When JetPack 3.0 is finally installed, even though the extremely skimpy nvidia documentation said nothing about inserting an SD card, I did so prior to starting the install.
So, when the boot and etc. are created during the install, is it put on the onboard EMMC (NVRAM) or does it all go to the SD card for future booting and running?
Or do we simply use the SD card as a disk to be mounted and used just like an SATA data-only drive?
/rant on/
For the $600 !!! I expected and want a really thick book that tells me all about the the TX2, it’s layout, capabilities, software ins and outs, and so on — instead of having to ask other people and DRAGGING dribs and drabs of info out of this web site and that web site, with no coherency at all. NVidia, you really dropped the ball there. If this project weren’t so important to me, I’d have returned the TX2 already. And what came in the box was barely 2 napkin-sized pages and printed in black on dark brown! Whose work of art was THAT??
Ok /rant off/ Btw, I have 7 or 8 other types of embedded boards in my harem, have been using QNX/UNIX/Linux since 1980, and studied AI theory since the mid 80’s. Even took a couple of stabs at SNNS, if anyone remembers that great set of tools. Started programming professionally in 1977. Still lots to learn, ain’t it great!
You don’t need an SD card, the device has 32GB EMMC storage and that’s what you flash to. That’s the same way many laptops work, Chromebooks especially. I’m using a Chromebook running GalliumOS with EMMC only to post this. EMMC is slower and cheaper than an SSD. I like EMMC over SD cards because when I’ve used SD cards for OSes, I tended to end up with corruption issues if the device lost power in the middle of something. I haven’t had the same issues with EMMC but time will tell.
I don’t think the documentation is much of an issue with the TX2, myself. The Linux version is standard and it isn’t hard to do a kernel rebuild. I’d like it if they added more drivers to the stock kernel but it wasn’t that difficult to add them myself. Given the volume, it is going to have some sharp edges. So far I’ve been pretty happy with it and its running servos, cameras, etc. The video processing speed given the low wattage is pretty nice.
FYI, having an SD card present on the Jetson during flash will not change what gets flashed. Changing the flash command itself would be the only thing causing this. The mmcblk0p1 target is the eMMC; in a case where SD card is flashed the target is mmcblk1p1.
Thank you esjs and linuxdev. That is extremely important information to know in this early stage. Other SOC’s can boot from either emmc or SD card or a combination of both. And one has to know what’s going on so as not to do the wrong thing. :)
If you have a serial console you can actually interrupt boot within U-Boot and explore some of the environment (just type “help” to see commands in U-Boot). Several of the U-Boot environment variables are executed as macros and are what determine boot order and boot device search. You can edit those variables to test new things and then boot with the edits being only temporary, or you can save the edits first for permanent change (flash would still overwrite such changes).
I assume it gets hit heavily when running AI tasks.
Why? A model is typically no more than a few hundred megabytes in size, and it’s read-only; read it from disk and put it on the GPU.
If you store output information from inference, you will probably start writing a little bit, but typically the output write rate would be low.
Recording video (even encoded) is likely to be higher data rate than any AI/inference work you might do.