StarFive VisionFive 2

>don't know how slow it is in comparison to even the cheapest Arm SBCs right now. I own a VisionFive 2 with this SoC. CPU sits between rpi3 and rpi4 performance. GPU is easily 3x that of rpi4. I/O speeds are similar to rpi5. This is not anywhere as slow as you're making it to be. And certainly it isn't slower than the cheapest ARM SBCs. But yes, you can get a much faster rpi5 and it is cheaper. Current RISC-V SBCs and laptops are for enthusiasts and developers, who know full well what they're getting. Next iteration will be faster. The next, faster still.

This is JH7110, a chip from 2 years ago. It was first seen in VisionFive 2, the first mass-produced RISC-V SBC. It's also available in other boards, such as MILK-V Mars, Star64 and the Roma laptop. A bit old now, but has the advantage of having good upstream support, which is likely why its SoC was selected. I use my VF2 as a small home server, with ZFS disks. It runs Debian trixie with Debian's generic RISC-V kernel package.

I use this on VisionFive 2. Works well (seems no worse than archlinuxarm), and has most Arch packages. I am hopeful as the arch ports RFC[0] passes, RISC-V and other ports will be taken more seriously, and hopefully eventually become fully supported. 0. https://gitlab.archlinux.org/archlinux/rfcs/-/merge_requests...

It's not a "metric" it's real-world use. Whatever the causes, the CPU core is not the only component in an overall system. I use both of these boards every day. Constantly, as it's my work. Building my RVV 0.7.1 gcc 9.2 snapshot takes 111 min on the U74 VisionFive 2, 122 minutes on the C910 LicheePi 4A. Running DotNET "LINQ" test suite takes 4m30s on VisionFive 2, 5m30s on LicheePi 4A. Doing `emacs --eval '(kill-emacs)' hello.c` takes 0.72s on VisionFive 2, 0.75s on LicheePi 4A. Close, but the C910 does not win. Yes, the C910 significantly beats the U74 on every micro-benchmark. Dhrystone, coremark, my "primes" test https://hoult.org/primes.txt , memcpy... https://hoult.org/TH1520_memcpy.txt https://hoult.org/JH7110_memcpy.txt It doesn't translate to the real world, at least in the TH1520 SoC, on the LicheePi 4A sbc.

> It still is imho the better RVA20/RV64GC SBC available Agreed. Certainly the safest bet. The Pine64 Star64 or Milk-V Mars might be as good or better (and Mars is a little cheaper, but also missing features), but the VisionFive 2 has the most development and support. > other chips might be faster somewhat but they use more power and need active cooling TH1520 is a bit disappointing. It benchmarks faster on Dhrystone, Coremark, my primes benchmark etc, but on real world work it is often slower. I've found it about 10% slower building the GNU toolchain. Richard Jones has noted the same thing building Fedora packages. On the Dotnet "LINQ" tests the vf2 takes 4.5 minutes while the LPi4A takes 5.5 minutes. The vf2 has NVMe (which I'm not using). It has faster I/O to the same external Samsung 2 TB USB3 SSD which I use on both boards. The 2 MB of L2 cache on the vf2's JH7110 vs the 1 MB on the TH1520 probably also contributes to real-world performance. > I am very hopeful for Milk-V Oasis Yup. I don't believe the promised timeline, but SOPHGO have proven they can make a decent SoC, the Oasis will be much cheaper than the Pioneer, and 16x Arm A78 equiv is much more useful for most people most of the time than 64x A72 equiv. That's 13 SPECInt2006/GHz vs 6, per core, so significantly better on anything using up to 16 threads, and no worse for anything that uses no more than 32 threads. The SiFive cores will also be a lot more standards-compliant.

I got my VisionFive 2 late January last year. It is a great board, which stays cool (~60C) with just passive cooling (XU4Q's is compatible). It still is imho the better RVA20/RV64GC SBC available; other chips might be faster somewhat but they use more power and need active cooling. Upstream support (linux kernel, uboot, opensbi) is very good[0]. I am very hopeful for Milk-V Oasis[1] (SiFive 16x P670 + 8x X280) and other new boards with faster CPUs and full RVA22 compliance plus 1.0 Vector extension. 0. https://rvspace.org/en/project/JH7110_Upstream_Plan 1. https://community.milkv.io/t/introducing-the-milk-v-oasis-wi...

