Radxa 6 under development

99% of users - living just fine with 24/7 systems with kernel that was release 2 decades ago and patched like there is no tomorrow (btw almost all enterprise systems, like HPe 3par, Cisco UCS, etc use it)

Willy - I DON’T BELIEVE, crappy kernel will ruin my data. Not hardware, not crappy software running on top on linux, not users, but kernel!

Like come on, at least made a non-laughtning argument why you bought x86 instead of using rock 5b. Since the only difference is to have ability to use windows and some proprietary blobs, created only for x86 anyway

I got FS corruption in the past that was caused by heavily patched BSP kernels. When you see vendors play with memory management, with barrier around DMA operations, with the block layer and so on, you should really be scared about what your data will become. Very often BSP kernels are made for a few specific applications and are just not expected to be used outside of these, so vendors don’t even consider those and have no clue whether their patches will break perfectly valid use cases. I’m sorry but no, no more of this crap.

Yeah sure. Thankfully all my 5 rock pis with “bsp kernels” are still fine even after 3 years of working including power loss, fs corruption (bless fsck.ext4) caused my losing 6/10 disks in raid5 (power issue) and reapplying re-compiled BSP kernel for Netgen nic support, so basically even more *outside of user case*

But sure, BSP kernel caused your unrecoverable FS corruption.

But if you are trying to use something like zfs/xfs/btrfs - good luck. There is even vanilla kernel bugs, that render them useless with data corruption

I don’t think Willy is alone and I am not going to turn this into a fan boy argument against someones opinion.
Willy seems to have the opinion he doesn’t trust the BSP kernel based boards and that is just his opinion.
I am the same as the pressure to get working kernels with limited resources and times often means resorting to quick fix hacks.
Many of us accept early adoption can often stretch into years and often these hacks are the only solution available.
Its not so bad on small edge boards of $10-30 as you may find specific working solutions or simply shelve.
When you start getting past $100 and keeping with the thread and providing feedback to Radxa about how we feel about higher cost more powerful boards, the BSP provided by small companies with limited resources starts to become unattractive when there are alternatives.
When you get standard kernels running big name hardware that is developed, tested and supported in such numbers it gives a level of assurance that is the gold standard for running servers and anything less is just a hobby.
Enjoy your hobby but also respect that others may deem it so and be more critical of employing devices where the level of assurance is practically non existence apart from the bias of a few observers who have.

I am a hobbyist, but there gets a level such as Intel N100 based boards and above such as Mac Mini’s where its extremely hard to justify levels of spend on something running a BSP on a unique board by a small vendor such as Radxa.
That is just honest opinion and have no interest in arguing that as its not going to change.

I struggle with the price point of the Rock5b to be honest and anything above that needs a very unique and strong selling point than just a bit bigger and better.
Its not that I will not spend its just likely I would source alternatives.

1 Like

Mainline kernel has been running fine for a few months with no issue on my Rock5 cluster.

1 Like

That’s a great feedback, thanks! From time to time I watch the status of mainline and I thought it would be about time to give it a try. You’re just encouraging me into that direction :wink:

@hipboi without disclosing the SoC nor vendor, I’m wondeing about the price point of the SoC alone. When you compare intel’s N97 and N100, the former just has a 6% higher turbo, 33% more GPU cores, twice the TDP, embedded options available and the rest is the same. The chip sells for $128 vs $55, so they’re clearly segmenting their market. Then if you compare N100-equipped mini PCs to N97 ones, you’ll see that the N100 ones start around $130 while there’s nothing below $230 for the N97.

It doesn’t mean the N97 doesn’t sell, I think it does, and it means that buyers are valuing the difference between the two to pay $100 more. Thus I think that if the SoC of the ROCK6 is very expensive, it’s important that the board provides features (and/or performance levels) that will make buyers compare it favorably to an N97 or above rather than an N100. I know it’s not easy, as N97 and an ARM devices are completely different beasts. That’s also why I think it must really boot on a regular distro out of the box if it’s going to be priced like a powerful small PC.

And if it remains expensive due to the SoC, trimming the board to its minimum could definitely help, including thinking about going the carrier board way from the start maybe, so that the expensive part (RAM+SoC) exists in stock with very few combinations and users decide on the baseboard depending on their choice of eth/pci/usb/hdmi/m2/sata that best matches their use case (probably just 2-3 boards). It also allows to perform incremental updates to certain parts (e.g. power supply, PCIe splitting etc) without having to wait to sell a large pending stock.

