StarFive VisionFive 2 Quad-Core RISC-V Performance Benchmarks

[crashtan]

Btw if someone wants to test out some RISC-V boards, there's www.cloud-v.co, it allows you to set up a CI pipeline on real boards and some QEMU instances. Don't have physical access to hardware though.

[coder]

computational user space apps--which almost by definition interact very little with the kernel. If the goal is to get as much work done as possible, it's best not to waste any CPU time in the kernel, is it not?

Even a "computational" userspace app is still susceptible to the kernel's thread scheduling, frequency scaling, and probably its memory management.

We saw this with the last Asahi benchmarks on the M2, where supposedly the kernel didn't know how to schedule the big vs. little cores and I'm not sure whether its frequency scaling was optimal.

Worse than that, as you can see, most of the cores implemented so far that us little people can get our grubby little hands on aren't super high performance in nature.

Even with an unoptimized toolchain, the Lichee Pi 4A achieves higher DMIPS/MHz than the R.Pi 4's A72:

https://sipeed.com/licheepi4a

[coder]

Age is absolutely relevant because there are many newer/faster/better CPU generations available. CPUs use technology available at the time of their design. Process technology has improved quite dramatically since the first FinFET generation, enabling more complex designs, larger caches and lower power.

If the designs were both targeted at 28 nm manufacturing process, that effectively nullifies this point.

The difference between the cores in Pi 4 and Pi 5 is 3 years or 3 CPU generations - look at the huge speedup that gives!

A couple of things to say about this:

1. Don't shorten Orange Pi 5 to "Pi 5". You'll just confuse people. It has nothing to do with Raspberry Pi, other than ripping off the "Pi" part of the name.
2. Orange Pi 5 is made on a much better process node (Samsung 8nm, I believe), which is partially responsible for its improvements.

age matters as well. Nobody should be surprised that a modern RISC microcontroller outperforms say a 486.

Or that a Core 2 outperforms a Cortex-A53. So, you're overindexing the importance of age.

[coder]

If your complaint is that I'm out by a year or two on a 10-plus year range, I'd say that's the argument that's "just not that strong".

My complaint isn't limited to you. It's just that all the haters seem to pad their estimates in a favorable direction. Maybe you did so by mistake, but it can't be true for everyone.

Happily though it's irrelevant anyway, since https://en.wikichip.org/wiki/28_nm_lithography_process will tell you that commercial 28nm production did indeed start in 2011, making this one of the rare occasions that I actually remembered a date correctly.

Not according to the review I found of this Apple iPad from 2012. It said:

"Unfortunately, the new iPad’s release schedule meant that it missed the 28nm shipping window by a couple of months"

https://www.anandtech.com/show/5688/...-2012-review/2

https://www.aliexpress.us/item/3256804950133483.html (from Google just now) is an N95 NUC for $127 - that's with 8GB of RAM, a 128GB SSD, and a case with a heatsink and fan.

Okay, that's indeed a good deal. It feels like inventory liquidation, given that it's marked down from $239. China's economy is currently in a rough spot, so it's plausible they overproduced for their domestic market and are now dumping some inventory at a loss.

Given that the bare board in this article was $115 for an 8GB SoC with no storage, I'd say that easily puts the Alder-N board itself at barely more expensive, and quite possibly not even at all so.

Michael cited Amazon.com. If you're making cost arguments, you really should compare AliExpress prices only to other AliExpress prices, for a variety of reasons.

> Also, they come rated as low as 6 W, but the quad cores like to turbo up to about 25 W or more.

Again, not sure what the "they" is here - I'm guessing from the "25W" part that you mean the Intel cores, but nobody was talking about power consumption until you brought it into the topic just now.

I know, but if you're posing Alder-N as an alternative to other SBC's, then power consumption is something people should be aware of, especially given Intel's tendency to quote artificially low values, here.

Not that it's not a point worth raising, but given the earlier objections it looks more like moving the goalposts.

It's about trying to have an informed discussion of the tradeoffs. That's all it was. I mention it, because I was surprised to see just how much more than the rated power they could actually consume, even without hitting the iGPU.

These aren't SPARC/Alpha chips trouncing 486s: they're mostly replacing microcontrollers, and to move up from there they need to go through what's now ARM territory,

You should check out the Sipeed link I just posted. I think the Lichee Pi 4A is going to break some assumptions about just how quickly RISC-V is catching up to ARM. Hardware Unboxed already booted the thing and ran some Linux desktop apps on it

[PerformanceExpert]

If the designs were both targeted at 28 nm manufacturing process, that effectively nullifies this point.

The same process only implies similar cost/power/performance per transistor. A newer design always has an advantage over an older design. Even if the core and process are identical, performance varies depending on chosen cache size, memory bandwidth, frequency etc.

Orange Pi 5 is made on a much better process node (Samsung 8nm, I believe), which is partially responsible for its improvements

Orange Pi 5 it is!

Absolutely, that's the point I'm making about newer designs using better technology. And you only get these gains from a newer design (ie. porting Cortex-A72 to a new process won't give much speedup no matter how amazing the new process is).

Or that a Core 2 outperforms a Cortex-A53. So, you're overindexing the importance of age.

Well Cortex-A53 outperforms various low power Core 2 parts. So an in-order core can outperform an older out-of-order design - as long as it is old enough.

[PerformanceExpert]

Even with an unoptimized toolchain, the Lichee Pi 4A achieves higher DMIPS/MHz than the R.Pi 4's A72:

Dhrystone is regarded as a bad benchmark by pretty much everybody. Someone ran SPEC on it: 7.64 SPECint2006 at 1.85 GHz vs 9.22 for Pi 4 at 1.5 GHz. So on complex code Cortex-A72 has almost 50% higher IPC. Now it's possible there are issues with his board or setup that affect performance but the C910 was claimed to be competitive with Cortex-A73. So I don't see how you could say this shows "just how quickly RISC-V is catching up to ARM".

