Hacker News Reader

Comment by hajile | Hacker News Reader

"We'll pay you to not switch to those cheap RISC-V cores coming down the pipeline".

How much cheaper are RISC-V chips in reality?

I get there might be a saving for not having to pay ARM a royalty of some kind, but are RISC-V chips cheaper in practice as a result?

I was under the impression (rightly or wrongly) that the arm royalties per chip were really small.

I think one of the biggest fears from ARM would be the popularity of the Raspberry Pi and their community.

There are better boards than the Raspberry Pi (strictly speaking specifications here). I took that path of playing with a lot of alternative boards, my biggest issue is lack of support, some boards had kernels never updated, etc...(YMMV).

If Raspberry Pi released a RISC-V board I have no reason to believe the community would not be just as strong. Sure in the beginning they would support both, but eventually the ARM support would wither.

> some boards had kernels never updated

As much as I love tinkering, this is why I stick with RPi boards and not others. RPis do everything I need, and have amazing support in software.

Is it truly amazing? I was under impression that Raspberry requires some blobs to run properly. Is there detailed specifications for Broadcom chip they're using? I was under impression that it was NDA and not possible to obtain for ordinary mortal. So may be it's good because of sheer number of people tinkering with it and smoothing rough edges, but it could be better. Please correct me if I'm wrong.

I would better phrase it as every other board has amazingly bad support.

Sometimes it is a matter of choosing the one that sucks the least.

> Please correct me if I'm wrong.

My memory told me it was the GPU that needed the blobs. So I asked at DDG

https://duckduckgo.com/?t=ftsa&q=binary+blobs+and+the+Raspbe...

Turned up this: https://wiki.debian.org/RaspberryPi and it says...

> All Raspberry Pi models before the 4 (1A, 1B, 1A+, 1B+, Zero, Zero W, 2, 3, Zero 2 W) boot from their GPU (not from the CPU!), so they require a non-free binary blob to boot

So the 4 (and I suppose the 5, if it ever actually comes...)

Goes on to say:

> Since then, Broadcom publicly released some code, licensed as 3-Clause BSD, to aid the making of an open source GPU driver. The "rpi-open-firmware" effort to replace the VPU firmware blob started in 2016. See more at https://news.ycombinator.com/item?id=11703842 . Unfortunately development of rpi-open-firmware is currently (2021-06) stalled.

So there you are. Not wrong, are you, but not strictly correct, depending on "...to run properly" definition

https://github.com/librerpi/rpi-open-firmware has updates 3-months ago

You're right. Pretty much all the low level stuff below the kernel in a pi is closed source.

Want your own custom boot rom so you can start up in half a second rather than the default 3 seconds before linux gets loaded? - sorry, we can't share the code for that with you, nor the specs for you to write it yourself!

There’s a few popular bare metal projects. Pi 1541 springs to mind. So it is possible, tho perhaps only with earlier generations

That's not what they meant. Yes, anybody can write bare metal alternative to Linux ( at the very least by looking at the Linux codebase). But still that Pi 1541 depends on the bootcode.bin, fixup.dat and start.elf binary blobs, which the OP was complaining about.

It wasn’t clear to me what they meant. I’m not familiar with the details of boot.bin but my read of what GP was implying was you had to use the rapsbian kernel and drivers. Thanks for the information.

Same. I would prefer to use some of the alternatives, but at the end the support level is important.

That said, I'll tolerate some regression in support to switch to a RISC-V based competitor.

Beaglebone?

And also lots of the alternative boards target RPI users ("like raspberry pi but better, cheaper etc"). If RPI switched then it would probably make sense for many of the other boards to switch to in order to stay comparable to RPI.

If they want to compete they have to agree on something like ACPI so the OS vendors can target their boards without a separate distribution for each board.

Most ARM boards already use Device Tree with Linux, there isn't any new agreement needed. What is missing is getting the drivers for each board into the same source tree.

NetBSD provides a filesystem image containing one kernel that will boot on all supported ARMv8 boards, you may need to write a board-specific build of u-boot to the start of that image. A Linux distribution could do the same as this.

Is Raspberry Pi really that big for ARM?

