i would like to have a all in one road warrior * just travel and connect * router, switch, wifi ap, vpn server client, and vpn obfuscation server, openvpn 256 or 512 over shadow sock or any other kind of censorship circunvention, also firewall that is my use case model, and that is why i am asking here to the experts what could be fine, aparently banana router is very suitable
You had marked this a solved, so I was aiming my post at the slightly diverged dialog.
But, I will suggest to you that you also include parameters like bandwidth for each service, bandwidth needed for VPN, encryption type/algorithm to be supported and how much growth for the future for each item, that you are willing to pay for now.
Hopefully others can continue to recommend based on these updated criteria.
2 Likes
a legacy for the ones stepping behind our path!
1 Like
I'm afraid I don't agree. I have both the BPI-R2, based on the MT7623 which is a 1.4GHz ARM Cortex-A7 (successor to the A9) board, and the BPI-R64, based on the Mt7622 which is a 1.35GHz ARM Cortex-A53 board. The 64-bit A53 board gets almost exactly twice the integer performance per core as the 32-bit board, at essentially the same clock frequency.
My version of OpenWrt does. For example...
...Wireguard on Linux uses the existing crypto-api, which makes use of NEON. From lsmod on my BPI-R64:
chacha_neon 20480 1 libchacha20poly1305
poly1305_neon 20480 1 libchacha20poly1305
Also wireguard isn't the only new generation of crypto being used. For example, I use Syncthing on my main router as the hub for all my devices' file sync. Syncthing uses several primitives, including AES and chacha20, which both see a major boost on the A53.
The A9 is not a poor CPU. I use my MT7623-based device as a client bridge in my office, and it is also a decent little file server for my LAN. But single-core performance is important, as generally this will be the rate-limiting link due to interrupt handling. And at twice the performance-per-megahertz, my main house router's A53 is a major asset.
In any case, I think this discussion may be a little distracting for the OP.
3 Likes
you can all feel free to discuss here, if you want, until mod closes this thread, any information will be deemed as useful, have fun
The a7 was the sucessor of the a5 and was suceeded itself by the a53, this is from ythe power optimised in-order branch of the arm A family, so not representative of the out-of-order a9. Sure inside the power efficient branch successive designs improve upon each other, but that does not really change the point that within each generation the performance cores like a9, a15, a72 generally make short work out of the efficiency cores, and often the performanco cores still are not outmatched by the next generation efficiency cores. Like a9 often being better or equal to a53...
Well, great for your use case then. And then I agree decent choice. But the argument was about AES instructions....
ADDITION: a9 also has NEON units (and a7 might as well), so not sure this is relevant for the original a9 versus a53/55 discussion.
And now imagine you would actually compare a9 instead of your a7 with the a53 and you would not see a doubling of the integer performance.
Gentlemen, I am out of this discussion, I said what I wanted to convey, and I feel this is a waste of all of our time.
1 Like
Regarding A53 cores, I found it interesting that running the coremark benchmark (an available openwrt package) on my EA8500 (1.4 GHz ipq8064 Krait 32 bit dual core - ARM A15 based) yielded 5715 Coremark for 4.08 Cm/MHz, but my Belkin RT3200 (1.35GHz MT7622 Arm A53 64 bit dual core) yielded only 4677 Coremark for 3.46 Cm/MHz. I concluded my old EA8500 wasn't showing too much age ;-).
1 Like
Ah, one more example of arm's performance cores holding their own against the next generation efficiency cores 
That is however not all that surprising given the different primary goals of these two lines of CPU's. And this also does not make the efficiency cores bad CPUs... but it does show that 'newer must be better' is a tad too simplistic to judge relative performance of arm-based router SoCs, at least without benchmarking one's own relevant use-cases.
RockPro64 with an external NIC works very well in general and gives you a bit more options about what distro you want to run and plenty of performance. It's been ages since I tried OpenWrt but it should be fine on that platform too (images are available).
1 Like
With a dual a72 and a quad a53 this will offer adequate performance for 1 Gbps links (completely side-stepping our performance versus efficiency core discussion, simply by offering both efficiency AND performance cores from the same generation, as well as both neon simd (which a9 also offers) and cryptography extensions...) and a real PCIe slot to allow adding more ethernet ports easily via a NIC.
1 Like
Would you please not to mislead people by putting the Cortex-A9 NEON and the Cortex-A53 NEON in the same row. NEON on a Cortex-A9 is able to execute only 64 bits at a time, while NEON on Cortex-A53 can execute 128 bits at a time.
Moreover, it is not even a fact that your target (mvebu/cortex-9) is compiled against the NEON instruction set.
Could you share with us the output of the following command:
lsmod | grep neon
Thanks for the details. Yet the argument was a53 introduces something novel over a9 (or rather a7) and that was simply not factually correct.
And to add insult to injury your words were:
And that reads like you argue that v8's NEON is a novel improvement over v7, and not that the width of the vectors was increased from 64 to 128. And as a general statement unrelated to NEON the width of a vector unit does not necessarily tell you how fast it is, there are implementations that present themselves as N wide but internaly are N/2 wide and simply process the halves sequentially...
Which might or might not be true, you have no information about my compilation options, BUT this is pretty much beyond the point anyway IMHO. I am not trying to compete with you here, I am just trying to rectify some pretty fan-boyish sounding statements about different ARM cpu designs.
I am decidedly not saying the a9 is the best arm of all time, I am just noting that a53/a55 are not overwhelming superior to a9, and that this is no real surprise given the position of the two on ARM's cpu portfolio. Now, I might have been not as friendly and generous in this thread as I would have liked, but maybe we still could find an amicable way to end this thread?
Here you go:
root@turris:~# lsmod | grep neon
chacha_neon 32768 1 libchacha20poly1305
curve25519_neon 28672 1 wireguard
kpp 16384 1 curve25519_neon
libcurve25519_generic 20480 2 wireguard,curve25519_neon
P.S.: this used to be the level of discourse between intel and amd aficionados, I had not encountered that discussion style while discussing arm until now.
1 Like
This topic was automatically closed 10 days after the last reply. New replies are no longer allowed.