What I like about the VisionFive 2 is that the Ubuntu server image works great: https://wiki.ubuntu.com/RISC-V/StarFive%20VisionFive%202 I had to use UART to install the image to SSD but it works quite well and I am happy with it.

By all means, use a good power source. Yet note that a good power source doesn't mean that these SBCs will suddenly draw 15W sustained. Average draw will be much lower than any of these NUC-like devices. e.g. my VisionFive 2 is below 4w.

Any board based on StarFive's JH7110 is currently best in this regard, I think. Datasheet & reference manual for this SoC is available. Especially their VisionFive 2 board. I've even downloaded a schematic for it (although older revision than actual board I have). And they're pretty good about upstreaming drivers. That said: what you probably care about is documentation for integrated peripherals (esp. GPU), and existence of open source, mature drivers for those. RPi is very good in this regard. Afaik the only binary blobs there is some GPU/SoC firmware, and (maybe) some boot code. RPi's in general are very well supported & documented, and its software ecosystem is very mature compared to anything RISC-V based. Could you pinpoint what you think is lacking there? Other ARM based boards may offer more bang/$. Likely at the cost of documentation or driver support (Beagleboard being an exception).

>Since when is x86 legacy? Since the first commercial RISC chips obsoleted x86 in the 80s. >Have you actually used one of those? Yes, JH7110 on VisionFive 2. >They are excruciatingly slow. Less performance than most Raspberry Pi. Not my experience, owning VisionFive 2 and all raspberry pi B boards (1,2,3,3+ and 4), all of which have been used heavily. The CPU is much faster than RPi3b+'s while using way less power. Only a little slower than RPi4, and this is assuming the RPi4 has a huge heatsink like mine does. JH7110 doesn't require a heatsink nor throttle under load, for it stays under 75C even when running builds 24/7, yet is designed to survive 125C. Whereas everything else in the SoC is better. Especially the GPU, which in the RPi4 is anemic, and can barely keep up even for 2d at 1080p. The experience is much better on VisionFive 2 for even simple tasks like web browsing, and it's not even close. Night and day. Nevermind the upstream support effort[0] getting this far in just a few months (indiegogo delivery was in March), and follows standard RISC-V boot/firmware interface specifications, whereas Raspberry Pi is all bespoke and never cared much about upstreaming support. >Also competitive ARM based alternatives (such as Snapdragon) require many closed source blobs to function. Here I just do not understand what point you're trying to make. 0. https://rvspace.org/en/project/JH7110_Upstream_Plan

I've also found the 1.5 GHz in-order VF2 does remarkably well vs the 1.85 GHz OoO LPi4A on software build tasks, though not as extreme as Richard shows. I'm lazy and using the original Image-55 on my 8 GB VF2, and the Debian that came preloaded in the eMMC on the LPi4A. My mass-production LPi4A arrived yesterday, I haven't tried it yet, tests are on the beta board that arrived a couple of months ago. On pure CPU core + L1 cache tests (e.g. https://hoult.org/primes.txt ) the LPi4A is considerably faster. The LPi4A is also much faster on memcpy tests. https://hoult.org/TH1520_memcpy.txt https://hoult.org/JH7110_memcpy.txt My build test is an RVV 0.7.1-enabled snapshot of the gnu toolchain (gcc 9.2) that I use on the TH1520 and SG2042. Newlib, non-multilib (just rv64gcv) build. I used the same Samsung 2 TB external USB3 SSD drive for src/build trees on both boards. https://github.com/brucehoult/riscv-gnu-toolchain VF2: real 107m52.116s user 348m49.800s sys 16m27.529s LPi4A: real 122m16.099s user 400m17.737s sys 23m8.567s Neither board is thermal throttling at all. I have the supplied fan sitting loosely on top of the LPi4A CPU (I also try tests without the fan ... it throttles a lot, taking 166m on this test). The VF2 is without a fan -- doesn't need it. The LPi4A has more L1 cache, but only 1 MB of L2 vs 2 MB on the VF2. But the LPi4A's RAM is twice as fast as the VF2's L2 cache! At least for streaming -- I haven't t…

