Milk-V Reveals Technical Specs of Milk-V Jupiter RISC-V Mini-ITX PC
linuxgizmos.com
linuxgizmos.com
Not everything runs on risc-v, but plenty does and you can look at Ubuntu's repos to see for yourself.
https://gitlab.freedesktop.org/mesa/mesa/-/tree/main/src/ima...
https://elixir.bootlin.com/linux/latest/source/drivers/gpu/d...
But if you want support Right Now, plug an AMD video card into that PCIe slot.
I think that is exactly the stage in the life-cycle that the ecosystem is currently in. This is for those people to start working. The previous level of getting the Linux kernel working is done, now it's onto the drivers.
After this will be ironing out power management and Wayland issues.
Only after that will it be ready for the end consumer. Thousand miles, first steps sort of thing, but yeah, it's not ready for Linus to drop and then benchmark CS2 on.
This whole process of bootstrapping RISC-V ecosystem did not start yesterday; it's been going for several years now.
By virtue of having 8 cores it's the fastest cheap RISC-V SoC currently available (only the 64 core SG2042 in Milk-V's own $2500 "Pioneer" workstation is faster).
It's also one of only two currently shipping SoCs supporting the latest RVA22+Vector ISA specification. The other one, which shipped in November, is single core with 0.5 GB RAM, so this is far more practical as a PC.
So you can be pretty sure there is going to be a LOT of attention and work on this SoC.
At least until it is overtaken by something faster.
Two generations of faster, but still inexpensive, SoCs are expected this year: 1) SiFive P550-based quad core with about 2x the per-core performance, 1.5x from IPC and 1.5x from MHz, also RVA22 but lacking vector; 2) SiFive P670-based 16 core SG2380 with about 3x the per-core performance, 2x from IPC and similar MHz to P550, full RVA22+Vector. This SoC will leapfrog currently shipping (or known) Arm SBCs such as the Pi 5 and RK3588 boards (Rock 5, Orange Pi 5).
There is also the 64 core SG2044, an SG2042 updated to RVA22+Vector specifications. It will still have about the same per-core performance as this SpacemiT SoC (but a LOT of cores), and I don't expect it to be cheap.
The biggest problem with RISC-V at the moment is that we're seeing -- and will continue to see for some years -- a 2x or more per year improvement in performance of inexpensive mass-produced hardware, so ANYTHING you buy is sure to be quickly eclipsed.
Similar to the 1990s in x86 land, which started with everyone using the 33 MHz 80386 and ended with 1.2 GHz Pentium III and the announcement of Pentium 4 and Amd64/x86_64.
Except current RISC-V hardware (e.g. this machine) is already at late Pentium III performance per core (but 8 cores and 64 bit and better than AVX vector/SIMD), with early Core-i7 performance (but more cores) coming THIS YEAR.
Everyone is quickly converging on the RVA22 instruction set spec (with or without the optional Vector), a big step up from the previous RV64GC (now renamed RVA20) common base ISA. The RVA23 spec (which among other things makes Vector compulsory) will be ratified soon, with hardware in maybe 12-24 months.
The openness is at the __ISA__ level. So you don't get a duopoly like being forced to choose between Intel & AMD. It means that anyone can start up a company and produce a replacement with proprietary graphics.
Yes, open source graphics would be nice, but that will have to follow a more open-standard CPU. So let's not put the cart before the horse, nor let the best be the enemy of the good. This is the next step.
ALso it has a PCIe connector where you can (presumably) add your own GPU if you really want/need to develop next-gen 3D applications for RISC-V
What's the alternative? I'm not aware of performant, open and cheap graphics options.
Note that they are already doing that. ¯\_(ツ)_/¯
We've got 60 years of experience with RISC ISAs showing that they are as good as any other approach, since Seymour Cray's CDC6600 in 1964 (and of course his later Cray 1 in 1975 would also be called RISC today). By the late 70s researchers at IBM were also converging on the same ideas, as were academics at Berkeley via experience with the DEC VAX. Multiple RISC ISAs, with mostly only cosmetic differences between them, hit the market commercially in 1985 and 1986 -- almost 40 years ago.
It's pretty clear what works, and for example Aarch64 and RISC-V were developed in parallel with each other [1], without either being aware of the other, but came up with very similar designs, both taking lessons of what to do and what not to do from MIPS, SPARC, Power, Arm32, Alpha and others.
In contrast, the first modern programmable GPU was the NVIDIA GeForce 3 in 2001, just 23 years ago. Even today leading GPU vendors have not settled down on how a GPU should be designed. Every NVIDIA generation is radically different from the previous one and NVIDIA, AMD (ATI), Intel, Imagination GPUs are very different to each other -- and the exact details are still often closely kept secrets. NVIDIA doesn't let anyone program their CPUs directly, they provide a pseudo machine code (PTX) and different closed source compilers from PTX to each generation's real architecture.
The chances of anyone coming up with the "forever" GPU design in the near future, let alone making it open, seem pretty slim. The whole field is still too new, changing too much.
