$95 Banana Pi BPI-R2 Pro 5-port Gigabit Ethernet router board is powered by Rockchip RK3568 SoC

Banana Pi BPI-R2 Pro is an update to the Banana Pi BPI R2 router board that replaces MediaTek MT7623A quad-core ARM Cortex-A7 processor with a much more powerful Rockchip RK3568 quad-core Cortex-A55 processor.

The Banana Pi BPI-R2 Pro board looks very similar to the first generation R2 board with the same dimensions, 2GB RAM, HDMI & DSI display interfaces, five Gigabit Ethernet ports, one SATA port, two USB 3.0 ports, and one mPCIe socket, but it also adds one M.2 socket, support for MIPI CSI cameras, and extra storage with a 16GB eMMC flash.

Banana Pi BPI-R2 Pro specifications:

  • SoC – Rockchip RK3568 quad-core Arm Cortex-A55 processor @ 2.0 GHz with Mali-G52 GPU, 0.8 TOPS NPU
  • System Memory – 2GB LPDDR4 (option for 4GB)
  • Storage – Onboard 16GB eMMC Flash (option for 32 and 64GB), MicroSD card slot, 1x SATA III port, SPI flash for network boot
  • Networking
    • 5x Gigabit Ethernet RJ45 ports (1x WAN, 4x LAN) using Mediatek MT7531BE switch (changed from RTL8367RB-VB-CG in thw August 2021 announcement)
    • Optional 5G, 4G LTE, WiFi, and Bluetooth via M.2 and/or mPCIe card + micro SIM card socket
  • Display interfaces – 1x HDMI port, 1x MIPI DSI connector, 1x MIPI DSI/LVDS connector (selectable by software), 1x eDP connector
  • Camera – 1x MIPI CSI camera connector
  • Video
    • Decode – H.265, H.264 up to 4Kp60fps
    • Encode – H.265, H.264up to 1080p100
  • Audio – I2S header, speaker header, 3.5mm headphone jack, built-in microphone
  • USB – 2x USB 3.0 ports, 1x micro USB port used for debugging via CH340 (but other parts of the specs shows it as a USB 2.0 OTG port)
  • Expansion
    • 1x mini PCIe socket (PCIe + USB)
    • 1x M.2 Key-E socket (PCIe + USB)
    • 40-pin GPIO header with 28x GPIOs, UART, I2C, SPI or PWM, and power signals (+5V, +3.3V, GND).
  • Debugging – 4-pin UART console, micro USB Debug UART (TBC)
  • Misc – IR Receiver; Reset, power, and u-boot buttons; LEDs; fan header
  • Power Supply  – 12V/2A via DC power jack, 2-pin header footprint, or  5V via micro USB port
  • Dimensions – 148 x 100.5mm
  • Weight – 100 grams

The original BPI-R2 router board had a built-in MT6625L wireless module with WiFi and Bluetooth that is missing with BPI-R2 Pro. But that may not be a bad thing, as I had read people had issues with the driver, and the expansion slot provides more flexibility to the users for their preferred wireless module(s) if needed.

Banana Pi BPI-R2 Pro supports Ubuntu 21.04 (Not supported by Canonical anymore), OpenWrt 21.02, and Debian 10 all with a Linux 4.19 kernel. You can check the BSP on Github, or find the OS images and more details in the Wiki. Rockchip RK3568 is much faster than the MediaTek MT7623A found in the first BPI-R2 board, but lacks network acceleration, so we’d have to see how both platforms compare under various networking workloads. The multimedia capabilities will be much better on the Rockchip board with 4K support, and up to three displays. We’ve also seen initial RK3568 support added to the recent Linux 5.14 release, which may be a plus.

The new Banana Pi BPI-R2 Pro sells for $95 plus shipping with 2GB LPDDR4 memory and 16GB eMMC storage, which compared to the $91 that was asked for Banana Pi BPI-R2 board last, year, but the company has now removed it from its Aliexpress store.

[Update: This post was initially published on August 30, 2021, as the board was unveiled, and updated once it became available on Aliexpress]

Share this:

Support CNX Software! Donate via PayPal or cryptocurrencies, become a Patron on Patreon, or buy review samples

37 Replies to “$95 Banana Pi BPI-R2 Pro 5-port Gigabit Ethernet router board is powered by Rockchip RK3568 SoC”

    1. Wouldn’t the faster cores and features like Crypto extensions mitigate the lack of network accelerators? I’ve also read some accelerators are not supported in OpenWrt, but I can’t remember for which chip/board that is.

