StarFive VisionFive 2 Quad-Core RISC-V Performance Benchmarks

[brucehoult]

Noob question: what makes it slow at the end? Lack of software optimisation, fabrication node (thermal/frequency limit) or general state of development of the risc-v platform?

What makes you say it is slow?

It is precisely the speed you would expect it to be, given its dual-issue in-order micro-architectural design -- similar to ARM's A53 and A55 -- and that it doesn't have SIMD/Vector because 1) not all computing requires that, and 2) the RISC-V Vector spec was not completed until three years after the U74 core was announced.

It is what it is, and it does what you'd expect. You don't expect an A53 to be performance competitive with A72 and A76, so why is this being benchmarked against A72 and A76 instead of against A53s and A55s e.g. Pi 3, Rock64, Odroid C2 or C4 etc.

Absolute performance isn't everything. These in-order dual-issue cores use less electricity and produce less heat per unit of work done. They are commonly used as "efficiency" cores in chips that also have some faster cores. But you can also still buy low end phones and tablets with four A53 cores and nothing else. This chip sips so little power that even running all four cores at maximum it doesn't get anywhere near a temperature that would cause throttling (at least in normal home/office environment). It needs no heatsink or fan. Unlike, say, the Pi 4, which doesn't come with cooling and throttles heavily under load if you don't add any.

[Almindor]

What's worse than no benchmark? A bad one. It's really disappointing that Michael let this slide using ancient kernel. It's useless to the point of being damaging.

[Michael]

What's worse than no benchmark? A bad one. It's really disappointing that Michael let this slide using ancient kernel. It's useless to the point of being damaging.

It's the kernel image that is the default for the VisionFive 2 default OS image... If a newer version of the StarFive patched kernel was in better shape, I am sure StarFive would have switched to it as their default kernel by now.

[coder]

This chip sips so little power that even running all four cores at maximum it doesn't get anywhere near a temperature that would cause throttling (at least in normal home/office environment). It needs no heatsink or fan. Unlike, say, the Pi 4, which doesn't come with cooling and throttles heavily under load if you don't add any.

Wow! My Pi 3 easily throttles on all-core workloads, even after I installed the best copper heatsink I could find, using high-performance heatsink compound.

Do you have any idea what manufacturing process node the StarFive JH7110 SoC uses? I searched, but was unable to find this information. I know the Pi 3 used 40 nm.

[ayumu]

Do you have any idea what manufacturing process node the StarFive JH7110 SoC uses? I searched, but was unable to find this information. I know the Pi 3 used 40 nm.

According to brucehoult elsewhere:

https://www.reddit.com/r/RISCV/comments/15svco6/starfive_visionfive_2_quadcore_riscv_performance/jwigmrw/

It seems to be 28nm. Not sure what his source is.

[brucehoult]

According to brucehoult elsewhere:

https://www.reddit.com/r/RISCV/comments/15svco6/starfive_visionfive_2_quadcore_riscv_performance/jwigmrw/

It seems to be 28nm. Not sure what his source is.

I dunno. Lots of articles say so, presumably based on a press release:

https://forum.banana-pi.org/t/starfive-announced-2-high-performance-risc-v-products-jh7110-soc-and-visionfive-2-sbc/13820

Regarding thermals .. I was reading the datasheet a year ago [1] and noted:

The most interesting thing I've noted so far is 125 ºC maximum junction temperature, and 8.1 ºC/W θⱼₐ and 5 W power rating. Together the last two mean at maximum power the junction temperature is 40.5 ºC above ambient temperature, or 60-70 ºC in a normal environment. Without a heatsink. So, cool enough.

My experience with actual boards confirms this.

[1] https://www.reddit.com/r/RISCV/comme...mment/in98v6s/

[arQon]

Wow. Pitiful performance. Hopefully they're better in 3-5 years.

"Better in 3-5 years" is inevitable, but "adequately better" is a lot further off than that. The Pi4 that's absolutely destroying this thing (wow, that's a sentence I never thought I'd write!) is a 6 year old chip on a 12 year old process - it's not like it's cutting edge tech. An Alder-N Atom is about the same speedup again, and only slightly more expensive, so it's got an even longer way to go than it looks from just this article.

Since RISC-V only has a fraction of the market ARM had a decade ago, and the TAM for anything other than controller use is almost non-existent, progress is likely to be even slower than that was. You'd need a sugar daddy with FAANG money to burn for it to catch up to even just current ARM in 10 years, let alone where ARM will be by then.

This board in particular is disappointing though: there are a lot of use cases for a dual-NIC box, but atm I'm not sure I'd put money on it even just being able to drive both of them at line rate, let alone so anything with the data. I've worked with old ARM chips in network gear, and even just the TCP checksum work can eat 100% of a core in high traffic. (Hell, I remember it taking 30-40% of a Xeon when comparing).

By all means be enthusiastic - but I think you're gonna want to temper your expectations.

[ayumu]

My experience with actual boards confirms this.

I still have a (passive) heatsink on mine, because I'd rather the heatsink get warm than the PCB.

The other components (like the SSD below) might not be built to the same industrial temperature standards.

CPU seldom touches 60C even when compiling for hours at ~30C ambient temp.

[coder]

The Pi4 that's absolutely destroying this thing (wow, that's a sentence I never thought I'd write!) is a 6 year old chip on a 12 year old process - it's not like it's cutting edge tech.

Why do the naysayers always seem to exaggerate? If your argument isn't strong enough to sell without exaggerations, then maybe it's just not that strong?

The Pi 4 launched just over 4 years ago. The process node it used is just over 10 years old.

An Alder-N Atom is about the same speedup again, and only slightly more expensive,

Its performance is Skylake-caliber, but show us a board that's only slightly more expensive. Also, they come rated as low as 6 W, but the quad cores like to turbo up to about 25 W or more.

But, this board was never about beating x86. It's just a low-power tinker board, and a mere stepping stone down the long, winding path towards world-domination (cue evil laughter).

By all means be enthusiastic - but I think you're gonna want to temper your expectations.

I think you're in for a few surprises.

[CochainComplex]

What makes you say it is slow?

It is precisely the speed you would expect it to be, given its dual-issue in-order micro-architectural design -- similar to ARM's A53 and A55 -- and that it doesn't have SIMD/Vector because 1) not all computing requires that, and 2) the RISC-V Vector spec was not completed until three years after the U74 core was announced.

It is what it is, and it does what you'd expect. You don't expect an A53 to be performance competitive with A72 and A76, so why is this being benchmarked against A72 and A76 instead of against A53s and A55s e.g. Pi 3, Rock64, Odroid C2 or C4 etc.

Absolute performance isn't everything. These in-order dual-issue cores use less electricity and produce less heat per unit of work done. They are commonly used as "efficiency" cores in chips that also have some faster cores. But you can also still buy low end phones and tablets with four A53 cores and nothing else. This chip sips so little power that even running all four cores at maximum it doesn't get anywhere near a temperature that would cause throttling (at least in normal home/office environment). It needs no heatsink or fan. Unlike, say, the Pi 4, which doesn't come with cooling and throttles heavily under load if you don't add any.

I opened my question with the "noob" question because I have no clue about riscv arch. So this is by no means meant to be an insulting question. I was saying slow because I was comparing the benchmark and the price point with the given other options.