2gig suggestions

AT&T isn't very transparent about it, but it's all Fiber and they do up to 5gig.

ok, but you'll receive a fibre modem, the ISP side might be scaled to 5G that's fine, but still the "client service" side can be tricky: they can give a fiber modem with 1 fiber port + 5 x 1Gb ethernet ports and the 5g fibre plan is actually shared across the physical 1g ports; or they can give you a device which indeed has 5g ethernet lan port as the service termination access point ... very-very different setup as latter case you'll need your client devices to have 5g ethernet port also to utilize full capacity. in former case isp might say you can put 5 distinct users all with max 1g ethernet, but all together is 5g traffic.

Edit: i think AT&T use BGW320 router/ONT/wifi combo device (https://help.sonic.com/hc/en-us/articles/1500000066642-ARRIS-BGW320) which looks has a 5G ethernet port.

If you do buy new stuff, the markup for 2.5GBASE-T network cards isn't that much (LAN switches are another topic), so if your ISP is offering 2 GBit/s contracts it doesn't hurt to be prepared. SQM is a software feature, it works on all hardware - but it needs considerable performance for highspeed links (so you do want fast router hardware for your situation, as in non-Atom x86_64).

As it's always easier to upgrade a contract, it does make sense to start out with a 1 GBit/s one and just have your end of the hardware prepared for more. But you need to decide what you want and how much you're willing to pay monthly (if the markup is negligible for 3 persons, why not).

Edit: a single-port 2.5GBASE-T r8125 Realtek card costs between 20-25 EUR (compared to 8-10 EUR for a 1000BASE-T r8168 one), speeds above that (and multi-port cards) drive up the prices quickly.

Would a MikroTik RB5009 be breaking the bank?

1 Like

this would be the best solution, but you can't find it anywhere at the moment.

The mini PC with a N5105 will easily support 2Gb with SQM. in a french forum someone posted some screenshots with ESXI and openwrt runing in a VM in this box :

see here :

hummm, he used SQM only for the upload, because he was having some upside/down between 500/800Mbps. with SQM he could stabilize the upload to 800Mbps.

Later he said :

For fun I did some download tests, I still had about 2100Mbps after several speedtest (without SQM).
I set SQM to limit my download speed to 2030Mbps, I put "2140000" at random and each speedtest was limited to 2030 afterwards.

I can clearly see the CPU impact, I consume 30%! (IDLE 70% during the speedtest).
And it works extremely well (see screenshot).

so we can say that with SQM on a 2Gbps connection we consume about 30% of CPU...And this is in a virtual machine.

And here : https://lafibre.info/remplacer-livebox/remplacement-de-la-livebox-par-un-routeur-openwrt-18-dhcp-v4v6-tv/msg957739/#msg957739
he made a stress test, and with SQM, during the stress test he made a speedtest. So even with the CPU at 100% he is close to 2Gbps. The network part has priority.

I'm really sceptical about those results…

In my local testing with iperf3, I hit the cliff (one core at 100%) at 830 MBit/s on my celeron j1900 (Rohde & Schwarz Cybersecurity gateprotect GmbH GP-7543, four Intel I211 1000BASE-T/ igb network cards). Both have the same 2.00 GHz base frequency, while the burst frequency of the n5105 is higher (2.90 GHz) than the 2.42 GHz of the j1900, all performance gains would be down to the process shrinkage and core improvements. While you might be able to convince me that the performance gains made in 7 years of Atom development would have closed the gap for sqm/ cake at 1 GBit/s wirespeed, I find it hard to believe that they are sufficient for 2.0-2.5 GBit/s, at least for comparing apples with apples (sqm/ cake, up-/ and downlink shaping) and not less intensive configurations (codel, only uplink shaping).

1 Like

but you're comparing a cpu dated from 2014 and one dated from 2021, even if the frequency is the same 2Ghz, the passmark is different :

i don't know, maybe there is something that i didn't understand...

1 Like

i contacted the guy from the french forum, i hope he will be able to clarify the things here.

I'm explicitly -from my very first response- mentioning my frame of reference, you will have to draw your own conclusions from that.