      1. No, that’s unfortunately not how routers work.
        It’ll most likely be fine for 802.11n type networks, but once you move to 802.11ac and get a fast internet connection, this thing is simply not going to be able to keep up.
        The SoC cores would be decent enough for something like a NAS, but this SoC lacks key router features like hardware NAT and Wi-Fi offloading.
        The Wi-Fi offloading depends on what Wi-Fi modules you use though, as some from Broadcom and I think some more recent ones from MTK has integrated processing that will offload the main SoC, but this only applies to the last generation or two.
        Even high-end router SoCs from Qualcomm features packet processors that handles things like switching, routing, traffic management, QoS etc. plus a dedicated NPU for the Wi-Fi and even a separate security engine that can be used with more advanced firewall software.
        Slightly faster processors can’t make up for things like that.
        Obviously this board is limited to Gigabit speed, so it might just be able to handle that, if Banana Pi makes sure they optimise their drivers, since this seems to be needed on all of these types of chips.
        Personally I would not build a router around this chip.

        1. > Banana Pi makes sure they optimise their drivers

          Are you kidding? 🙂

          They do nothing relevant other than copy&paste from somewhere. The predecessor R2 sells since 2017 and hardware NAT barely works there just since recently. And if not some very dedicated individuals (especially Frank Wunderlich) would have done all the real driver work the R2 would still be a paperweight.

          1. Yeah, hence the comment, as I’m expecting nothing from them.
            I guess you missed the if…

        2. If adding related hardware, discussion is about admins’ (server or commercial side?) data handling also?
          Studies, statistics and reasoning on that item?
          (Did count/guess clicks on one reputable web page once for having all ‘legitimate interest’ options objected, before being optimized to a ‘one button reject’ to that interest: gave up counting above 900 clicks, manually, for that one web page?
          Compared to that, few pages show what information is visible to sites.)

        3. Well, for me it’s the exact opposite. I am particularly careful to stay away from bogus (and sometimes even insecure) acceleration features. They usually rely on totally crappy drivers and bogus silicon, and you progressively have to disable them one after the other until the point you’re left with the tiny CPU whose size was justified by all the monstrous accelerations available to offload it.

          I can understand the need for acceleration for very specific workloads involving lots of small packets with stateless protocols but usually that’s pointless or even counter-productive.

          You were speaking of hardware NAT, but the real cost of NAT is the table management, typically something you absolutely want to keep in software. Some poor network stacks might force to recompute full checksums but usually IP headers are updated while doing differential checksums so that you don’t need to rescan the whole packet. Some home-grown proprietary stacks might not always do that, though.

          1. Uhm, I don’t think you quite understand. It’s not possible to do the speeds we’re hitting with 802.11ax without the built in accelerators, there’s nothing bogus about it. If that’s how you feel, maybe you should stick to wired networking. You clearly have no understanding of how modern routers work. We’re not talking corporate rack size hardware here if you haven’t noticed, there’s no way these SoCs could do what they’re doing without the various co-processors that are being used.

            1. Who cares about wifi specific cores on something that doesn’t even have wifi ?
            2. Hardware NAT is trash, it’s just a trick for CPUs that are too weak to route at wire speed (like the mt7621a), once you enable it you lost most of the useful functions of a router, like a proper firewall, QoS etc. Yay.
            3. The CPU here is MORE than enough to route several gigabits purely in software.

            This board is trash but not for those reasons.

          2. > It’s not possible to do the speeds we’re hitting with 802.11ax without the built in accelerators,

            But why ? How many PPS do you need to forward at these speeds ? The dual-core cortex A9 in my clearfog happily forwarded 1 Mpps through the whole kernel stack last time I checked, and had no trouble reaching 1.488 Mpps saturating GigE with the smallest packets when done at lower layers.

            Here we’re talking about a CPU that has basically the same per-core performance with twice the number of cores so from a software perspective we can expect the double (and yes, I know that a Rockchip’s internal I/O architecture is far behind a Marvell’s). So unless you provide some elements to figure what’s so special about 802.11ax that should impact the routing function, I’m sorry but I don’t understand.