In the past, one of often cited reasons for the lack of success of ARM server offerings was that every developer machine is x64, no software is ever tested on ARM and no dev has a machine to do so. As the Raspberry Pi got more performant it brought a lot of people running software on ARM, leading to lots of software publishing ARM builds, toolchains getting exercised, bugs ironed out, etc.

With Apple having moved to Arm there are definitely ecosystem benefits brewing too. I can run a Debian VM out of the box with UTM, compile an Arm64 package using standard Linux toolchain and have it run directly on a RPi. It's like having a supercharged RPi development environment with me all the time.

I wouldn’t be surprised if the existence of many ARM docker images for hobbyist projects were indirectly due to the popularity of the Raspberry PI (in anrditi to Apple switching to ARM, more recently).

for hobbyists yes. Usually people stick to the technology they first started experimenting with, and RPi is that platform for many future experts.

If you are going to learn computer architecture, you will learn something cheap you have on hand

The technology is linux, gpios etc. not which instruction set the CPU has. That is completely irrelevant and raspberry switching to USB-C is a much bigger change from a user standpoint than switching to RISC-V would be.

Assuming performance and software support is comparable. Which obviously won't be the case for a long long time.

But there are few things as irrelevant as the CPU instruction set. (Part from specific extensions, like AES support enabling quick crypto etc.)

> The technology is ...

The technology is what people know. Using a different SoC board, camera, ... requires adds more time to gain the same level of knowledge.

Doctors will latch onto a single product solution so they don't have to spend the time learning how to operate an alternative. Hospitals need to stock consumables based not on the best products but on what doctors know.

Airlines retain the same air crafts to reduce time spent learning to operate an alternative. Boeing marketed this as a sales feature with 737 MAX, no extra flight training required!

Software developers will often stick with the same language, even though others better fit the domain problem. Few seem willing to take the time to try and play with new concepts, languages, and operating systems.

Trying new and different things drives innovation not the world.

Exactly my point. Changing the instruction set on the CPU won't change the API for the camera etc.

Some things will surely change because the hardware backed features work in a different way, but mostly it will be similar enough.

Going from a raspberry pi to an orange pi could be a much bigger leap than switching to an raspberry with RISC-V that has mature software support.

Very much disagree. Of course linux and GPIO is important, but the widest use case for them among myself and people I know is as a build box and/or something to test ARM software on. One person uses it to learn ASM. From my small and surely non-representative sample, the architecture is maybe the most important thing. So I don't think we can confidently say that CPU arch isn't relevant.

But instruction set, in practice and in this case, is tightly coupled with form factor and MIPS per watt. I work in mobile robotics: my low end choice is RPi, high end an NVIDIA board. While I can see Risc-V challenging ARM here, they don’t yet. (Excepting an ultra low power/low compute edge-case.) I just don’t see any CISC architecture that’s available today competing.

> But there are few things as irrelevant as the CPU instruction set.

I’m going to have to disagree with that statement…

… but only because I’m in the middle of goofing around with ARM assembly on an RPi as we speak.

I ... do not know.

The Amiga 500 didn't make people stick to the m68k. They went IBM PC as did everyone else.

And they are still bitter about it !

There aren't many other options if you want or need a system with physical access near your desk that is also designed to run GNU/Linux. I think the only other options are less-known SBC-based systems, or systems where GNU/Linux is an afterthought at best. With the Raspberry Pi 4, it's quite likely that one of your colleagues already got the locally preferred distribution to boot on it. Lots of people use them to reproduce and fix generic AArch64 issues, even if they have remote root access (including the ability to install another OS) to much faster lab machines.

I think a lot of the motivation here is to avoid RPI leading to further interest and development on the RISC-V side. SIFIVE may be willing to make something perfect for the RPI, but the real risk is if RPI goes RISC-V, then software gets ported and tested on RISC-V. People start liking it and hacking it. And then RISC-V suddenly gets a lot more interest from ARM licensees and their competitors. Which, at the very least, will lead to higher licensee leverage in negotiations with ARM on renewal.

Bingo. This is what they mean by "strategic investment." It's not because they are just lovers of the education and maker market. They'd be negligent to their shareholders to not make this investment.