What I don't even attempt to do, is believing a single (weighed/ aggregated) benchmark or trying to put too much weight on those results for the specific use case of networking performance (routing + sqm/ cake in particular). But even if I did believe your benchmark - and if the requirements for 2.0-2.5 GBit/s could be calculated in linear fashion (which is quite a stretch) 830 MBit/s (j1900 sqm/cake - and that's really with the back to the wall), so the n5105 would have to be at least three times faster. Under the presumption that your benchmark is correct and would apply to the networking performance, that would just about fit - not much margin for comfort. And the real-world experiences (speedtests) you quote seem to underline that (despite reducing the requirements, it doesn't quite achieve 2.5GBASE-T wirespeed); forgive me if I miss the finer details, my french is limited.

Obviously there is a big performance delta between the Atom based j1900 and the Intel-core based c1037u (which doesn't even blink an eye with 1 GBit/s wirespeed sqm/cake), showing how much potential there would be to find for Intel in silicon improvements (both are of the same vintage). Obviously Intel had to play catch up with ryzen, so they might have shifted so performance to the newer Atom designs as well.

Would all of this suffice for the inherently single-threaded/ integer and I/O centric use case of providing 2.0-2.5 GBit/s routing performance with sqm/ cake? I don't know - and I'm not going to speculate (nor am I going out to buy hardware for performance tests). All I do know, is that a core i3/ i5 should suffice (even the 7 year old ivy-bridge base supports that interpolation). And the r4s performance figures (which comes with two a72 cores plus four a53 cores, but 1000BASE-T ethernet) do not instill much confidence in being able to route at 2.5 GBit/s wirespeed with sqm/cake, while the (not-yet-supported) r5s comes with 2.5GBASE-T network cards, it (afaik) lacks the a72 performance cores, so probably not going to cut it. Not enough data to guesstimate the RPi4's performance (especially as that would need two USB3 2.5GBASE-T cards).

The newer Celeron chips use DDR4 RAM instead of DDR3 on the J1900, maybe that is a significant difference. They also have a much faster internal GPU, which is completely irrelevant for a router but it may get included into overall scores.

I am the author of the SQM Benchmark with N5105.
What you missed is that the processor has a turboboost function.
This was enabled by default on my mini PC, so you have to compare with an 11th generation 2.8 GHz Intel processor.

But, to limit the energy consumption, I chose to deactivate this functionality and to work with a 2 Ghz processor.
I can confirm that I was only using 40-50% CPU during the 2Gbps speed test with SQM enabled, with a 2Ghz CPU clock.

Also, because this mini pc was overkill for openwrt, I installed vmware esxi and created an openwrt VM with 2 CPUs reserved and 1GB of Ram. I don't use SQM for download, so even with only 2 processors, it's still more than enough.


So you're actually suggesting buying a no-name box, with firmware on it that you possibly can't trust. Also because it's a noname box there will be no firmware / bios updates in the forseeable future. Then there's the matter of not knowing the build quality and if / or there are counterfeit parts being used. These are all big alarm bells going of in my opinion. Yes , it's dirt cheap, and it is for a reason.

Why not get something decent like this: https://www.supermicro.com/en/products/system/IoT/1U/SYS-110C-FHN4T

$265 bucks, approx 3.5x less than the Supermicro.


Yea...I should've also clarified we wanted something quiet, since it'll be three of us in a relatively small space, but I was under the assumption I wouldn't be setting up an entire server rack just for a router lmao

The lenovo M720q/M920q :+1: still waiting for my 4x 2.5G realtek card, i'll post some benchmarks when i'll receive it...

Everything that everyone need to know about modding this mini PC :

Depending on the CPU inside, a M720q can be found at 180$ for a I3 8100T up to 250$ for a I5-9400T
Then add 60$ for the quad port 2.5Gbps Pcie card or 130$ if you wan INTEL nic with 4x I225
20/30$ max depending on where you buy it for the Pcie Riser + Rear baffle/bracket
And finaly depending on your ISP, an external ONT with a 2.5gbps LAN port, which will cost about 40/50$ but there are currently none to be found anywhere.

btw about the 2.5Gb ONT, i saw in Taiwan they have offers where they lend this device :
it's called "Smart optical converter".

I see something else in that link Linksys E5750-AIS AX5400 WiFi 6, it comes with 2.5Gig ports. need more info about the hardware and future compatibility with OpenWrt

Close to the WAX206, except for the wifi :

1 Like

looks like this one...

Powerful 1.5GHz triple-core processor :face_with_diagonal_mouth: i'm affraid it's a Broadcom