I would look at JH7110, released earlier this year, used in VisionFive 2, Star64, as well as PineTab-V, and with an excellent official effort towards upstreaming software support[0]. I own a VisionFive 2 and can recommend it. There's also the TH1520, a newer chip that's a bit faster and has pre-spec V extension. This is available in the Sipeed Lichee Pi4A, but the support for this chip is very green still. CPU-wise, JH7110 performs between rpi3 and rpi4. TH1520 is faster than rpi4. Else (GPU, I/O, cryptographic acceleration, hardware video codec acceleration and other peripherals), they are both significantly faster than rpi4. 0. https://rvspace.org/en/project/JH7110_Upstream_Plan

My experience with VF2 has been great. I love the standard RISC-V boot process. My board runs Arch from NVMe. The weakest link seems to be the GPU driver (Imagination Technologies), as it is proprietary and not very good. We know they are working on a mesa3d driver, but I am not sure it will cover the GPU variant used in the VF2 SoC.

Today, I am only looking forward, with RISC-V architecture. ARM SBCs are the past. I love my VisionFive 2. It is no Orange Pi 5, but it is fast enough, and it is RISC-V proper.

This is a decent review of VisionFive 2 as it is right now (early days). I noticed a few mistakes/omissions. From the top of my mind: - There's mention of bad cryptography performance. But there is a cryptography acceleration engine. This is not enabled in the current test kernels, but yet patch is being reviewed by Linux upstream. In contrast, the Broadcom SoCs in the other boards VF2 was compared against have no such hardware. - There's mention of unusably slow youtube, but again, there's a hardware video decode block that's still not usable, yet it is considerably more capable than the one in the Broadcom SoCs. - Related, Firefox is very slow because JS is interpreted. A RISC-V JIT implementation landed recently and will be present in the 111 release due in a few weeks. - The SoC GPU is glossed over, but it did deserve more attention. Claimed to be 4x the performance of RPi4's SoC, Imagination Technologies is working on an open driver. This is a public effort as of early 2022, with some code already in Mesa3d upstream. I am happy that the author did stress that this is early days, and that if the review/benchmarks were to be repeated in the future, the results would be very different. Get this board (I love mine) if you're curious about and want to play with RISC-V. There's currently nothing faster than the VF2 available at any price, and yet VF2 is available at ~$100.

The modern equivalent to that experience is through cheap microcontroller development boards. I recommend the RISC-V based ones (e.g. ESP32-C6, CH32V, GD32V, K210). Some SBCs double as development boards, with huge GPIOs capable of a lot besides basic I/O (PWM, I2C, SPI, UART, I2S...) The new, RISC-V based, "VisionFive 2" is a lot of fun.

It is not! That is super misleading. For normal program code it is closer to a Pi 4 than to a Pi 3, similar to the all the very popular A55s board that have come out more recently than the Pi 4. The only way it is slower than a Pi 3 is if the Pi 3 program is using SIMD (Neon), which the VisionFive 2 lacks. The worst part of this Pi 3 comparison is that the Pi 3 has only 512MB or 1GB of RAM, which is extremely limiting in the modern world. This RISC-V board comes with a *minimum* of 2GB and is available with 8GB for $37. The RAM difference alone makes many things possible that are impossible on a Pi 3, and many other things much faster, regardless of the raw CPU speed. And then you have the M.2 NVMe SSD, something that neither the Pi 3 nor Pi 4 support, which again makes a whole raft of things much faster, even if the single lane means it can "only" do something near 400 MB/s (vs SD cards at 40 or 80 MB/s)

You're not giving us much to go on! What, exactly, do you want to do with RISC-V? Depending on your answer, my answer could be anywhere from a $1.50 CH32V003 board (and $3.50 USB/UART adapter) with 48 MHz CPU, 2 KB RAM and 16 KB flash for your program to a $2500 2 GHz 64 core 128 GB RAM Milk-V Pioneer. In between, sensible choices include Linux-capable boards including: - Milk-V Duo, 1 GHz CPU, $9 for 64 MB RAM, $13 for 256M - anything ranging from the $34 Milk-V Mars CM, the Mars SBC, Pine64 Star64, VisionFive 2, and on up to the $180 Sipeed Lichee Pi 4A with 16 GB RAM. They are all essentially the same production quality and processing power, varying only in RAM (2, 4, 8, 16 GB), WIFI or Ethernet and how many, eMMC or M.2 NVMe options etc. They all take SD cards if that's what you want. Also the CanMV-K230, which has the same speed core, but only 1 of them not 4, and only 512 MB RAM. But it is the only board right now with RISC-V Vector extension 1.0. You get what you pay for. It's also possible that "Install riscv-gnu-toolchain and QEMU on the computer you already have" is the right answer for you. Or "Run RISC-V Ubuntu (or other distro) in docker (which contains QEMU)". e.g. bruce@rip:~$ docker run --platform linux/riscv64 -it drujensen/riscv-ubuntu bash root@8db6fe081ba0:~# cat >hello.c #include int main(){printf("Hello World!\n"); return 0;} root@8db6fe081ba0:~# gcc -O hello.c -o hello root@8db6fe081ba0:~# ./hello Hello World! root@8d…