          3. > A9 … Here we’re talking about a CPU that has basically the same per-core performance

            …at the same clock but the A55 in RK3568 run 20% faster. IIRC you have both VIM3L and ODROID C4 in your lab? Have you done any comparisons between S905X3 (AKA ‘another quad A55 with not that great internal I/O’) and the Armada 38x wrt pps in routing/NAT scenarios?

          4. I have a VIM3L and a StationPC-M2 (they finally RMAed it after the linux-based flasher bricked it). However they both have a single GigE port and it’s a crappy dwmac or something of similar level. It’s not bad per-se, it’s just to a NIC what a needle is to a drill. I don’t even remember if it supports SG, TSO nor interrupt mitigation! At least mvneta can do all the basic stuff correctly, even IPv6 checksums and that helps a lot! I’d like to run some measurements but I’m too short of time these days, will do that in a few weeks at best I guess :-/

            What should be great will be the new Clearfog equipped with the CN9132 and 10GE. That’s basically the same thing as the A7040, but in a clearfog’s awesome enclosure 🙂 I’d probably stop carrying my mcbin everywhere with me if I had one!

          5. > StationPC-M2 … have a single GigE port and it’s a crappy dwmac or something of similar level

            Well, I guess that’s the same thing RK3568 is equipped with twice 😉

          6. Except the Marvell Armada chips have network acceleration co-processors that you clearly were unaware of. It’s a chip that was designed to be used in routers and NAS appliances. This general purpose chip from RK was not.
            Yes, it has more native processing power, but if that was enough, why would Qualcomm and Broadcom go through the trouble to add these features if they weren’t needed?
            This might work ok as a a simple 2×2 router, but I doubt it’ll work well beyond that.
            I’ll happily wait for you to get one and report back though.

          7. > Except the Marvell Armada chips have network acceleration co-processors that you clearly were unaware of.

            I am perfectly aware of them (in fact you only have them in some models, those using the mvpp2 driver can for example use the buffer manager and a few other features), I just don’t use them. The old armada chips also used to support using the DMA engine to forward packets. It was slower than when done in software! It would just save CPU cycles, which was convenient on the Feroceon (single core) if you didn’t need performance!

            No, really, I’m just using the basic trivial stateless stuff that’s documented over only 17 pages in the datasheet and that supports standard scatter-gather, TSO, checksum offloads, packet type classification etc. Nothing outrageous, basically the same stateless stuff that a good old e1000e NIC already supported, except that the neta doesn’t need to pass via PCIe since it’s directly attached to the internal bus, thus it has a very low latency to the descriptors stored in the L3 cache. There’s nothing highly difficult there, really!

            We can make criticisms about Marvell chips (particularly on the scarce documentation) but when it comes to I/O efficiency they just remind everyone that with a properly balanced set of extremely simple features and low-latency busses, you don’t need 2000 CPU cycles to forward a packet between two NICs. And by the way, on the Armada8040 in the mcbin, I have no difficulty pushing 19 Gbps for HTTP traffic through the two 10G ports. And that’s with only 4 poor A72 cores and no copro either.

          8. But regardless of coprocessors and such things, by experience I do *not* expect to get the same network performance from an RK nor AML chip as from a Marvell one. Just because they’re clearly not designed for low-latency communication with their internal high-speed devices, and that their NICs are simply there for connectivity and were not designed to deliver performance. I won’t blame them for this, it’s not their target market.

          9. The 12 years old 32 bits Cortex-A9 doesn’t have the same single core performances than a 64 bits Cortex-A55 that is less than 3 years old.

          10. > The 12 years old 32 bits Cortex-A9 doesn’t have the same single core performances than a 64 bits Cortex-A55 that is less than 3 years old.

            Quite the opposite, they’re precisely identical because A9 was the high-end by then, almost matched by the low-end A53 later, that was slightly improved by the A55. What differs, though, is the power consumption, the memory controller which is way better in the A55, and a large number of extensions found in the A55 (crypto etc).

            But the point above was not about the cores themselves but the surrounding I/O architecture that is specific to the SoC and not to the core.

