5g/2.4g Router, 1000mb WAN/LAN

Hardware Questions and Recommendations
21

Yes, and the link points to your post. I used it just to illustrate how modern arm64 platforms are more efficient than x86 based routers and equally capable for this task. arm64 CPUs are very powerful and way more efficient, it is a no-brainer for the industry. You just have to see how Apple and Microsoft are quickly steering into the arm64 architecture.

We have a similar setup. Yes, I am also using SQM, an encrypted Wireguard tunnel for the whole lan, AdBlocking, and encrypted DNS requests using unbound. The average CPU load is 5% with a peak of 25% when using the whole bandwidth.

1 Like
22

Any good references to read up about this?

I'm intrigued. Can you elaborate on that?

23

It is not a secret, just google around and you will see how arm64 architecture is taking over the x86 dominance. Probably the most relevant data comes from Apple, which is migrating their desktop and laptops from x64 to arm64. regarding performance this is a sample graph:
https://www.extremetech.com/wp-content/uploads/2020/06/GeekBench-Mac.png

Regarding power, only the ultra low power Atom x86 platform can reach the power levels of a high-end aarch64 platform. And, at that level, the new aarch64 SoC outperform Atom. You can see how high-performance computation farms are switching also to arm.

Well, you can use Wireguard (or OpenVPN) both as a server, allowing you to connect to your lan from an external device, or as a client, connecting to a VPN service, and routing all the traffic through that interface to hide all your traffic and data to your ISP while anonymizing your IP. I prefer Wireguard as it is more bandwidth efficient. It requires more CPU than OpenVPN, but with this router, I have plenty of idle CPU cycles.

2 Likes
24

The problem with arm64 still remains, unless you're as big as Apple, Amazon, Facebook etc., you can't get the really good ARM hardware, but have to take what's left over for the low-cost/ low-performance SBC market. The ARM market is much more relying on made-for-purpose designs than x86_64 (yes, both are converging to some extent, but from different starting positions) and lacks the flexibility for the mere consumer buying in single digit quantities to customize the hardware to their demands (as in, add a RAM module here, a network card there - let alone software flexibility such as SBSA compliance), which leads to the workarounds (USB ethernet cards dangling from the RPi4) referred to by @Lynx.

Let's do a simple comparison.

I'm currently using a ~2013 vintage celeron 1037u (ivy-bridge) mini-PC (197 *197*29 mm³) with two onboard Intel 82574L 1000BASE-T ethernet cards, which is totally bored while doing SQM at 1 GBit/s. 15 watts idle isn't spectacular, but still bearable - baytrail-d (~6 watts) or haswell (~10 watts) would drop those figures, but the c1037u just was 'available' for cheap, when I needed something urgently.

If you go back to the 2013/ 2014 RPi offerings, you'd be stuck on the ARM11/ ARMv6 RPi1 (BCM2835) with 1*1 GHz, 512 MB and the dreaded USB2.0 system bus - you don't want that (anything below the RPi4 released in 2019) as router. Its immediate competition (Allwinner A20, 2*916 MHz cortex A7) wouldn't win performance contests either, nor would the late 2014 vintage Rockchip RK3288 (4*1.8 GHz A17, but whose mainline ecosystem only started quite a bit later than those of the former).

The RPi4 and the NanoPi r4s are now viable alternatives to x86_64 for 1 GBit/s with SQM, but availability and pricing doesn't leave them as much of an advantage as one might think. RPi CM4 + dfrobot baseboard/ case or the r4s/ case set you back around 150 EUR, the first new micro PCs with (usually) four ethernet cards start around 200-250 EUR - power usage being on a comparable level, x86_64 wins in terms of software compatibility. If you look beyond 1 GBit/s routing demands, x86_64 still wins easily due to mere availability and scalability (if Atom doesn't cut it, i3, i5, i7, i9 or ryzen will - need 10 GBit/s, just add in a PCIe card). Yes, we've seen announcements for ARMv8 servers with 10GBit/s ethernet ports for the last 10 years, they still remain unavailable to mere mortals and/ or blow up the budget.

Don't get me wrong, I'm hopeful for the ARMv8/ ARMv9 ecosystem, but I just haven't really found an ideal solution (be it as general purpose computer or fast-just-works router (ipq807x and mt7622bv do meet expectations, but performance is just sufficient, without headroom (e.g. for SQM) left), yet.

