If you're in the business of selling computers, a factor you can not ignore
in all this is geopolitics.

Large regions of the world (China, the global south, the EU) are looking to
RISC-V to break the
control of the US over architecture. The word they use is sovereignty. This
movement started under Trump 1, when the RISC-V Foundation
moved to Zurich and became RISC-V International. As you can guess, it's
exploded in importance since Jan 2025.

There is reason to believe that future RFQ will mandate RISC-V, even if the
result is a more expensive or less performant
system, just to avoid any dependence on US technology. US companies are
preparing.

China has taken a strong lead with the one student one chip program,
https://ysyx.oscc.cc/en/project/intro.html.
They are training close to 20,000 students a year in designing RISC-V
parts. In the EU, we see their efforts at
https://openchip.com/openchip-unveils-ber10-europes-new-sovereign-chip-is-here/
. Brazil has an effort too.

RISC-V is following the path that SPARC/MIPS, then x86, then ARM followed
before it, starting small and climbing the value chain.
Few remember when the 386 topped out at 12 MiB/s memory bandwidth, and was
considered a joke for the high end;
when x86 achieved 30 GiB/s memory bandwidth 15 years later, people took
notice. For reference, Cray was
the gold standard into the 90s, with 12 GiB/s on typical systems.

China just took the lead again with Line Shine:
https://en.wikipedia.org/wiki/LineShine. This is a non-GPU machine
with a tensor unit. There is a lot of expectation that, in the 2-3 year
timeframe, those ARM parts become RISC-V parts.

Overall, something like 17B RISC-V parts shipped in 2025, in systems most
people don't know about. RISC-V long ago
overtook x86 and ARM in many applications.

So, yes, you can ignore RISC-V for a while yet. But not for long. I choose
not to ignore it, even with its limits and flaws,
because I've been here before, first with workstation processors for HPC,
then x86, then ARM.

Further, RISC-V comes with a few nice ideas that make new approaches to old
problems pretty easy. I'll be talking
about one such idea at supercomputing 2026.


On Mon, Sep 21, 2026 at 7:57 AM Jiji Freya Daniel Maslowski via 9fans <
[email protected]> wrote:

> You don't need to use the vendor firmware at all. Most SoCs I have
> seen allow you to run custom firmware.
> I'm one of the developers behind https://github.com/oreboot/oreboot
> and we're targeting many RISC-V platforms. We offer an SBI through
> RustSBI, but we're also designing oreboot such that you start
> executing in M-mode, which is where you can then set up your own
> exception handler and define any interface to S-mode, or just put your
> kernel there if you don't need an MMU. You can only use virtual memory
> once dropped into S-mode, that is.
>
> On Mon, Sep 21, 2026 at 7:16 AM <[email protected]> wrote:
> >
> > On Sunday, September 20, 2026, at 11:13 AM, Jiji Freya Daniel Maslowski
> wrote:
> >
> > There is a software abstraction to cater for general purpose software:
> SBI. If you don't like it, ignore it and create your own stuff. It's
> trivial to do, just tedious.
> >
> >
> > This confuses me.  I ended up doing 'a bunch of' research to figure out
> what SBI was, and finally arrived back in Mr. Collyer's article.  He
> clearly has a strong dislike of it, and yet, couldn't get around having to
> use it.  I conclude that either you know how to get around it even if he
> didn't, or that you are referring to more recent tech where it -can- be
> circumvented.
> > 9fans / 9fans / see discussions + participants + delivery options
> Permalink

------------------------------------------
9fans: 9fans
Permalink: 
https://9fans.topicbox.com/groups/9fans/Tbf390d3a51f65cdf-M9d397fd9d35a4506ce66486e
Delivery options: https://9fans.topicbox.com/groups/9fans/subscription

Reply via email to