The 2nd ethernet
ROCK 3B ideas and discussion
Yeah I never noticed the Telegram group discussion but only returned after seeing the cnxsoft review of the Radxa Zero.
I think we should just call them the pcie 2 slot and pcie3 slot, 2 on top / 3 on bottom for me.
Look PCIe 3.0 is double the performance of PCIe 2.0 and the JMB585 is actually a PCIe3 2 lane controller that likely will work exactly the same with 1x PCIe 3.0 as 2x PCIe 2.0.
That is already 800MBs as it was on the Rock Pi4 and cpu wise we are dropping down what is not much more than Pi3.
That CPU will never be able to utilise the bandwidth its capable of and it has no way to route that bandwidth to anywhere.
Until actual use and someone does a benchmark we shall not know but if you have a sample and sata controller it would be great to give it a test.
The top slot is an E key as likely A+E key wifi/bt modules can fit there but there are 2x USB 3.0 that can equally be used.
I guess the bottom 3.0 single lane being B-key can provide for NVME or B key 4/5 modems up to its 1GBs rating of a single lane.
Like a say about a wheely bin users are not going to feel any additional performance from dropping the extra 1gbe Nic because a quad A55 will not cope with that sort of bandwidth.
A 4K video is only 50Mbps and the current 800MBs is actually huge for its CPU.
It seems because you where in the Telegram conversation you seem to be struggling with the RK3568 layout so I will post a diagram for you.
Its from https://www.cnx-software.com/2020/12/01/rockchip-rk3568-processor-to-power-edge-computing-and-nvr-applications/ that provides more fuller info.
You can not connect a 5Gbe nic themselves as they only exist as USB3.0 currently and only give approx 3.5Gbs and are expensive.
The RK3568 has a QSGMII 5GBs MAC but no embedded 5GBs Phy and a model with a Phy chip onboard will be the only way to supply it.
Also it will be the only way anything will be able to route the bandwidth of 2x PCIE 3.0 lanes to anywhere useful and even with full blown 5Gbs its still only partial but at least it does something with that bandwidth as its so totally pointless with the CPU available its the bandwidth of anything but its USB 3.0 connector in device mode.
I am saying scrap the Rock3A format as at a very similar price it gives nothing that the Rockpi4 already gives.
Provide a Rock3B as above but make the Rock3A like I said and add a 5GBs Phy onto the QSGMII otherwise the bandwidth of the 2x PCIe 3.0 lanes is totally pointless as the CPU will not cope and there is no interface apart from using as a USB dongle that comes close to being able to consume that level of I/O.
With the 5Gbe Phy then from 5G router to a much faster Nas/NVR it does give possibility to use a large portion of that bandwidth of the 2x PCIe3
JMB585 ainât the only SATA controller around. E.g. RockPi4 was initially suggested to be paired with m.2 2242 b-key to 4x SATA Marvell 88SE9235 boards utilizing 2x PCIe 2.0 lanes, which are very easy to source. They are more compact, very reliable and donât run as hot as JMB585.
Kobol is making new revision of their Helios64 NAS with RK3568 instead of RK3399. Weâll need to see benchmarks, of course, but I doubt theyâd go this way if RK3568 performance was as poor as you describe.
It gives two separate PCIe slots, which RockPi4 never did, and allows much lower z-height installation due to not needing a massive heatsink. Itâs performance is lower indeed, but weâll need to see extensive benchmarks or better test ourselves to see for what it is good, and for what it is unacceptable, before jumping to conclusions.
Frankly I would expect a board with RK3588 in 3B form-factor, instead of RK3568. Rock 3A form-factor fits limited RK3568 IO much better, I believe. Just an opinion, though.
It would seem because of choice Radxa prob because its part of there NAS sata offering the JMB585 is going to be choice for the Rock3.
If you choice something else that is PCIE 2 and you only get have bandwidth that is your problem.
The Helios64 NAS had both GbE and 2.5GbE ports because it would be pointless to have that sata bandwidth without being able to get it offboard.
At a guess I expect they will do exactly what I have suggested and put a $20-30 Phy and couple it to the QSGMII 5GBs MAC and that is why they are looking to at the RK3568
You are validating my suggestions with your own objections.