VisionFive 2 is "the one" as far as I'm concerned. Very cheap and crucially it has a slot on the back for an NVMe drive.

The biggest misstake is to not upgrade the GPU further. The competition has always been the Jetson Nano. Vision Five 2 is now my favourite passively coolable SBC.

"Buy Now" -> "Coming soon" God damn it. As far as boards you can actually get your hands on I think the VisionFive2 is the most interesting. Wi-Fi 6, PCIe M.2 slot, a GPU, 2 Ethernet. Not sure what else has become actually attainable lately though.

At least 4 companies are working on server-class chips, targeting "Xeon-like" performance. Of course without actual hardware to test you should take those claims with a grain of salt for the moment. If you want a cheap, low-end single board computer you can use right now, then I'd go for either the Vision Five 2 or the Sipeed Lichee Pi 4A. Performance won't be that great, but it'll run Fedora, Debian, Ubuntu or Yocto distros pretty much out of the box. I'm working on a Hifive Unmatched vs VF2 vs Sipeed vs Raspberry Pi 4B performance comparison right now, which will appear on https://rwmj.wordpress.com

I'll be sad if the Raspberry Pi 5 is anything else than RISC-V. In part because 5 is the perfect number for this, but mostly because it would mean that the foundation has already lost touch with reality and the direction the industry is going, and is drifting into irrelevance. SBCs that are compliant with RISC-V specs, have good driver upstreaming efforts and are highly competitive with the Raspberry Pi series such as the VisionFive 2 and Star64 have already hit the market. They aren't as fast as the rpi4 in the CPU side, but the CPU is still faster than the 3b+'s and everything else in the SoC including the GPU is much better, with higher power efficiency and using standard usb-pd power supplies. Nevermind the fact they use standard RISC-V boot process rather than bespoke garbage. Whereas the Sipeed LM4A TBA this month has a RISC-V CPU that is significantly faster than the one in rpi4 and 400 and is still competitively priced with options under $100.

>the increased cost of that chip that even with 2 additional CPU cores doesn't quite beat the Pi 4 The 2 additional CPU cores throw me off. I don't know what chip you're talking about there. >Not just because RISC-V's CPU is still not all there (but could be in another generation or two), If you're thinking performance, Ventana's Veyron should be available in 2023H2 in server form. While we don't know how competitive that will be, there's also Tenstorrent Ascalon (Jim Keller's team). That's releasing next year (2024), with performance competitive with (projected) Zen5. There's no gap left to close at that point. Of course, if the target is just a "Pi 4", the LM4A that's releasing the present month does already significantly outperform it. >The Raspberry Pi 4 fixed that problem in an unannounced hardware revision years ago. Revision 1.2 in 2020. AIUI, what they fixed is a spec violation. They still require 5v at very high current. This is e.g. in contrast to the VisionFive2, which will take any usb-pd phone charger and will negotiate 12v, yet will also work with one of these 5v rpi supplies. >Also, be careful what you wish for. RISC-V, despite all the hype, is just an instruction set. A high-quality RISC ISA you don't need to pay a license to use, and which has put serious efforts in having high-quality boot and platform specs in place before significant hardware availability, thus much better positioned than ARM in this regard.

VisionFive 2 (same SoC, has been available for months) has a NVMe slot instead of the PCIe slot. It's still possible (with an adapter) to use PCIe cards. It is already possible (with current u-boot patchset) to boot from NVMe. It has been possible for months to have root on NVMe. It is indeed much faster than your average SD card.