Milk-V Pioneer – native RISC-V computer | Hacker News Reader

Milk-V Pioneer – native RISC-V computer

milkv.io

95 pointstheshrike7964 comments

Sophon SG2042, is that a reference to the Three-Body trilogy? Kinda disconcerting if so, for people that know the storyline... The teardrop logo also seems to be a reference...

https://en.sophgo.com/product/introduce/sg2042.html

I hope that if they end up holding the sword, they're also willing to actually push the damn button when needed.

for safety reasons

if you have the sword the button cannot be pushed, only stabbed (?)

OMG why why why WHY did you disable the scrollbar on your web page?

Just why?

UI standards exist for a reason!

I agree that is terrible.

On Windows 11, I confirm Chrome shows it, Firefox hides it.

They're setting scrollbar-width: none; which only Firefox currently supports [1].

Maybe it's done by their website generator? It's Docusaurus who claims to take accessibility very seriously...

[1] https://developer.mozilla.org/en-US/docs/Web/CSS/scrollbar-w...

links and lynx forever...

It's encouraging to see the milestone achieved of being able to buy a strong desktop computer based on RISC-V as an off-the-shelf product. Maybe there is a future for FOSS-based CPUs :-)

Anyway, I like their cheese grater idea:

https://milkv.io/assets/images/workstation-425fa8f62d1320fdf...

I wouldn't necessarily call it a 'strong desktop'—it has many limitations, and is more geared towards those working on RISC-V enablement, or hackers/tinkerers who want to learn more about RISC-V on a decent little workstation.

It's performance is abysmal compared to even tiny low power desktops, and it does have a good deal of IO, but a lot of it is severely constricted by the slow-ish CPU cores, and lackluster (so far) software support.

This is a machine that's more meant to get more people into the ecosystem, to help so future hardware can be that 'strong' RISC-V desktop.

Linus Tech Tips recently did a video on this box (I think last week?), and it has a lot of good demonstrations.

Performance per core is the same as (even a little better than) Arm A72, as found in the Raspberry Pi 4 or -- more appropriately -- the Graviton 1, both from 2019, except it has 16x more cores than the Pi 4 and 4x more cores than the Graviton.

That puts shipping RISC-V hardware four years behind Arm.

By mid 2024, shipping RISC-V hardware will be competitive with Arm A76 SoCs such as the RK3588 (shipped on boards such as the Rock 5 and Orange Pi 5 1.5 to 2 years earlier) or the Raspberry Pi 5, which is only about 6 months ahead (I received mine, ordered at announcement, this week).

By the end of 2024, shipping RISC-V hardware will have 16x Arm A78 equivalent cores, on a board that realistically will cost $300.

People seem to forget that before the introduction of A76 and the Neoverse cores ARM was incredibly slow compared to x86 with the exception of Apple's chips. High performance ARM cores are a recent innovation of the last five years.

How could it be same as ARM Cortex-A72 without SIMD (vector) instructions and being in-order ?

The SG2042 SoC in the Milk-V Pioneer has Out-of-Order THead C910 cores with very similar design to an A72, and with dual 128 bit vector units (RISC-V V extension draft version 0.7.1).

Cortex A76 hardware was first available in 2018. Just because A76 happened to appear on an SBC 18 months or six months ago using this as a milestone doesn’t give a realistic impression of the relative timelines of the two architectures.

Cortex A76 hardware was absolutely not available in 2018. That's when the IP for the core was announced. It takes years to get from that to mass-production hardware.

The dates that SBCs came out is relevant, and a very consistent data after the announcement of the CPU core.

C910 core announced July 2019, SBCs delivered June 2023 (Lichee Pi 4A) and December 2023 (Milk-V Pioneer). ~4 years

U74 core announced October 2018, SBCs delivered February 2023. ~ 4/12 years

A76 core announced May 2018, SBCs delivered in mid 2022 (Rock 5) to December 2023 (my peordered Pi 5 arrived this week). 4 - 5 1/2 years

A72 core announced February 2015, SBCs delivered June 2019 (Pi 4). 4 1/2 years

A53 core announced October 2012, SBCs delivered March 2016 (Pi 3, Odroid C2). 3 1/2 years

Very high priced prototype hardware based on "shuttle run" chips can be available as much as two years earlier. High end phones are also often around the same time, as one or two privileged chip designers often get access to beta core IP well before announcement and can drop final RTL into an already designed SoC.