>but the real risk is if RPI goes RISC-V, then software gets ported and tested on RISC-V.

It's already happening. RISC-V doesn't need the Raspberry Pi.

Yet it would indeed be an accelerator, but I'm not sure how big. ARM are certainly very afraid.

Cheaper is not even the primary reason to use RISC-V over Arm.

The ability to modify the design for your application and be able to apply a plethora of cores where they are needed at only the cost of silicon area.

Apple has been moving their management cores on their M-series parts from Arm to RV and they have an architectural license.

Everyone has an architectural license for RISC-V, you can add your own instructions, change the mix of available instructions. A whole parametric RV32 or RV64 will be available on every node at every fab.

This move by Arm is absolutely to block RPI from moving to RISC-V. They have mindshare and distribution.

Other than geopolitical and totalitarianism concerns, is that a good and bad thing that let you freely change your instruction … implementing no charge I understand, but everyone has their own architecture …

Art is not just about no limit, but what you do with the limitation including illuminating there is a limit. Not just about beyond it but of course you could.

Not sure, but really small royalties on $4 computers can still be accumulate. They may also prevent $4 computers turning into $1 computers even if huge volumes would call for this otherwise.

Moreover, those royalties put fences on what might otherwise be more open and collaborative. It would be sad, if the tinkering goals set out originally by Raspberry Pi were curtailed by ISA license restrictions. RISC-V inherently has an edge over ARM in that dimension.

SiFive would practically give away their latest P870 cores (close to A78 performance levels) to get them shipping by the millions in the Pi because of the free advertising, free exposure, and massive boost it would give to the development of the RISC-V ecosystem.

Once such a switch was made, Pi could even consider using free and open source cores for their low-end devices where margins are slim (an area where RISC-V has already been gaining ground rapidly).

Where can I buy one of these p870 cores and how much do they typically cost? (Or boards containing them?)

These are the newest OoO cores with full 1.0 vector extensions, right?

As a normal consumer you can't. But the predecessor P670 should be available in the SG2380 board in 2024Q3 [0]. It will have 16 P670 and 8 X280 cores, and will cost $120 to $200 (without included RAM).

[0] https://forum.sophgo.com/t/about-the-sg2380-oasis-category/3...

>P870

These cores have just been announced, and are available for licensing. You can contact SiFive if you're interested.

If what you want is hardware somebody's already made, that's going to take 2-3 years as per tradition.

>These are the newest OoO cores with full 1.0 vector extensions, right?

Yes, but so are multiple generations of predecessors. It seems that hardware based on P670[0] and X280[1] (both match that description) will be available for purchase in less than 10 months from now[2].

0. https://sifive.cdn.prismic.io/sifive/7be0420e-dac1-4558-85bc...

1. https://sifive.cdn.prismic.io/sifive/9405d3d0-35a1-4680-a259...

2. https://forum.sophgo.com/t/about-the-sg2380-oasis-category/3...

A tougher question as so many factors that I hope Asianometry on YouTube cover one day.

  Whilst the RISC-V instruction set is free from licence/patent fees, the design of those chips will be made by a company that will need to recover costs, so there will be a cost.   Compared to ARM who already offer up free core designs for use for free like the cortex-m0 and others.  I know the Raspberry PICO uses a cortex-m0x.

Though, many do seem to blur the lines between the instruction set and the core design you get in the chip you buy.

Over time, but that will be when you get open-source RISC-V chips that compete with the designs of the market offerings other companies make and sell. Which as a mindset may cause the evolution of RISC-V issues as people could get burned due to expectations exceeding the reality and finding the get what you pay for those to still hold stronger than envisioned. Yes, eventually open-source CPU cores will get there, but production costs and scale will still become a factor, as to many aspects that all add up to the final cost. This is even on a level of software stack tried and fully robust of equality, which is still behind ARM.

ARM may well kill off x86 as a mainstream before RISC-V fully bites it.

Look at how long ARM took to get where it is, and started with microcontrollers, used in many things. That is where RISC-V will and is starting to get traction, but scaling beyond that, whilst could be faster than ARM's history, it's not as clear-cut as many foresee.

