Benchmarks Of The 24-Core ARM Socionext 96Boards Developerbox

[-MacNuke-]

*People say it doesn't have branch prediction at all, but the technical reference manual for it says otherwise.

It does do branch prediction only for instruction fetches. But since it does not execute them it is immune to Spectre and Meltdown.

[boboviz]

Not bad for a developers box. I sometimes think people mis the point of developers boxes They are for software development.

All developers i know use Raspberry 3 or Arduino. For less than 50 USD.

[starshipeleven]

It could be interesting if it cost 12 USD, not 1200

1200 is pretty affordable for a devboard.

12$ is for mass-produced tinkerer toys.

[starshipeleven]

All developers i know use Raspberry 3 or Arduino. For less than 50 USD.

If they can do fine with a Raspi, then they aren't developing intensive applications, so your comparison is invalid.

Really, this thing is trading blows with the Jetson from Nvidia which isn't cheap either, or getting close to low midrange desktop processors depending on task, while using a 5w SoC.

[boboviz]

Really, this thing is trading blows with the Jetson from Nvidia which isn't cheap either.....

Uh, Jetson Tx2 costs the half and goes the double.

[coder111]

Hmm. It would be interesting to run PTS on a high-end mobile phone and compare the results. Say something that has a Snapdragon 845 in it.

I wonder if it would it be possible to run PTS in a chroot environment on android? Has anyone attempted it?

With regards to this particular CPU- it's another attempt to shove many low-end cores into a single chip and call it a day. I'd rather see some improvements in single-threaded performance for ARM. I guess it could still be used as a dev box or some very specialized server, but it won't compete with x86.

[Zola]

These are low end cores even for a cell phone.
If you're not just trolling, you're reading too much into this.

The A53 is a small in order core & with a bad design flows (corrected with revision 2).
Thing that you need to understand is how SMP is done on FP units. NEON (128 bit) SIMD find on A53 is rather slow, old and dose not meet IEE duble precision requirements. For comparison A55 NEON SIMD blocks is 40% faster & has duble precision certification. Still ARM is behind Intel regarding SIMD's but not for long judging by recent SVE SIMD announcements (Fujutsu A64FX) & the 512 bit one is only the first one to come as they will go up to 2048 bit.
This is an SoC which is far from being built on cutting edge manufacturing processes, with an old core's, doubtful CPU topology and interconnect, bad implemented memory controller & ridiculous price.
Imagine an A55 pared with 512 bit SVE SIMD that would use 50% more power & achieve 3~4 x more FP performance. Now put it up on normal FinFET manufacturing process & bump the clock speed to sustainable leakage level for a FinFET based structures which is 2~2.2 GHz and you have a gain of at least 6~8 x performance. Now you have projections what ARM is capable to achieve & with an SoC that is in power consumption in line with ultra portable laptops requirement. So I hope you see a HPC many core CPU future more clear now.

[Michael]

Hmm. It would be interesting to run PTS on a high-end mobile phone and compare the results. Say something that has a Snapdragon 845 in it.

I wonder if it would it be possible to run PTS in a chroot environment on android? Has anyone attempted it?

With regards to this particular CPU- it's another attempt to shove many low-end cores into a single chip and call it a day. I'd rather see some improvements in single-threaded performance for ARM. I guess it could still be used as a dev box or some very specialized server, but it won't compete with x86.

I've done PTS on Android many years ago as a quick experiment. Recently someone was working on PTS on Android via termux and whatever else, but I haven't done anything recently due to no spare Android hardware / no commercial customers paying for it to allow me time to work on it.

[coder111]

I'm just curious what is the state-of-the-art performance of ARM, which these days seems to be mostly high-end mobile phones.

With regards to PTS on Android and branching into mobile phone hardware reviews- it would be interesting direction to take for Phoronix. However I do understand you probably lack both time and funds to do so. And I'm not sure whether it would be worth the effort. PTS specializes on Linux & more or less open hardware which mobile phones certainly aren't.

Besides- while CPU benchmarks should run in chroot environment without that much effort, all the 3D stuff would probably not work without spending lots and lots of effort and you'd end up with apples to oranges comparison anyway...

Which reminds me- I should get Phoronix Premium. I've spent enough time on your site already, and I really appreciate all the effort that went into making it.

I've done PTS on Android many years ago as a quick experiment. Recently someone was working on PTS on Android via termux and whatever else, but I haven't done anything recently due to no spare Android hardware / no commercial customers paying for it to allow me time to work on it.

[Weasel]

Yeah, it kinda is bad. Look at its kernel compilation benchmark, in comparison to a dual-core i3! Not to mention that developers are often doing incremental builds where you actually want decent single-thread performance.

Exactly this. People who just do "./configure && make" think they're "developing" but that's not the case how it's done at all. We don't compile from scratch all the time, only the few files that changed and then link them (which is cheap without LTO).

Only with LTO it gets compiled from scratch again, but that's only done once at release time (if you're even using it, most projects still don't). Which is not "during developing".

Lastly, seriously who cares how low power it is? Did you look at the price? You're not going to save $1000 in power bills with it.