Some things are also announced long before they are delivered, while others are announced when stock is already available.

But actual delivery of SBCs to customers occurs a remarkably consistent amount of time after formal announcement of CPU cores.

https://youtu.be/vaMxTSm53UU

Link to the LTT video

I wonder what the firmware blob situation is with this. How much documentation is available for this SoC? It would be nice to have more platforms with fully open source firmware.

$1,500 is a bit much. I'd love this as a replacement for my aging NAS/webserver, though.

64 is a lot of cores, getting that many cores in something like a Threadripper isn't cheap.

But I wonder what the relative performance is. Four channels of 3.2Ghz DDR4 is theoretically a good amount of aggregate memory bandwidth to keep those cores fed...

Why 64-cores for a NAS though?

The XuanTie C920 cores in the SOPHGO SG2042 are an older design without support for Bitmanip, cryptography or Vector 1.0. You could compile especially for it though, and take advantage of T-Head's proprietary ISA extensions and the draft version of Vector supported, but it would still be far from comparable to a Zen core.

Are those extensions really what matters in comparing to a Zen core?

I assumed these cores are lacking in the OoO/speculative exuction/branch prediction department. It makes them smaller so it's cheaper to have many of them, but won't be as fast as any state of the art x86 cores.

Special extensions don't generally automagically get utilized even when they are present...

I don't think Bitmanip is special but pretty much required for reaching feature-parity with ARM and x86 so that you could start comparing them. RISC-V is otherwise quite spartan, which might be desirable for embedded MCUs, but for application-class processors you'd want more. For example, an indexed load which on ARM and x86 is one instruction could on RISC-V require four dependent instructions without Bitmanip and two with.

The vector extension is RISC-V's closest counterpart to SSE/AVX, although it is not used for scalar floating-point.

BTW, the RISC-V foundation has published official Application Profiles for application-class processors. Those are probably going to be baseline for binary distributions in the future — they include Bitmanip and more. The RVA23 profile has Vector. Google's requirements for Android (with Google services) will also include vector cryptography.

Of course these cores are approximately 4x slower than current x86 cores, but you get 64 of them for the price of an x86 with 12-16 performance cores.

If you want 64 cores in, say, a Threadripper 7980X then you're looking at $5000 just for the chip, with motherboard and RAM etc on top of that. Yeah, it's 4x faster, but it's also 4x (or more) the price.

Fewer faster cores is better than many slow cores though. It's just harder to do.

From the benchmarks in the paper that I think everyone is basing this on: https://ui.adsabs.harvard.edu/abs/2023arXiv230900381B/abstra.... The 4x slower single-core seems to be based on a EPYC 7742. A modern 7950x has approximately double the single-core performance compared to that EPIC chip, with very close multi-core performance.

So the comparison is more like 4x faster (or 8x if single-core matters) for 1x the price, since you can easily get a similar rig with a 7950x for $1500.

$1500 is competitive with, say, a Ryzen 9 7950X with about the same aggregate performance, remembering that for your $1500 you're getting motherboard, CPU and 128 GB RAM.

that bundle gives you 1/10 of the performance of a typical x86 server system.

let's don't forget about that.

Well you could still replace an older x86 server system by this RISC-V machine.

by paying thousands of $ just because it is risc-v?

sorry, but I can't afford such price tag at the current laughable performance ratio.

Because it is higher OPS per watt than your older x86 server, you will save in the longer term.

3 pci slots in the photo and two in the specs. Am I to understand it’s 2 @8 and 1 @16?

The specs were changed in several ways before production. This was in a blog post a few months ago.

While I wanted to try and support this machine, I cancelled my order for it once I found that its vector support was outdated and would be a PITA from a toolchain POV.

Boards with a newer SOPHON chip based on SiFive P670 and X280 cores (much faster and V capable) can be preordered, TBA 2024Q3.

You have a URL for that, for pre-order?

https://www.reddit.com/r/RISCV/comments/17aq0tq/first_peeks_...

What do you mean "once I found"?