We can stop this pointless arguments because we can wait and see what Kobol do and we can wait for benchmarks.
Thank you.
It had a (poorly implemented) 2.5GBe hooked over USB3.0, and internal USB3 hub, which ainât a good engineering decision for a NAS. One can plug a USB3.0 dongle into any Rock board and achieve the same result.
Our arguments will definitely stop being pointless once you learn about heterogenous multi-processing and how to use a search engine.
How do the OCD types always zone in on me!?
Thank you for your wonderful insights, but please stop.
PS is this cevap or whatever his name was with a new account?
@stuartiannaylor Iâd like to add that youâre full of it⌠on this bright note - have a wonderful evening, weâre done here.
It is you isnât it Cevap with a new account, I clicked before I edited the last msg.
Yes have a good evening mate.
But with the Mali 52 GPU might we assume that the 3B will have much better graphics and video playback capabilities and speed than does either the Raspberry Pi 4 or the recent Rock Pi 4B+? And if so, will this be evident from the onset, or will it require OSâs and/or software in general that works in conjunction with the 3Bâs GPU?
Yeah it does as the graphics is better but the platform of 4x A55 is quite a bit less than the Rockpi4 or RPi4 which make borderline desktop experiences irrespective of claims.
The Rock5 RK3588 would of likely filled that role but the RK3568/RK3566 are not general application SoCs and trying to fit them into a ill fitting shoe will just return that.
The RK3566 is an extremely well priced edge device that has a NPU and GPU that can act as a co-processor with Arm-NN it also has a PCIe 2.1 x1 where it could provide an extra AI co-processor, 1Gbe NicâŚ
The RK3568 is an amazingly priced gateway SoC but packaged as a device rather than a gateway you just have to compare it price wise with what it can do even with products in the Radxa range.
I am only saying this as I think this is a huge opportunity lost probably by wearing blinkers to general application SoCâs when Rockchip serve up 2x amazing edge/gateway socâs that are symbiotic in use.
Its a shame that is all and still have the Radxa Zero of interest as @$35 + shipping vs the Pi4-2GB I know which one I will be getting even with the horrendous VC4 presented as VC6 graphics and due to the weird nature of booting via GPU the PCIe only supports 32bit and CM4 Coral accelerators donât work.
I would still get a Pi4-2GB or CM4 and this is what confuses me as if you are going to directly pitch the RK3568 in the same place they have you beat already from price to performance and 3rd party market product.
Edge devices sales and development has now over taken the cloud and you get this perfect low cost pairing, scrap one and butcher the other by making it a general purpose application SoC pitched directly at somewhere between a Pi3 & Pi4 that have huge advantages of economies of sale.
Sales and community activity will be an indicator if its too late realise what amazing edge & gateway devices those could be.
I am really interested in Edge devices for a number of applications and maybe can see an extremely lucrative market that Radxa seems to of missed.
I have only mentioned for feedback and will end my discussion but what could of been extremely interesting is actually no interest at all, hence why I will stop.
Mentioned with a hope maybe it could be rectified 
The Khadas VIM3 Pro makes for a relatively tolerable overall desktop experience, and my hope is that the RK3588 will be superior to the VIM3âs Amlogic A311D CPU in all areas (including price).
Actually it doesnât, GPU on RK3566/RK3568 uses Mali G52 1-core 2EE(1 Shader core and 2 Execution Engines) capable of 38.4GFLOPS and 1600Mpix/s fill rate when 800mhz clock frequency. In comparison Mali T860MP4 GPU found on RK3399(RockPi4 A/B/C/B+) is capable of 100GFLOPS. FYI, Itâs not the same as Mali G52MP6 GPU found on amlogic S922X/A311D SoCâs.
Mesa/Collabora agreed to call RK3566/RK3568 GPU as âG52Lâ to distinguish from other G52 variants, where âLâ stands for Lite.
2 Likes
Yeah its the RK3588 we need and damn those silicon queues!
Doesnât matter what GPU it has when its CPU architecture to back the GPU that is needed as in the RK3588
@Larry with edge devices the esp32-s3 hopefully will become really interesting and then currently as a central device my best option is some of the embedded X86 platforms as not all NN frameworks run complete on Arm.