[Quackdoc]

Btw if someone wants to test out some RISC-V boards, there's www.cloud-v.co, it allows you to set up a CI pipeline on real boards and some QEMU instances. Don't have physical access to hardware though.

thanks for the tip

[PublicNuisance]

The board has both an eMMC socket (Odriod/Pine64 defacto standard) and NVMe via an M.2 slot.

What? VF2 has nvme 2240 slot.

My bad, must have missed it in the specifications list.

[arQon]

My complaint isn't limited to you. It's just that all the haters seem to pad their estimates in a favorable direction. Maybe you did so by mistake, but it can't be true for everyone.

Like I say, precision tends to decrease when you're talking about that sort of timeframe, so I wouldn't be too hard on people for it. The only reason I got it right off the top of my head is I remember building an IVB machine in 2013. 22nm 10 years ago meant 28 had to be at least 11 years, and Ivy had been around for over a year by the time I did my build, so I had a decent basis to guess from.

> Okay, that's indeed a good deal. It feels like inventory liquidation, given that it's marked down from $239. China's economy is currently in a rough spot, so it's plausible they overproduced for their domestic market and are now dumping some inventory at a loss.

Could well be - but prices are what prices are, regardless of reasons. You wouldn't have argued that e.g. a 3060 was "the best $300 card" two years ago when they cost $1000 or whatever if you could even get one at all, would you?

> Michael cited Amazon.com. If you're making cost arguments, you really should compare AliExpress prices only to other AliExpress prices, for a variety of reasons.

Fair, but that would have required more investment than I felt it was worth. AliExpress was first in the search results, so that's what I went with.

> I know, but if you're posing Alder-N as an alternative to other SBC's

I wasn't really - a NUC may have some key similarities, but they're still very different animals. I was just objecting to the claim that the S5 board is significantly cheaper and/or only slightly slower, when the pricing was surprisingly close and the performance difference is several multiples.

> then power consumption is something people should be aware of, especially given Intel's tendency to quote artificially low values, here.

Oh, very much so. You'll recall a conversation we had here not too long ago where I reminded you Intel is even worse about it than you think. (and seems to be continuing that trend, though it's worth mentioning that one of the Alder-N chips appears to be salvage part that runs at 10W rather than 7W, so that's nearly 50% over the nominal draw even before you get to the 3x or 4x multipliers on "under load" or "using the IGP").

I think it's fair to say that the S5 got crushed by the Pi4, and the gap from that to even the every bottom of the Intel range is so wide that you're probably looking at OoM differences, so it's no surprise that Alder-N is actually in a different power class, despite Intel's claims. Even so, if you were to set PL1/PL2 on the Alder parts to enforce the 7W limit, you'd have something at least 4x-10x faster.

>You should check out the Sipeed link I just posted. I think the Lichee Pi 4A is going to break some assumptions about just how quickly RISC-V is catching up to ARM. Hardware Unboxed already booted the thing and ran some Linux desktop apps on it

I've heard this song before though, and I'm fairly sure you have too. There's so much more to bringing up a new CPU than the HW design, and in fact that's that *quick* part. It's nice that there are R5 parts around to evaluate on peer process nodes, since e.g. 28nm vs 7nm is just too far apart to be able to make good estimates of relative performance, but it doesn't change the fact that R5 is still massively behind x86 on performance, and massively behind ARM on production.
So let's look at this from the other side: what is it that makes you believe R5 is going to be relevant in, say, the next decade, other than "Because it would be cool if it was"? Where's the actual innovation / improvement over ARM? (Which, you'll recall, is technically a RISC design itself - until the usual inevitabilities set in and it grew SIMD ops etc).

[coder]

Like I say, precision tends to decrease when you're talking about that sort of timeframe, so I wouldn't be too hard on people for it. The only reason I got it right off the top of my head is I remember building an IVB machine in 2013. 22nm 10 years ago meant 28 had to be at least 11 years,

No... one is Intel's process node, while the other is TSMC. They did not advance in lockstep.

Could well be - but prices are what prices are, regardless of reasons.

It depends on the point you're trying to make, with them. If you're saying the premise of this product is flawed, because right now it doesn't represent the best value out there, then you really need to look at why x86-based alternatives are selling for what they are. I think what you've found is a market anomaly, due largely to the historic dip in global PC demand and China's flagging economy. So, I wouldn't read a broader meaning into it.

Anyway, I think we'd all agree that this product doesn't offer a great value for someone who's not specifically wanting RISC-V. If you're comparing it to non- RISC-V products, then you've already missed the point.

it's no surprise that Alder-N is actually in a different power class,

Please compare it with a more contemporary ARM SoC, like one MediaTek Genio 1200. Then, I think you'd see that Alder Lake-N's power consumption still isn't entirely justified by its performance. If you turn down Alder-N's power limits to match the MediaTek SoC, then the latter will certainly pull ahead on performance.

I've heard this song before though, and I'm fairly sure you have too. There's so much more to bringing up a new CPU than the HW design, and in fact that's that *quick* part. It's nice that there are R5 parts around to evaluate on peer process nodes, since e.g. 28nm vs 7nm is just too far apart to be able to make good estimates of relative performance, but it doesn't change the fact that R5 is still massively behind x86 on performance, and massively behind ARM on production.

WTF? I didn't say the Lichee Pi 4A was faster than Alder-N or ARM. My point was that it's a big leap beyond the VisionFive 2 and already boots & runs desktop Linux! All I claimed was it's "going to break some assumptions about just how quickly RISC-V is catching up to ARM". In other words, the gap isn't as big as you seem to think it is.