ARM did not start with microcontrollers - it started with desktop CPUs. In fact, it was a very long time before it got to microcontrollers!

For "coming down the pipeline" they're essentially free.

Today, the c910 is an Apache 2, hardware proven out of order core on GitHub here https://github.com/T-head-Semi/openc910 a little slower than an RPi3's core.

What’s the price per unit?

There are no license fees if that's what you're asking.

Additionally it's very competitive in PPA metrics for it's gate count, so cheaper than similar cores in terms of wafer area as well.

I think the point is, very few people are looking to buy some RTL, they want the silicon. So the question is what's the price of the silicon? My understanding is anything comparable to the SoC in the RPi boards is still quite a lot pricier.

Pretty damn cheap.

https://www.aliexpress.us/item/3256805346421328.html?gateway...

Every other place seems to claim that the RRP is $119. So this must be a mistake/scam? Or did the price go down this much?

Good question. I looked it up and apparently the royalties are around 2% of the price of the chip, so it probably won't make a huge difference. They do also charge an up front "membership fee" which seems to vary from $200k to $10m depending on the chip.

RISC-V is very very popular already, but most RISC-V cores are not user accessible. They're controllers in hard drives, embedded management cores in SoCs, etc. Definitely nice to not have to pay for licensing that, and verification doesn't matter so much since you are more likely to be able to work around bugs in software.

I suspect it won't make a huge difference to visible CPUs. Probably the biggest impact will be flexibility.

Regardless of cost, working with almost anyone other than broadcom would probably be seen as a huge plus as well.

Unpopular opinion, backed by the current status quo (https://www.jeffgeerling.com/blog/2023/getting-risc-v-again-...)

It will take 3-5 years for RISC-V to catch up with ARM in any way that is meaningful for the Raspberry Pi's target segments. Maybe some industrial applications, but I really doubt we'll see RISC-V silicon with enough clout and power efficiency for a long while.

I thought we were getting close to Pi4 speeds with the latest stuff from SiFive and there's some GPU support.

> 3-5 years

> The Raspberry Pi 4 Model B was released in June 2019

The math checks out.

The acceleration is unlike anything ARM has ever had.

                           x86                       ARM                      RISC-V
    Introduction         8008 (1974)             ARM1 (1985)              Raven-1 (2011)
    first OoO            NexGen NX586 (1994)     ARM A9 (2007)            BOOM (2015)
    first multicore      AMD Athlon X2 (2005)    Nvidia Tegra 2 (2010)    EOS16? (2012)
    first 64-bit         AMD Athlon64 (2003)     Samsung Exanos 5 (2013)  Western Digital SweRV? (2017?)
    Bleeding Edge          --                    Apple M1 (2020)          TBD (Tenstorrent Ascalon 2024?)

Catching up is faster than doing something the first time (and it's even easier if you don't have loads of legacy edge cases slowing down progress).

x86 was 20 years to OoO, ARM was 22 years, and RISC-V was just 4 years.

x86 was 31 years to dual-core, ARM was 25 years, and RISC-V was just 1 year.

x86 was 29 years to 64-bit systems, ARM was 28 years, and RISC-V was at most 6 years.

ARM took 35 years to catch up to x86. RISC-V seems set to do that in 13-14 years (if you only start counting when RISC-V had privilege systems, that's like 5-6 years and if you only count since vectors were added, it's just 3-4 years.

Put everything on a graph and RISC-V is blowing the doors off with how fast it has gone from unknown to next-gen tech.

While I don't think your assessment is inherently wrong, part of that I think is that fabbing custom silicon is easier than ever before.

Don't get me wrong, being open is doing wonders as well but just in general, nowadays fabbing ASICs (and the experience that goes with it) is so much more accessible to smaller, less risk averse players than when ARM (or x86) really started getting big.

Simply put: technology advanced to superscalar, OoO cores & 64 bit multi-core cpus. These advancements took decades to pan out.

Both x86 and ARM had to 'ride the wave' when that tech came along.

RISC-V otoh can pick up the result of those advancements, piece it together for a new ISA, and go. No waiting-until-tech-advances other than the wait for design work & prepping those designs for mass-production. Both of which can take considerable time. But not decades.