None cloud, big data AI can work really well but the central home âBrainâ still needs considerable power but as opposed to edge devices its cost is singular.
I have a 2nd user i5 4th gen NuC for backup desktop and my main desktop is a I7 6700 so its all about opinion but for me my previous RockPi4, current RPi 400 and prob that Vim pro would not make the grade of what I call a usable desktop.
Its not much of a bother as where I have been looking is MQTT edge devices and gateways, wireless multimedia, NAS & NVR and feel that the RK3566/RK3568 could of been perfect but are being forced to be something that they are not. Very much not in the case of the RK3566. The NPU & Mali are great AI co-processors that probably will make some really great devices that maybe could of been better and more cost effective.
I am just going to play with the Radxa Zero instead that feel is a better fit at its price point to what I am going to play with.
Might make a great network KWS & wireless audio receiver where the Pi0 really struggled, in fact it beats the Pi3A+ which is my fave builders device.
Desktop when the RK3588 does arrive I will prob give one a try if its a bit more cost effective than that Vim3 pro.
All of this now has me questioning as to just what nitch (if any) the RK3566/3568 CPU complete with itâs effectively (or figuratively) castrated Mali G52 GPU is trying to fill?
If you had the RK3566 then it makes a lot of sense as a low cost IoT AI device with NPU and GPU co-processor as the ARM NN libs allow you to use a Mali in similar way you use CUDA.
You can partition and delegate certain layers to NPU/GPU and lighten the load on CPU.
The higher cost gateway/co-ordinator RK3568 makes less sense on its own than say as a high bandwidth MQTT broker or the extremely strong parallels as a NAS for its NVR connected devices.
They where never âdesktopâ or âgamingâ socs even if they do have extremely strong video functionality, the RK3566/RK3568 where this sort of symbiotic pairing for IoT that seems to make less sense as a singular SBC.
1 Like
what do we lose if we implement 2x PCIe 3.0 on bottom m.2 slot, PCIe 2.1+USB+SIM on top m.2 slot?
After reviewing all your discussion and targeting 3B as a general purpose SBC, here is our final decision:
- PCIe 3.0 x2 on the bottom for m.2 M key
- M.2 B key for PCIe 2.1(muxed with SATA), USIM and USB, support both 4G and 5G cards, as well as 2242 SATA SSD(need to test)
- SDIO WiFi/BT will be on board, WiFi 6/BT 5
- remove the SATA 7P connector, since we are a compact SBC, the 2.5inch SSD is not good with enclosures, if you want to use 2.5 inch SSD, we can provide a M.2 B key to SATA breakout board
I hope everybody would be happy.
If you have other ideas, we can implement with the CM3+, the planned carrier board for CM3+ is Radxa E25:
- 2x 2.5G 8125B, one via pcie3 x1, one via pcie2 x1
- MINI PCIe for AP WiFi cards, via pcie3 x1
This can make E25 a small router/CPE device.
6 Likes
These changes look fantastic for a multi-purpose SBC!
I definitely see a use case here even with single radio, however have to ask - would it be possible to put both 2.5G on a single pcie3 lane, like on this dual RTL8125 card, to have additional minipcie with pcie2 link for second radio? Also, will there be additional GBe port(s) on this (e.g. 1GBE for WAN, 2x 2.5GBE for LAN)?
Yes, there is.
If Radxa may release some RK3588 networking SBC or carrier board for RK3588, with 2x or 3x slots for wireless radios - then E25 is perfectly fine with single radio.
One more thing worth investigating would be SDIO radios, like you did on Rock 3B. They all seem to have quite limited signal strength, but if price of adding one is not very high, or it is possible to add a e-key m.2 slot wired to SDIO without increasing price or complexity of the board much, such radio may be acceptable for 2.4GHz band if user wants to have 2400Mbps 5GHz or 6GHz-only radio in miniPCIe slot. But again, if there will be upcoming RK3588 SBC / SoM + carrier board with 2x-3x minipcie slots for radios, this is kinda irrelevant - those who are happy with single radio will get the E25, others will step up to RK3588 product.
May i confirm will the b key be routed with USB2 or USB3?