Scaleway RISC-V Testing (EM-RV1-C4M16S128-A)

182

Scaleway is a French cloud/server provider with a long history of providing new, and fun options in their Scaleway Labs section. They’ve long had ARM-based instances available, but what they’ve not done before is enter the RISC-V market. This changed today as they introduced the Scaleway RISC-V Elastic Metal servers, and it’s those that I’m going to have a quick look at today.

Based on the Alibaba T-Head TH1520 SoC (the same used in the Sipeed LicheePi 4A, amongst others), it’s a quad-core Xuantie C910 processor running at 1.85GHz. Scaleway is pushing the limit of the SoC and includes 16GB of RAM, along with 128GB of eMMC storage. Whilst in the Labs section, it “only” comes with a 100Mbit network connection, though both IPv4 and IPv6 are available.

At a monthly price of €15.99, or €0.042 hourly (both excluding VAT), it’s an affordable way to get your hands on a dedicated RISC-V system to play around and see how things are progressing. Let it be known though, that there is no SLA on these machines. That doesn’t bother me though, as long as it stays online for my brief testing!

Scaleway RISC-V: Initial Setup

Scaleway offers 3 distributions at the time of my testing, with Ubuntu 23.10, Debian Unstable, and Alpine Linux Edge available for installation. I chose Ubuntu in this case and that came as Ubuntu 23.10 (GNU/Linux 5.10.113+ riscv64) once deployed.

Beyond that, there’s not too much to it. You’ll need to create an SSH key before you’re able to deploy, but that’s the most complicated part of it!

Benchmarking the RISC-V TH1520

A lot of the usual suspects I’d use for benchmarking on ARM-based SBCs are not yet available for RISC-V, but we do have a couple available in the form of UnixBench and Geekbench for compute/memory testing and they’re a straightforward enough comparison for most readers without diving into hundreds of specific benchmarks.

The usual disclaimer applies here, though, as the operating systems used differ, therefore there is almost definitely a difference in the underlying software being used and this is why technically, these are more system benchmarks than CPU or memory benchmarks. The data from the VisionFive 2 that I’m comparing here comes from a previous benchmark run on the same benchmark versions. This isn’t meant to be an incredibly scientific comparison, it’s more a glance at the numbers and verifying some that Scaleway provided themselves.

UnixBench (5.1.3 riscv64)

RISC-V support in UnixBench has been with us for a while now and I’ve previously run RISC-V UnixBench tests so it’s good to get some more data.

We’re comparing against the VisionFive 2 here (which runs a StarFive JH7100 SoC at 1.5GHz) as it’s the closest thing I’ve laid my hands on (I have a LicheePi 4A but time is scarce..) but surprisingly, in UnixBench, the TH1520 falls behind on the overall System Benchmarks Index Score.

That doesn’t tell the whole story, though. RISC-V is still catching up and the difference between different SoCs and core designs is vast, even for chips released around the same time. In quite a few tests, the Scaleway Labs instance is more than twice as fast, and in others, it’s 50% slower. I’ve included a more detailed graph with the individual test results below to highlight this.

	VisionFive 2 (Single-Core)	Scaleway Labs (Single-Core)	VisionFive 2 (Multi-Core)	Scaleway Labs (Multi-Core)
Dhrystone 2 using register variables	547	1186	2191	4718
Double-Precision Whetstone	263	486	1051	1938
Execl Throughput	453	243	1708	763
File Copy 1024 bufsize 2000	468	387	1777	1407
File Copy 256 bufsize 500	395	260	1482	985
File Copy 4096 bufsize 8000	584	793	1868	2449
Pipe Throughput	266	299	1059	1198
Pipe-based Context Switching	112	73	465	581
Process Creation	171	95	1367	451
Shell Scripts (1 concurrent)	667	440	2575	1765
Shell Scripts (8 concurrent)	2308	1576	2377	1717
System Call Overhead	518	336	2094	1342
System Benchmarks Index Score	418	364	1532	1329

Geekbench 6 (6.2.2-LinuxRISCVPreview)

Scaleway tested this themselves and got a score of 380 on the multi-core side of things, though they didn’t post single-core results so I’ve run it myself and the results are interesting

Their own results graph has the VisionFive 2 for comparison and they claim their multi-core score was only 148, whereas in my VisionFive 2 Review it clocked in with 222. Sure, it’s still not mind-blowing but the increase from 148 to 222 is 50% so it’s quite a shift. This can likely be explained by a difference in system setups but without knowing exactly how they ran their test, or where they took their numbers from, it’s hard to say.

eMMC

To round off the testing we’ll see how the eMMC module performs.

It’s identified as a Y2P128 model chip from “0x00009b” which is an SK Hynix chip and interestingly, is the same chip found in the Lichee Pi 4A. Hmm, are they just Sipeed’s SoM after all?

Block Size (KB)	Sequential Read (MB/s)	Sequential Write (MB/s)	Random Read (MB/s)	Random Write (MB/s)
4	30.41	27.09	16.36	27.09
16	81.64	64.54	43.32	66.44
512	253.77	194.77	197.49	178.78
1024	268.04	204.86	214.86	200.22
16384	260.42	221.42	254.97	216.57

Whatever they are, they provide OK performance. It’s noticeably slower than a traditional SSD/NVMe drive and even at a 4KB block size, it’s around that of some microSD cards that I’ve tested. At larger block sizes, however, we do see it pulling ahead of what a microSD card would be able to offer.

General Testing Thoughts

It’s a little unknown as to what kind of setup Scaleway has (I’ve asked for more information and photos, I’ll add them if they oblige) but we do get access to temperature sensors that were reporting around 50 degrees Celsius under a single core benchmark, and up to around 75 degrees Celsius on all-core workloads. From what I could see though, this didn’t result in any throttling, and performance remained consistent throughout. At idle, it hovers around 41 degrees Celsius.

Whilst not particularly important, the 100Mbps network connection makes sense in the Labs, though I wonder if they’ll introduce 1Gbit further down the line. I’m really curious as to what they’re using here! Are they just LicheePi 4A SoM (System on Modules) in some kind of custom cluster board? Are they of their design? Please text back and let me know, Scaleway, your product page mentions 3D Printing and artisanal approaches but I want to know more ☹️

Conclusion

Whilst I’m not sure most of my usual readers (shout out to all 5 of you!) will be swarming to get one of these instances themselves, it does show that RISC-V is on more people’s radars these days, and popularity/demand is growing.

When looking for RISC-V compatible benchmarks, I often saw the developers saying they simply didn’t have access to hardware, and whilst the SBC market for RISC-V has boomed, having hourly instances like this may prove very useful in having more developers support the RISC-V/riscv64 architecture going forward.

Maybe I should dust off the LicheePi 4A and see what the rest of it is like, I’m feeling all interested in RISC-V again.

UPDATE: Sipeed LicheePi Cluster Confirmed?

A fellow Twitter user kindly linked me to a tweet from a staff member at Scaleway that shows their RISC-V servers in the rack and yup, I think we can now confirm that they’re just running Sipeed’s LicheePi 4A SoMs in a cluster.

Rackage à DC5 terminé. Câblage #Atari. Les DEL sont cramptés. Une belle densité de 672 serveurs par rack. La date de lancement est fixée. On peut partir en vacances. cc @Scaleway_fr #dedibox #baremetal pic.twitter.com/yQGIJ3C7sG

— Seblu (@seblu84) February 16, 2024

If you look at the I/O layout, you’ll see it matches that on the diagram below, which is of Sipeed’s cluster board for the Lichee Pi 4A modules. Case closed!

Header photo credit: @seblu84 on Twitter/X