        4. This is a gigabit device. You really _don’t_ need hardware NAT to hit gigabit internet speeds. Software flow offloading is enough for the the mildly clocked A55 cores to handle that very easily (hell, even overclocked single core MIPS from many routers can handle that). If you’re after alternative firmwares like openwrt, you wouldn’t be relying on these accelarators anyway, since they’re not supported. And if they were, they’d break SQM anyway.

          So yeah, as a gigabit device, it’s just fine. Even with a fast AX card on board, you’ll ultimatively still be limited by gigabit interface, so there will no real bottleneck to speak of. You’d have to do some serious SQM, WiFi and Gigabit routing at the same time to put it on it’s knees.

          Those other real routers have those accelarators for good reason; they usually have pretty weak main core(s), so they _need_ them.

    2. I don’t know. Check out the OpenWRT forum. Raspberry Pi 4 and NanoPi R4S are more or less the best ARM-Devices. AFAIK they do not have any router features. Still very fast. Also for CAKE.

      1. Except the Marvell Armada chips have network acceleration co-processors that you clearly were unaware of. It’s a chip that was designed to be used in routers and NAS appliances. This general purpose chip from RK was not.
        Yes, it has more native processing power, but if that was enough, why would Qualcomm and Broadcom go through the trouble to add these features if they weren’t needed?
        This might work ok as a a simple 2×2 router, but I doubt it’ll work well beyond that.
        I’ll happily wait for you to get one and report back though.

        1. Why would a company add something to a product or not doesn’t mean anything as an argument. Anyway. In the sbc world we have tons of people making routers even with CM4s/RPI4s, devices that not even have extensions. This soc can make affordable routers… how much cost an armada or a “propper” open spec router chip? Tons of money. This sbc cost less that many cm4 routers plus carrier boards. They wont compete with your armada bc they have a complete diff price tag.

    3. Well, that’s how it is with the copy & paste banana sinovoip team. You order whatever and they deliver top notch crap since 1988 (or so)…

  1. > 1x SATA 2.0 port

    RK3568 supports SATA 3.0 (6Gbps). I know it’s copied from their ‘wiki’ and as such the fault of SinoVoip’s copy&paste monkey who still just vomits letters and numbers into their wiki instead of doing proper technical documentation but at least here it should be reported correctly.

  2. > 5x Gigabit Ethernet RJ45 ports (1x WAN, 4x LAN) using Realtek RTL8367RB-VB-CG switch

    If the rendering’s traces are correct it seems like the WAN port is attached to an own PHY using RK3568’s 2nd GMAC so only the LAN ports should be behind the Realtek switch.

    1. It’s also possible they connected it to one of the switch’s RGMII ports. Looking at the switch’s product brief, when in 100 Mbps mode, it only supports half-duplex! That will remind 1995-era duplex trouble to all those who want to connect 100M devices there such as low-end TV boxes.

      1. > It’s also possible they connected it to one of the switch’s RGMII ports.

        Well, if they would be interested in cooperation with open source communities they would’ve long released schematics (since they claimed ‘sample will ready next week’ few days ago). But no, they’re even too stupid to provide the datasheet link in their ‘documentation’ for the Realtek switch (same URL as PMIC datasheet link).

        1. I think that by now their long history of inexistent cooperation leaves no hope for a short-term change, and their products will remain limited to serving as weights to block doors in draughts 🙂

  3. > Power Supply – 5V/3A via DC power jack, 2-pin header, or Micro USB (OTG) port

    That’s copy&paste from their wiki and most probably BS due to the power header next to the SATA port labeled 5V/12V. Maybe the result of SinoVoip letting their copy&paste monkey still do ‘documentation’?

    On the R2 page (the copy&paste source for R2 Pro) still is written ‘5V@2A via DC Power’ while it’s 12V in reality. This refusal to provide any correct information/documenation for 6 whole years now is reason enough to never ever touch this Banana crap again.

  4. Just wondering, i think RK3568 provides USB3 ports, why no design connects them to Minipcie or M.2 B key ( even behind a hub? ) since 5G is around the corner and most of the modems out there provides USB3

    1. The processor supports HDMI CEC based on the datasheet.
      But since Banana Pi has not released the schematics we don’t know if the HDMI CEC signals have been routed properly.

  5. The Wiki has been updated, and Realtek RTL8367RB-VB-CG has been replaced by Mediatek MT7531BE switch

Leave a Reply

Your email address will not be published.

Advertisement
Advertisement