2,010 karma · joined October 20, 2019
My concerns are
- are there open source (Linux) drivers for that GPU available now?
- is this CPU really suited for battery-powered mobile applications? I couldn't find details on the power consumption right away...
An overview of the available devices can be found in this presentation:
https://crc.dev/blog/Container%20Plumbing%202023%20-%20vfkit...
I had a quick glance at the OS in the linked article. This seems to be based on a 32-bit RISC-V with MMU. However, AFAIK, all available RISC-V SoCs with MMU are 64-bit. The 32-bit cores are only used for embedded controllers (unless you want to start designing an FPGA-based system).
The 32 and 64 bit versions of RISC-V are _not_ binary compatible, but the differences are rather small. Porting the MMU code from 64 to 32 bit or the other way round is not very complex, see my RV32 port of xv6 at https://github.com/michaelengel/xv6-rv32 (the regular MIT xv6 version only supports RV64).
The major difference is that virtual address translation on RV32, sv32, uses a two-level page table (10 bit index for the first level, 10 bit index for the second and 12 bit offset) whereas there are several modes of translation for RV64. The most common one, sv39, uses 39 bits of the virtual address split into three 9-bit indexes (so you need a three-level page table for 4 kB pages) plus 12 bit offset.
If you make the modifications, running the OS on real hardware should not be too difficult. The Allwinner D1 is a relatively simply RV64 single code SoC (boards can be found for $20 upwards from aliexpress) and getting the CPU and a UART to work is not that difficult. You can check out my xv6 port to the D1 as a reference: https://github.com/michaelengel/xv6-d1
PC HD floppy drives also have 80 tracks per side, but use twice the data rate (storing 18 sectors of 512 byte each per track) and thus require floppy disks with a different magnetic coating that can handle the higher frequencies. This can also cause problems with data retention and reliability when using 1.44 MB HD disks in DD drives for the ST.
One problem with floppies formatted with early TOS releases (1.0 and 1.02) was that these used a Media Description Byte (offset 0x15) in the boot sector which confuses DOS/Windows system even though the file system format used is FAT12 or FAT16. This was fixed in TOS 1.04.
Details at https://www-user.tu-chemnitz.de/~heha/basteln/PC/usbfloppy/f...
While there's a lot of information about the described satellites online, I was unable to find details on these details.
See also https://en.wiktionary.org/wiki/keming and https://www.ironicsans.com/2011/11/keming_revisited.html
As I mentioned in my comment to that article, the JH7110 is not an ideal platform for a mobile device, since suspend to RAM is not supported.
See section 1.6 of the errata: https://doc-en.rvspace.org/JH7110/PDF/JH7110_Errata.pdf
"This issue will be fixed in StarFive next generation SoC JH-8100"... let's see.
The Musebook, featuring the SpaceMit K1 or M1 SoC, seems to be the better choice, but the device has not been available from aracetech for several months: https://arace.tech/products/muse-book-risc-v-laptop
I have one of the Musebooks here which works fairly well - it's not very fast compared to, e.g. a Raspberry Pi 5 - and the price of ~300€ is similar to the ARM (Rockchip RK3399)-based Pinebook Pro, but with better display and keyboard as well as being offered with up to 16 GB of RAM.
The MilkV Jupiter board uses the same SoC as the Musebook and is relatively inexpensive (and available right now) if you want to experiment with the K1/M1 SoC: https://arace.tech/products/milk-v-jupiter-spacemit-m1-k1-oc...
I picked mine up from the post office yesterday, it's 50% faster in Geekbench single/multi-core CPU benchmarks than my M1 Pro Macbook Pro and about as fast in GPU performance. Impressive.
So do you plan to offer access to QNX source code again in the future?
Additional components also have documentation that's hard to find, especially those based on Synopsys' DesignWare IP. These are sometimes at least... strange (or should I say buggy?) effects such as with their "standard" 16550-compatible UART: https://lupyuen.github.io/articles/plic
The Vivante GPU in Freescale's i.MX8 SoCs was reverse engineered and is probably the closest you can find to a somewhat documented GPU (STM32MP SoCs also have a Vivante GPU). In general, the i.MX8 is well documented - one of the reasons why it is used in Purism's Librem 5 phone.
You can "cheat" and improve the GPU situation a bit by using a RISC-V board with a PCIe slot and an old AMD GPU.
When there's no documentation available, all you can do is reverse engineer the functionality of existing open source drivers. Linux drivers are often quite complex, I recommend looking at Net/OpenBSD and Plan9 drivers instead. For some RISC-V SoCs (SiFive Freedom U740 on the Unmatched board and JH7110), Haiku is also worth looking at.
An alternative which provides documented hardware is to use an FPGA-based SoC builder, e.g. LiteX: https://github.com/enjoy-digital/litex
Of course, this results in a significant degradation of performance...
Another option would be to use a regular x86-based PC and rely on the documented interfaces, such as VESA graphics, but then you have to deal with x86 and all the baggage of almost 45 years of PC evolution...
VW had an attractive and relatively affordable small electric car, the E-Up, but its production was stopped last year.
So, good riddance, German car industry. Let's concentrate on future-proof technologies like... IT, cryptocurrency, NFTs, AI... wait...
Stephen also wrote a great blog article series on building your own RISC-V OS in Rust (https://osblog.stephenmarz.com), so I'm really looking forward to reading his computer architecture book (and probably using it in my course).
For a good overview, I can recommend to check out the ACM doctoral dissertation awards: https://awards.acm.org/doctoral-dissertation