RB962UiGS-5HacT2HnT (hAP ac) low performance

Hi,

I've just decommissioned my 2 glorious Asus WL500gP, they've been serving me well for (almost) 20 years, 24/24, thanks to OpenWRT. I set back their original minipci cards and wet-eyed placed them on the shelves in front of my eyes behind the monitors, to remember those good guys serving me so much pr0n.... ehm... the best money I have ever spent on electronics.

So I bought 2 RBM33G and 1 HAP AC.

The RBM33G armed with 2x WLE900VX each are a piece of cake; can work out at L3 multiple full gigabit streams and there's spare cpu power still. Flow offloading gives them an extra 10-15% headroom... so sweet.

The HAP AC instead looks pretty crappy. Iperf3 tells me 300-350mbit/s (L3 routing, no qos, no complex mangling; just routing between 2 interfaces) and sw/hw offloading does nothing; top reports a ton of sirq and sys usage. Am I doing something wrong or that's the QCA9558@720MHz normal throughput? I had better perf on the RouterboardPro with similar CPU!

I've to admit I'm using the ar71xx target as I couldn't figure out whether one of the ar79 targets could work on this board or not and couldn't open the box to avoid loosing the chance to return the product. I've already initiated the return procedures with the seller but I'd like to be sure I'm not missing something and, above all, I've no idea what to buy in place of that little router. Any advice on another board with similar specs and price (~120-150€)?
Over there I need 3+2 gigabit ethernet (3 routed+2 switched), wifi and poe optional (as I can add external injectors and radios; the poe is way easier and reliable with external passive injector rather than fighting with AVR programming the onboard POE switch); don't need a 5 gbit backplane but ... 1gbit is needed; openwrt isn't strictly needed but I prefer to buy openwrt-compatible stuff even if I don't plan to use it straight away: after a few months, always pops up some extra feature missing from default firmware, and openwrt goes to the rescue...

Can't find MT7621 (ie: flow offloading) devices with the right specs, and can't figure out what platforms support flow offloading. Overprovisioning: the pcengines APUs are cool but the mpcie slots looks a bit weird (USB only, storage only... whatever); and a bit more expensive than RBM33G. Other x86 devices (ex: chinese VESA-sized 'NUCs') on aliexpress have single ethernet or cost 3 times more; and I'd like to avoid USB-Eth dongles.
And can't figure out the openwrt support status for other devices, ex: CSS610-8G-2S+IN would be awesome (as I have a couple of 10G devices and soon there will be a NAS) but looks like a plain switch with no L3 chances (dedicated asic instead of a general purpose cpu, 64k flash?).

Currently the only option I have is to buy a third RBM33G, a 20$ gigabit switch, and a couple of POE passive injectors; but it's a PlanB, not a real solution. HELP!

EDIT: typos and links.

Have you considered the hAP ac2: Support for Mikrotik Hap AC2

The original hAP ac definitely has cpu limitations, this review convinced me to never purchase one: https://blog.quindorian.org/2016/02/review-of-mikrotik-hap-ac-and-great.html/

Thanks for you suggestion. Honestly I didn't look much at those cheap Mikrotik products because from the picture looked like an El Cheapo expensive product (asus, tp-link, and so on), and was missing the PoE-out feature. Silly, I know; but I try to stay away from commercial products if I can afford it.

Currently, in order to repurpose a quad-threaded atom N2600 board, I've found a second-hand rackmount case on ebay (50$); I have both an old gigabit pci card (+2 embedded on the mobo) and an N-radio pulled out from my WL500gPs (I use the radio for wan backup only, the RBM33Gs are doing the wireless distribution job). I think I'll go down that route because that will be my 'edge' router; I don't have a gigabit wan but my town could be enabled any time in the next months; and occasionally I move some files around. From my tests I get double openssl performance of an equally quad-threaded MT7641. A 20$ gigabit switch, and a 5$ PoE injector will do the rest; and I've recycled a lot of old trash. As a bonus I can keep low the number of targets to build: x86 and rampis. For 75$. And in the year 2035 we'll talk again.

I've spent 3 days digging a ton of SBCs but none of those mips/arm cpus can match a 5-10w Atom/Celeron. Even the 64bit Cortex have lower performance than many recent days x86_64.
All those recent x86_64 have AES_NI plug-and-play; on those mips/arms instead you must be lucky, or hands-on yourself to rework some old code miraculously popped out from some shit hole and ... having comments in russian or chinese... then, 6 months later, pray for upstream inclusion, so you don't have to upkeep a custom toolchain yourself.
I mean: a Banana PI R64 is cool and I'd spend about the same money. It's the perfect board for that spot in my net; but given the price, the time to get it here, the limited support by openwrt (have to build it myself and the install procedures are a bit more complex than others because of some weirdness in the way Sinovoip packed the default flash layout) ... why to bother?
The only reason for me to use one of those 'exotic' cpus is ... well ... the hope they lack hw bugs afflicting x86s (spectre&co): using one of those bugged-by-design cpus as a security tool makes my skin run away from my body... but that's all we have on the market so... I must staple my skin in place and drink daily to forget about it. An exotic cpu helps my illusion of safety.
Is it enough to buy an underachieving board? Don't think so.

Consider the overall figure too: all those modular boards make sense if the chance to refit/upgrade persist in time. In the real world this isn't true (any more). In our case you don't get mini-pci-e AX radios; it looks like they are moving to M.2 slot because x4 pcie lanes are needed to support more than a couple gigabit. You'll find crappy radios only on the pcie bus (ex: Intel AX200). And the same happens in the desktop/server world: every generation a new cpu socket, every generation a new memory speed (and every couple gen, a new memory standard), every 2-3 generations a new system bus (pcie2-3-4-5-6), or peripherial interconnect (USB2/3call-me-gen1/3-call-me-gen2/4), connectors, or disk format...

Market-wise this fragmentation/competition is good. Technology-wise this standard-less development is a global scale disaster. 20 years ago I bought my first macbook because I wanted to drop x86s (and SPARCs were too expensive); In this situation, gluing our guts to Daddy-Intel is ... damage control.

I apologize for the rant but I got a bit frustrated from spending so much time to catch up the whole picture for this upgrade.

No need to apologize, quite thought provoking, thank you