Building Debian For RISC-V Currently Relies Upon Nine HiFive Unmatched Boards

[phoronix]

Phoronix: Building Debian For RISC-V Currently Relies Upon Nine HiFive Unmatched Boards

RISC-V is now an official Debian architecture for the Debian 13 "Trixie" release to happen in about two years time. Over the weekend a brief status update was issued surrounding this newest CPU architecture to be supported by the Debian GNU/Linux team. Arguably most interesting is how they are currently building out the Debian RISC-V packages...

https://www.phoronix.com/news/Debian-9-HiFive-RISC-V-Boards

[schmidtbag]

Though the HiFive Unmatched performance with the u740 doesn't come close to modern x86_64 CPUs.

Gee and people think RISC-V is going to be a Zen5 competitor because of a single cherry-picked graph...

[aviallon]

I am wondering if they could simply use a qemu emulated system.
Even with the added emulation cost, the performance of a good AMD Epyc should be well enough for building much faster than a few RISC-V boards.

[wswartzendruber]

I am wondering if they could simply use a qemu emulated system.
Even with the added emulation cost, the performance of a good AMD Epyc should be well enough for building much faster than a few RISC-V boards.

I had just assumed (until now) that this was already the case.

[gnattu]

I am wondering if they could simply use a qemu emulated system.
Even with the added emulation cost, the performance of a good AMD Epyc should be well enough for building much faster than a few RISC-V boards.

Actually you only need an 5950x to outperform HiFive Unmatched, and actually 3x fast in complication.

But 3 times fast is still not fast enough because it still takes more than a single day(35hrs) to compile chromium on 5950x with qemu-user. On HiFive Unmatched, it takes more than 100 hrs. This is why they said the following:

...we are aware that the build time for some packages is quite important. We are actively working on acquiring next generation RISC-V hardware...

Complication can be done with qemu-user but they also need to do the testing and that would be tricky with qemu-user, and probably much easier with real boards.

[phoenk]

If performance was the main metric you cared about, you could just cross compile it from x86-64. You'd get native compiler speeds because you're not emulating anything, the compiler is just outputting risc-v code instead of native.

[CommunityMember]

I am wondering if they could simply use a qemu emulated system.

After the initial bootstrap, many distros want to have real hardware for dogfooding purposes. Some distros require builds to be performed on real hardware in order to achieve supported status (also, typically, ECC memory). And as most such build hardware is remote, having full remote server management is often a practical requirement for a build farm at larger scale, and most of these early development boards have limited (to no) remote management capabilities, although such server class hardware for RISC-V is in various stages of development and availability (and should be available before Trixie is formally ready to ship as stable (two years is nearly forever in IT)). I recall discussions elsewhere where distros have talked about plans for RISC-V and are waiting for those server class RISC-V systems to be available.

[ayumu]

Realize this HiFive Unmatched that Debian is using was released May 2021.

The cheap VisionFive2 SBCs released early 2023, simply based on a newer revision of those same U74 CPUs, is already significantly faster.

The recent XuanTie C910-based designs, such as Sipeed LM4A, are twice as fast.

That RISC-V Zen5 competitor by the lead architect of Apple M1 is due 2024, like Zen5 itself.

[ayumu]

Gee and people think RISC-V is going to be a Zen5 competitor because of a single cherry-picked graph...

It seems you're talking about Ascalon, but for some reason you're choosing to ignore many facts that you have been spotted arguing against, and thus you're no doubt aware of, such as:

• Tenstorrent CEO is Jim Keller, industry veteran with an ample history of success, which we have no reason to think he can't replicate again with RISC-V.
• Ascalon is led by Wei-han Lien, who previously led M1 at Apple. No reason to think he can't do it again with RISC-V.
• The "cherrypicked graph" isn't some synthetic benchmark, but SPEC 2017, which is an aggregated result which includes a wealth of real-world application benchmarks.
• Ascalon's a real result, whereas Zen5's projection has been widely considered too high a estimate by the tech press.
• "Gee..." HiFive Unmatched was released in 2021 and is based on pre-2019 state of the RISC-V specification, whereas Ascalon implements 2021's RISC-V, including ratified V, hypervisor and bit manipulation extensions.

Yet, you're deliberately making a strawman out of fine arguments by RISC-V proponents, and thus choosing to argue in bad faith.

Bad schmidtbag. Bad.

[gnattu]

Realize this HiFive Unmatched that Debian is using was released May 2021.

The cheap VisionFive2 SBCs released early 2023, simply based on a newer revision of those same U74 CPUs, is already significantly faster.

The recent XuanTie C910-based designs, such as Sipeed LM4A, are twice as fast.

That RISC-V Zen5 competitor by the lead architect of Apple M1 is due 2024, like Zen5 itself.

A 64-core 2GHz C910-based CPU(SG2042) is barely faster than 5950x with qemu-user in complication, and currently having compatibility issues (like nodejs crashing on that core, preventing chromium to be built). At current state the C910 based CPUs cannot be used for a build farm until the compatibility issues are resolved.

By the way, the SG2042 features a terrible inter-core bus with exceptionally high inter-core latency; it absolutely cannot provide adequate bandwidth for all cores. This issue is part of the reason why the performance of the 64 cores scales so poorly. Although chromium cannot be built on that particular cpu, we do have numbers for building Firefox. C910-based SG2042 used about 140 minutes, 5950x on qemu-user used about 170 minutes, and HiFive Unmatched took more than 11 hours.

Indeed, HiFive Unmatched is quite slow, but a 64-core SG2042 being only five times faster in compilation than a quad-core 1.2GHz U74 CPU, despite being based on 2.0GHz C910 cores, is simply unacceptable. If you have the opportunity to use an SG2042 machine, you will find its memcpy speeds so slow that you have to wait several seconds even for a simple dmesg in an ssh session. Unfortunately its on-chip bus is that slow. We have to wait longer for a real properly designed server-class RISC-V CPU.

Edit: Although SG2042 is not well designed to accommodate with its core count, it is still the fastest RISC-V CPU that money can buy at the moment. If the compatibility issues are resolved it could still be an option that is way better than HiFive Unmatched.

Edit2: Forget to mention, the SG2042 requires you to use a vendor kernel at the moment and mainline has still a long way to go. This is another reason preventing distros like Debian to use it.