Didn’t realise Radxa had already branched into Intel/AMD devices under their Palmshell brand. There is the AMD R1505G NeXT H2 micro server. So the X2L came from the SLiM X2L.

New board is the Radxa NIO 12L - MediaTek MT8395(Genio 1200) ?

Its not going to be a major leap over the RK3588 but the AI unit via (are they using https://github.com/aws-neuron/aws-neuron-sdk) should be interest
You have A78 vs A76 which should be interesting and G57 MC5 MP 5 vs G610 MC4 and again interesting…

No, since it doesn’t have any pcie

@Dante4 That is the Radxa link https://docs.radxa.com/en/nio/nio12l/getting-started

Yes, I can click on the link in the @mtx512rk post.

And there is no pcie

Ok no to buying one, now I get you. What will be interesting is what Mediatek and Radxa provide as a BSP.
The RK3588 has benefitted from a few iniatives sponsored by mediatek for open source drivers that have been common.
I have never owned a Mediatek board and for all intensive purposes its a bigger more performant RK3588(s) board, that for me is still likely to fall short with LLMs I would want to run.
Likely some are already looking at the documentation provided that looks at 1st glance quite comprehensive.

https://mediatek.gitlab.io/aiot/doc/aiot-dev-guide/master/hw/mt8395-soc.html

I guess you misunderstood what I said and what I answered.

@mtx512rk asked is new board (which is mentioned in first answer to this thread) is the NIO 12L, to which I answered that “no, it’s not, since it doesn’t have pcie”

Its looking very like it is as the quad A78 had no mention of PCIe.
The later and bigger Rock6 supposedly has Pcie4

Did some simple comparisons for MT8395 and RK3588 in the aspects that interest me.

MT8395 (TSMC 6nm) RK3588 (Samsung 8nm)
CPU 4x A78 2.2Ghz + 4x A55 2.0Ghz 4x A76 2.4Ghz + 4x A55 1.8Ghz
GPU 1st Gen Valhall Mali-G57 MC5 3rd Gen Valhall Mali-G610 MC4
VPU 4k90 decode + 4k60 encode 8k60/4k240 decode + 8k30/4k120 encode
NPU/APU Cadence Tensilica (Dual-core, 5 TOPS) RKNPU (Triple-core, 6 TOPS)
MEM Up to 16GB LPDDR4X 4266 Up to 32GB LPDDR4X 4266 / LPDDR5 5500
FLASH UFS 2.1 + eMMC 5.1 + SPI NVMe (PCIe 3.0x4) + eMMC 5.1 + SPI
HDMI TX 1x HDMI 2.0b (4k60) 2x HDMI 2.1 (8k60 + 4k60)
HDMI RX 1x HDMI 2.0b (4k60) capture 1x HDMI 2.0b (4k60) capture
DP TX 1x DP 1.4 (4k60) via USB-C 1x DP 1.4 (4k60) via USB-C
WLAN Onboard MT7921 chip (802.11AX 80Mhz) Most common cards via PCIe 2.0x1
LAN 1x or 2x 1.0GbE (It depends) 1x or 2x 2.5GbE (It depends)
USB ?x 3.0 + ?x 2.0 (It depends) ?x 3.0 + ?x 2.0 (It depends)
4 Likes

Did you find something out about it’s PCI-E lanes? I saw that the mtk devkit also had an M.2 slot but couldn’t find anything on the lanes connected to it.

Still not enough for PS2 emulation then, probably. Why are the GPUs so limited in SBCs?

I think the board has exhausted PCIe lanes due to the WLAN MT7921 and USB3.0 bridge.



https://www.mouser.com/datasheet/2/1074/MT8395_IoT_Application_Processor_Datasheet_v1_4-3237050.pdf

1 Like

Great comparison, I’ve done same on my head also checking how much UFS in this version is performing (and it’s about 2x faster than eMMC 5.1 and about 1x pcie 3.0). NIO2 seems to be interesting board but has several downgrades compared to Rock5, I’m mainly disappointed by 1G ethernet.

I also noticed something strange in docs:

Is this same thing as we see on pi5 with PD only with 5V modes? I hoped that nobody will follow them :confused:

1 Like