You should be looking at Rockchip's stuff if you're thinking about both hobbyist grade and industrial use. The RK3588 is still slightly ahead of the Pi 5, and several orders of magnitude above what RISC-V SBCs can deliver.

Software-wise, RISC-V is way behind as well.

>The RK3588 is still slightly ahead of the Pi 5, and several orders of magnitude above what RISC-V SBCs can deliver.

Not that many orders of magnitude. There's RISC-V hardware (actual boards for sale) that's faster than RK3588 (thus RPi as well) coming in 2024.

This includes the SG2380 Oasis board[0] in 10 months, and a possibly earlier JH8100-based visionfive 3 board.

And we aren't even talking about large implementations like Tenstorrent Ascalon (expected to be competitive with Zen5, M4).

0. https://forum.sophgo.com/t/about-the-sg2380-oasis-category/3...

Well, for industrial use I love the documentation and support of Raspberry Pi, unless you specifically need performance not available on raspberry pi.

Wow, has it been that long already. I'm getting old!

It'll take at least 5 years and 100s of megabucks to get to where rpi5 is today.

It's possible that some of those 5 years and some hundreds of $100M have been already spent under wraps by Quallcom or whoever else, though. SiFive fiasco recently may prove it isn't easy even if the dollars are there.

Billions of bucks have already been spent.

The first large hardware showcasing this is arriving in 2024, including Tenstorrent Ascalon, expected to be competitive with Zen5.

It's MUCH more "We'll pay you to familiarize new embedded engineers with ARM processors so they're more likely to pick ARM SoCs for their (eventual) commercial projects."

It's smart marketing. And yes, RPis going to RISC-V would eventually be damaging to ARM all out of proportion to the number of RPis sold.

That is the first thing that came to my mind lol.

So, perhaps a good time to start a new business selling inexpensive RISC-V based micro computers?

You'd already be late[0][1][2], but starting today is still better than starting tomorrow.

0. https://www.kickstarter.com/projects/starfive/visionfive-2

1. https://wiki.pine64.org/wiki/STAR64

2. https://milkv.io/

About that community and dedicated resources, though. And branding.

Luckily the first users of the Raspberry Pi weren't discouraged by the lack of all of the above.

They weren't discouraged because it was an almost entirely unique product in the market at the time. It was Raspberry Pi or nothing. Demand for such a product was obviously huge (even bigger than RPi anticipated), so with no other options in the market, all the weight was put behind RPi by default. Something new today would need to siphon away a bunch of the RPi community to get the same momentum, and that seems unlikely because for the most part RPi treats the community pretty well. Aside from the shortages and allocation strategies, they do a great job of having polished solutions with software that works well and keeps getting updated.

Maybe I'm remember incorrectly but I was in middleschool when the RPI 1 came out and my dad got me one.

There was tons of guides, tutorials, etc. when I got it, which I think was close to launch? The os image was already made, tutorials on how to set it up, use ssh, setup a basic webserver, etc were all there at or very near launch.

---

Software support is almost never an issue for sbc's, and because they run linux, there's a significant amount of resources of how to use them.

Usually they fall apart because they don't have on-going software/firmware support the way RPI has always had a very well supported, and up-to-date linux image ready to flash.

---

Although I will admit I think far more packages were compiled for arm back then compared to riscv packages today.

> Although I will admit I think far more packages were compiled for arm back then compared to riscv packages today.

You can track the debian build progress here:

past two weeks: https://buildd.debian.org/stats/graph-week-big.png

past quarter: https://buildd.debian.org/stats/graph-quarter-big.png

Woah that's massive progress this year, dang.

The community has become a lot less important, as arm devices become more and more generic and supported. Nowadays you don't need any special distro or kernel to boot on most boards, so I can't see the brand name pulling them through if they don't maintain a price/capability edge.

This has not been my experience. Straying from RPi has each time left me dependent on community forums and stuff just to get stuff working, and a couple of times has left me with a completely dead device and no ideas how to proceed

You and me both :(

You mean like Milk-V or sipeed?

Reply on news.ycombinator.com