It's not a secret that these CPU cores were announced in July 2019 (four years is a normal time span from core announcement to SBC availability), when the current vector specification was draft version 0.7.1 (which it implements), while version 1.0 was finalised only in November 2021.

You can't design hardware for a specification that doesn't exist when you design it.

You also can't go instantly from CPU core design to a finished product delivered to customers.

Deliveries of a board with an SoC with one low performance implementation of RVV 1.0 -- based off the four year old C906's 0.7.1 implementation -- have just started. High performance implementations will start to arrive in customers' hands only in late 2024.

  > L1 Cache I:64KB and D:64KB

it seems quite large to me...(?)

i wonder what kinds of workloads can benefit from that?

Join Milk-V WeChat Group

Join Milk-V QQ Group

How is one suppose to use these QR code on a phone? ( Not that I have QQ or WeChat )

Edit: I misunderstood the question. See below.

WeChat has a button to add a contact / group that will open a camera, scan the barcode and do the magic. Same is in Telegram. Not sure about QQ, but I would guess it's all the same.

I think what they mean is, how do you scan the QR code if you are using the phone to look at the page.

Perhaps WeChat can also open images from camera roll, so you can read QR from a screenshot?

Edit: I discovered something: on iPhone you can long-press a picture of a QR code (don’t even need to screenshot it) and iOS will give you the option of navigating to the URL of the QR code

I see.

> I discovered something: on iPhone you can long-press a picture of a QR code (don’t even need to screenshot it) and iOS will give you the option of navigating to the URL of the QR code

Nice! I just tested on an Android phone, and a long click on a QR code gives a menu that has an option "Search image with Google Lens", which in turn scans barcodes.

Neat trick, but I still desperately wish that there was a way (I certainly cannot ascertain how) to copy the QR code contents without opening the link. As far as I can tell, iPhone will only give you the option to open in Safari if it is a url.

>Perhaps WeChat can also open images from camera roll, so you can read QR from a screenshot?

It can scan QR codes from images in the phone's gallery.

On Android, Binary Eye is my go-to QR code scanner. It's FOSS, and I just found out that you can share QR codes to it and scan them! Great app.

https://f-droid.org/en/packages/de.markusfisch.android.binar...

https://github.com/markusfisch/BinaryEye

Ha yes, the worst thing is esims - the only way that works reliably to install them is via QR code, but Android will only let you scan them using your camera. Not helpful when you have the QR code on your phone.

It's almost impossible to create a QQ acount these days outside of china.

Does it run Doom?

yes, it does.

RPi Pico does.

"- White slim PC enclosure with handle"

Looks odd. I would have preferred a 19" rack.

Does it run Crysis?

This is a legitimate question. Unlike Doom though, it has not been ported to other archs.

You can use an emulator.

Wow, 64 cores. I'd love to see a beowulf cluster of those.

you need to have a tumor in the brain to fork out $1,500 to pay for a half baked dev machine that just deliver 1/10 of the performance of an intel/amd server processor.

The point is to use this as a RISC-V development machine, not to get performance parity with x86.

If you can't even get decent performance out of such so called "development machine", what is the point using it rather than just run qemu?

If it is not about performance, what is the point having multiple cores in the first place?

If the machine is so slow to the extent that you can be running 10x slower than an AMD server processor 2 generations behind, do you actual consider such environment to be realistic?

Everything is about performance.

  > what is the point using it rather than just run qemu?

risc-v qemu is not really fast enough for desktop gui work, even with virgl acceleration (it does make it barely usable though) its still too slow for most stuff like that especially gnome....

Core for core it's as fast as or faster than QEMU running on current x86.

An x86 machine with 64 cores costs $5000 just for the SoC, probably $7000 for a complete machine. This is $1500 for motherboard + CPU, $2500 for a complete system with 128 GB RAM.

and that's only the board and the cpu. the case costs $2,500

Yeah… I’ve never seen RGB coolers that expensive.

it is pretty much an online free IQ test.

That includes 128 GB of RAM, which costs $400 to $500. And a 1 TB SSD, about $100 for a good one. And a video card. Cases and power supplies aren't free either.

If you bought just the motherboard with CPU and built up something with equivalent specs yourself, from new parts, you might end up at less than $2500, but it won't be by much.

Reply on news.ycombinator.com