sized Jetson TX2. Due to the high level of similarity & compatibility the GFXbench package should be runable there with almost no fuzz, regardless who compiled it. So i have no big doubt it will be available there fully or at least in parts for usage. Just the media and method for providing it might have been adapted to the specific target.

@Michael: Are you going for any articles in the above mentioned area?