Hacker News Reader

Comment by snvzz | Hacker News Reader

This is TH1520, same SoC as the Lichee Pi 4A.

A chip that's oddly faster than JH7110 (VisionFive2, Milk-V Mars) in microbenchmarks, but slower in practice (e.g. gcc). Presumably due to smaller cache.

It is also less power efficient, and lacks the upstream support JH7110 enjoys[0]. I would look at Pinetab-V, a tablet-laptop based on that SoC, today.

Better yet, wait for Milk-V Oasis[1] tba this June, as well as other boards based on SG2380, the first announced RISC-V SoC with serious performance: 16x SiFive P670 and 8x X280, all RVA22 compliant, plus vector 1.0 (standard) extension.

0. https://rvspace.org/en/project/JH7110_Upstream_Plan

1. https://community.milkv.io/t/introducing-the-milk-v-oasis-wi...

View on HN

I did a comparison of a few RISC-V single board computers and the Lichee Pi 4A is definitely very slow on real world tasks, even though in theory it ought to be the fastest. I never got to the bottom of exactly why - it might be software issues or non-upstream drivers. At the moment I'd recommend the VF2 for general use.

https://rwmj.wordpress.com/2023/07/25/heads-up-lichee-pi-4a-...

Interesting, I had the opposite experience: the GUI was very slow and nearly unusable on most RISC-V boards that I have, except the LicheePi4a. But that is probably just due to the GPU (which unfortunately doesn't have open source drivers yet?), and not the CPU, and they're all running different kernels and distros. When I have some time I'll need to compare again with the latest distro available for each, and also a fully open source one.

You have to keep in mind that since RISC-V is an ISA standard, C[gcc] coded kernels are here for legacy support purpose. The C[gcc] language is not appropriate anymore, assembly coded is the right(tm) way.

What is this supposed to mean? You think someone is going to rewrite linux in assembly language?

re: GPU, mesa3d side (vulkan driver) is merged upstream. OpenGL/CL/GLES are via Zink.

powervr is working into getting the drm side merged to the kernel, too.

Something's definitely "up with" the RISC-V GCC port; in principle maybe there are just some passes or backend options that need to be flipped on in some way?

https://godbolt.org/z/cvc6j1sM6

Looks like clang does the optimization for -mno-strict-align, but gcc doesn't seem to support that fully yet: https://godbolt.org/z/xn8erKMTe

Edit: Apparently gcc assumes by default, that unaligned loads aren't supported, but with -mtune=size it somehow enables it: https://godbolt.org/z/d9P19aMnn

some discussion around that: https://gcc.gnu.org/bugzilla/show_bug.cgi?format=multiple&id...

The SG2380 is a bit weird though. Four of the P670 cores are clocked lower than the others.

The X280 cores are not RVA22 compliant. They are only RV64GCV, so they can't run the same code as the P670 cores. SophGo's datasheet lists these cores under "AI accelerator", paired with a "TPU", so they were never intended to run general purpose code.

X280s where design for AI applications, and have an interface for custom vendor extension. If I understood it correctly SophGo plans to add a matmul extension to it.

You wouldn't be able to run P670s and X280s on and schedule applications between them anyways, because they have different vector lengths.

>You wouldn't be able to run P670s and X280s on and schedule applications between them anyways, because they have different vector lengths.

Not trivial, but also not impossible.

e.g. I understand there's some bits to check whether vector has been used. It would also be abnormal for a process to be in wait (as opposite to ready) within a vector loop.

Having a chip like this one deployed will definitely enable experimentation and research on how to best deal with this situation.

Well, I don't like the SGxxxx SOC as they contain an ARM core (= royalties), I would prefer a CVxxxx SOC without an ARM core, and if they could get in the future displayport instead of hdmi and AV1 instead of h26[45] (hdmi and mpeg are not worldwide royalty free).

I am trying to buy some milk-v duo boards with the CV1800 SOC in EU (without a credit card to use online, and which I can contact with a self-hosted email), anyone can help?