[1] Aarch64 clearly started earlier, and unexpectedly published the ARMv8-A spec in October 2011, 1 1/2 years after the RISC-V effort started, by which time the fundamentals of the RISC-V design were firmly established, though details continued to change for the next 7 years. The first Android phones with Arm64 CPUs (e.g. Galaxy S6) hit in early 2015, with SBCs (Raspberry Pi 3, Odroid C2) following in early 2016. The first commercial RISC-V board, the Arduino-style HiFive1 shipped in late 2016. Note that Apple's iPhone 5s launched with an Apple-designed ARMv8 CPU almost 1.5 years before Arm's own designs.
[1] https://www.clockworkpi.com/product-page/devterm-kit-r01
[2] https://pine64.com/product/pinetab-v-10-1-8gb-128gb-risc-v-b...
What RISC-V board is on the open market and not supported by vanilla gcc/clang?
Really? IANA-KD, but I have read and seen grumbling from ARM developers about RISC-V "overreaching" and standardising more to prevent that kind of issue. Now, maybe the Chinese board manufacturers have gone rogue and ignoring these suggestions, but this fragmentation is something that RISC-V leadership was acutely aware and conscientiously made effort to address. I mean, I could imagine that at the embedded level sacrifices had to be made, but I'd be surprised at anything strong enough to power a ChromeBook and up to have this as an issue... but if you are better informed please share.
It's not their job. RISC-V Foundation is the entity in charge of specs, not manufacturers.
Platform specs exist, which cover boot and the environment a kernel will expect, among other things.
The manufacturers simply follow these specs. So far, quite successfully at that.
Unlike the ARM situation.
You didn't bring up extensions, but that's what people usually mean when they complain about fragmentation. It's not an issue in practice, just target RV64GC and you'll be fine. That describes the extensions which you can expect any general purpose computer style RISC-V CPU to implement.
I wish the booting story was better though, for both RISC-V and ARM. Manually editing device trees sucks. But that's more an issue due to the lack of UEFI on these SBC style platforms, not an inherent issue with ARM or RISC-V themselves (in fact, in the server world, ARM systems with UEFI is common).
0. bespoke boot process
1. a custom compiler
For the love of computer science, do not just repeat some FUD you heard somewhere. Always verify.
This is not FUD. The boot process isn’t better than on Arm, which is already terrible. End users need to find magic .isos, while developers need to search for magic vendor forks of everything: u-boot, kernel, OpenSBI.
Re compilers: Sure you can target RV64GC and everything more or less works, but your performance will likely be terrible. To squeeze more performance you might need magic or out of tree versions of GCC and/or Clang.
This one is a low end chip (single core, slow, based on the early 2017 specs) aimed at custom, embedded devices where the whole software stack (firmware image) is built from a single Makefile by the vendor.
I have a development board with one of them (only 512MB RAM, I managed to run a web browser on it at one point); I do indeed own too many RISC-V trinkets; this was the second RISC-V board I got.
Allwinner, same situation as their ARM chips (e.g. the A64 used in Pine64 I also own, or the A10 in the Cubieboard I also own). Most upstreaming work is done by linux-sunxi[0] instead of Allwinner themselves, which does not seem to care.
Not great in terms of already upstreamed drivers, which is why it is perceived as a pain. The hardware is also pretty limited, being a design that predates the base 2019 RISC-V specifications. But it can run Linux. The earlier Kendrite K210 (my and many people's first RISC-V chip) can run Linux too, but as an academic exercise. That one only has the builtin 8MB SRAM and its implemented privileged spec predates the first ratified one with significant differences.
Otherwise, standard RISC-V boot flow (i.e. "0" does not apply). In that sense, much better than its ARM counterparts, but you'll want to ensure you have the latest boot firmware in any event, not the early shipped one. This is a fairly early design too, so note the boot specs that were available at the time are far less reaching than what is available today, and were not even ratified yet. Despite this, the firmware has advanced somewhat, alongside the specs.
Runs generic RVA20 gcc compiled code just fine (i.e. "1" does not apply either) because the unprivileged spec ratified in 2019 didn't really chance since 2017. It does also have a pre-1.0 vector extension, which was very useful at the time to play with vector before the final version was available. GCC has support for this extension. Obviously, it is considered a "custom extension", and thus code using it will only work on processors implementing it.
TL;DR: One of the earliest designs, based on specs earlier than 2019, predating any ratified spec. Despite this, it manages to not meet 0 nor 1. Irrespective of all the FUD, RISC-V is doing great.
And, yes, that latest commit at the time was priv spec draft 1.9.1, which is quite different to what eventually got ratified.
For a while people trying to use the K210 thought the MMU was broken but it's not -- it's just an older spec.
Good luck cutting down any modern Linux kernel to boot in 8 MB RAM! Being able to do that might even predate Linux working on anything other than x86.
xv6 works fine though.
Edit: I was a little bit off. Linux 2.0 required 4 MB, recommended 8+, and already worked on Alpha, SPARC, MIPS, PowerPC, ARM, 68k. Linux 2.2 (1999) required 8 MB, recommended 16+.
The Pi 4 is five years old and is well-supported. Now. It was pretty awful in 2019.
The HiFive Unmatched (shipped this time in 2021, still available new) is well sorted these days.
I see a lot of complaints about the RK3588 boards stability and they're now two years old.
If you don't want to bleed, don't live on the bleeding edge, use something a few years old. Don't give in to the temptation of that 50%+ annual improvement in price/performance.