Very cool! But how do you turn it on? It seems that the power on and reset buttons are quite buried and difficult to press.
Rock 5 - Case + Cooling
Yeah been saying since dev boards got sent out remove the onboard switches and make like a normal mobo with jumpers.
That way you can use the jumper as a switch or have wired switches like a normal PC case.
what does the fan is doing here? isn’t it supposed to be passive
Does this case allow the use of an m2 ssd on the bottom of the board?
Since it is the Radxa Oficial I would say yes.
PS you may of seen these ones, but if you wondered if they would install with the fins width ways or length ways, then yes they do.
Just been testing it actually and seems to do a lot better than the Blue allnet passive but I think that is due to having no thermal pads and just thermal grease.
There is a problem where the cpu is offset slightly and I think it just lies flatter without a thermal pad.
Nice Looks a bit big!
Yeah you can get smaller if you can find one where the lugs allow you to position any direction.
Its the thermal pads on the blue one from allnet that prob don’t like the cpu offset, as its about same size but seems much better.
I can add my side blowing fan now as it was the only one that would have the vanes going width ways and super cool all above and below pcb.
Really don’t need a fan with this but I will prob be loading with some ML work using cpu/gpu/npu with a nvme so will get one even if just to try out.
I have one actually a 5v 20mm just have to make some sort of bracket to attach to the hex pillars and haven’t got round to that.
is that powerful enough? Thought guys doing that were using huge Nvidia RTX GPUs and using tools like CUDA.
Running ML we are at a very good level where the 6 tops NPU is supposedly more than Google runs on its Tensor NPU in its pixel phones (5 tops but very low energy). Also the Mali is comparative with the cpu in terms of running ML and a model could be partitioned over all 3.
For training the 160,000 hours of audio OpenAi’s Whisper was trained on, you want something more than just a Rtx GPU. If you can convert Whisper to int8 though you could have pretty much SOTA local speech recognition and translation running on a 6 tops NPU.
There is a great repo that has optimised and converted whisper to run on CPU and the RK3588 doesn’t do a bad job and the NPU prob could jump up a model size as prob x3 ML perf of the CPU.
Awesome repo, very easy install.
PS the NPU seems to only use 1.5watt whilst CPU can be 5+ Watts running ML x3 slower.
seems cool, didn’t know speech recognition was now available without sending data to remote server or using closed service like google.
Give the above a try I think you will be pretty amazed as the previous wav2letter is low load and awful Whisper works on a transformer model and the results are amazing they are both quite interesting to compare as night and day in terms of accuracy.
Like everything somehow it ends up being taken over in English you can speak multiple languages into whisper and the English transcription comes out the other end.
Some languages due to less data are much worse than others, then gender, age accent the rarer and less data it gets but it seems god like for what I have tried with the wav files I have fed and also some basic streaming.
Yeah local private no Google or Alexa that is the idea. There are some current Pi based OPen Source that really do a pretty bad job and try to brand themselves by hovering and refactoring and branding MIT licenced software / older closed software past sell by made free.
But I expect some sort of Linux framework will soon be the norm that is an open process where you just mix and match skill servers to core modules of ASR/TTS/NLP
thinking of what diy projects this could be used for . so everything is open source, no alexa, google and so on?
Also less consumer junk as you only need a single ASR than a collection of smart speakers and use wireless audio systems that are part of your media system.
The whole sphere of security, home automation and voice control could work really well on low energy arm boards such as AI accelerated RK3588.
You jump up to 6 tops and you really do have something that is low energy but extremely competent that Google showcase in there newer phones so its quite active around this sort of level and even less.
I’ve designed a case of my own and I figured I’d share it for others to use.
pretty cool, might I suggest using printables.com. Thingiverse has had some issues in the past… with security and peoples accounts.
Sorry for later reply~~
This fan is just the same cpu fan from radax official, and it can get power from cpu-fan pins.
Because the official metal case itself has a metel “fat leg” from top of case down to the CPU of rock5b, there is no space for a cpu fan.
I only use the fan for lower the temperature of the metal case ( so as the CPU and the nvme ). It works very well.
I recently purchased 3 of these cooling fans for 3 Rock5B’s, and hooked them up to the fan connection, turned on the Rock5B, and the fans don’t spin. Are there any setup instructions for getting the fan to work? This would be most greatly appreciated!
The fan is controlled by the OS. Not sure how Radxa is doing that in their distro.