EDIT: the next SGxxxx SOC is actually a full desktop SOC, not an "embedded" one (I was talking about the those). No ARM cores, but still a mpeg and hdmi block, and it seems the GPU hardware programing manual is not public.

Damn, best products are always in the future :-) Oasis looks very capable and juicy, looking forward to get my hands on it.

Same. This seems to be the SoC and board I have been waiting a long time for.

Do not forget to preorder, to get the unbelievably low price listed in their announcement.

It should be fun to play with it this summer, and further down the line it will become my home server, replacing an old atom board that's considerably weaker.

That's mostly enabled by the builtin I/O (pcie slots, sata ports...).

+1 on your initial comment. Exactly how I feel about the current situation.

MilkV Oasis with SG2380 would be the end-game for majority of developers, but they are definitely loosing money if they keep starting price at 120 USD. They don't have it frozen (they changed the SoC specification some weeks ago) thus I wouldn't be surprised to see this slip into 2025. I wouldn't be surprised if this outperforms MilkV Pioneer.

They're counting on most purchasers not preordering, and thus not getting the preorder discount.

Which is a reasonable assumption to make. SG2380 will surprise many as it launches and gets reviewed and talked about.

I placed an order for 4 coupons :D

> Do not forget to preorder, to get the unbelievably low price listed in their announcement.

I don't know much about the RISC-V ecosystem at the moment, but what do you consider an "unbelievably low price"?

They SoC appears to be £5 to pre-order (EDIT: £5 is for a 20% off coupon), but the motherboard to use it on appears to be ~£1200. Is this considered a good deal for RISC-V right now?

> but the motherboard to use it on appears to be ~£1200

Not sure what you're looking at as you didn't give a link, but perhaps you're confusing Pioneer and Oasis?

The 64 core Pioneer prebuilt with case, power supply, SSD, video card, 128 GB RAM for $2500 is quite reasonably priced against commercially built 16 core x86 machines, if your workload can keep 64 cores busy. Each core is around 1/4 the speed of current x86, but there are four times as many.

Don't forget 64 core x86 is $5000 just for the chip.

I clicked through the GP's second link and the pre-order link which took me here: https://arace.tech/products/pre-order-milk-v-oasis-16-core-r...

This appears to be the (voucher for the) "chip" but I don't see any board to mount it on other than the Pioneer.

Again, I admit I know nothing about the RISC-V ecosystem so apologies, I'm just curious.

To clarify, Milk-V Oasis is the board, not the chip.

The chip is SG2380, and it comes soldered onto the board.

Thus the preorder is for the board, which includes the chip.

This is a Mini-ITX board, and you're meant to provide RAM, storage, power supply, case and whatever else needed.

The unbelievably low price price milk-v listed was $120. Sipeed mentioned something abour $300.

I presume that the milk-v one might be just the cpu+motherboard, and sipeed included ram and gpu in the price, but thats all speculation.

If it's anywhere in the $300 range I'd be very happy.

Just for scale the specint numbers from the P670 are on Cortex A76 level, so you get what the rp5 has, but you get way more cores.

Interesting, and what form factor is this in? Are we talking mATX/ITX style board as opposed to an SBC?

It says mini ITX: https://community.milkv.io/t/introducing-the-milk-v-oasis-wi...

That makes sense, and sounds like an excellent deal.

> Just for scale the specint numbers from the P670 are on Cortex A76 level,

A78 level, 12 SPECint2006/GHz.

It's P470 (Horse Creek, if it ever appears) and Dubhe-90 that are at A76 8 SPECInt2006/GHz level.

And, for completeness, P870 is 18 SPECInt2006/GHz.

My question is, will the future Milk-V Oasis laptop have a TrackPoint like the Lichee Console 4A has, or will greater market demand for a more powerful RISC-V laptop lead to nixing TrackPoint for cost optimization?

Reply on news.ycombinator.com