3 Likes
25

As you said arm platforms are made-for-purpose, and they are way better for specific tasks than a general-purpose system. That is exactly why arm is a better solution for this specific embedded application: a wifi router.
If you add the cost of the R4S + case + Wifi AX card + 5 port hub, you end up with a pretty expensive monster with way more cables, more failure points, and more power consumption than a stand-alone system. Whenever an upgrade is required, then it would be time to update the system. No need for a huge investment when there are solutions that check all the boxes.
I was like you many years ago, And I understand your point of view, however, I departed from it when I discovered that it wasn't actually providing any benefit over specific solutions.
Unless you plan the head router to perform other tasks such as web serving, bittorrenting, personal cloud storage, etc in the same device, having an x86 solution is a way to throw the money into a black hole.

1 Like
26

My current wireless is fine, as i have a RE450 AP upstairs and the TP link downstairs, and TP Link AP other side of house. Most demanding items are hard wired on 1gb switches

This is why the R4S could go between the modem and current router and change to AP

If i ever do upgrage Laptop to wifi 6, alls i need to change is 1 AP. If more devices downstairs require it, then id change that AP

I get theres more cables, but that doesnt bother me. but this way its more of componants can be upgraded where needed. I dont see cable company offering more the 1gb any time soon, nor do i currently see need for it. So the R4S should be futureproof for now, the Pi4 Compute would be a option but supply is poor at minute :frowning:

but the Pi4 + header and case would then be same price as the R4S, CPU is hard to judge, RAM can be same, but EMMC over SD card

27

And that's actually what I disagree with, whose purpose? That of the manufacturer or that of the customer, those aren't always the same.

Fibre connections are on the rise, with that comes (at least the possibility of) 1 GBit/s WAN speed, this is not an exotic (but still high-end) use case anymore, quite a few OpenWrt users however demand an effective QoS solution for that, to keep latencies and buffers at bay (gaming, VoIP, video conferencing, etc.). Now please show me the made-for-purpose wireless routers that can cover this requirement at 1 GBit/s.

  • mt7622bv can do 1 GBit/s, but apparently not SQM/ cake at 1 GBit/s (if I read it correctly, it manages around 550 MBit/s with cake)
  • ipq8071a (no NSS offloading) can do around 600 MBit/s without SQM, I didn't check the values with SQM/ cake.
    • yes, NSS offloading would improve these figures - and probably even manage nsscodel in hardware, but that's quite far away from being officially supported by OpenWrt
  • the RPi4 can do this (many success stories)
  • the NanoPi r4s probably comes close (I don't have reliable figures for this device at 1 GBit/s and SQM/ cake)

You simply can't buy a made-for-purpose wireless ARMv8 router that can manage sqm/ cake at 1 GBit/s at this point. The SOCs are designed to be cost effective and (just barely) meet the performance requirements of the big manufacturers (who don't object about proprietary acceleration methods not available to OpenWrt), there is no headroom left for advanced features the users might want.

ipq8064 is a quite good example for this, it's simply a secondary use of the apq8064 SOC designed for smartphones (e.g. Samsung Galaxy S4), it depends on NSS acceleration to achieve its rated performance. What can't be accelerated, or if NSS offloading isn't available at all (as in OpenWrt), will get a considerable performance penalty.

Neither the RPi4 or the RK3399 SOC powering the r4s have been designed with networking uses in mind, they just happen to be fast enough and have PCIe to support the necessary I/O throughput.

As mentioned before, the 9-year-old low-end c1037u ivy-bridge (according to Intel, "Vertical Segment: Mobile") can do this (1 GBit/s routing with SQM/ cake) easily (extrapolating from the CPU usage, it should easily manage 2.5GBASE-T, maybe even 5GBASE-T (probably a bit tighter), as well), but you can't buy an OpenWrt capable ready-made (made-for-purpose) router that would support this right now.

Likewise there are no (non-Apple/ non-MacOS) ARM options for the desktop or workstation market, so far at all. Prices and availability are only provided 'on request' at best. While the Apple ARM silicon does show what could be possible, it's not a contender for the i3/ i5 price range of whitebox desktop systems (nor -realistically- runs anything but MacOS so far (yes, efforts are underway, but not quite there yet)).

4 Likes
28

I hope the CAKE developers might implement the CPU optimisations referenced in their To Do list.

2 Likes