Title says it all. I’m wondering if anybody has or can provide a performance comparison between these two newest boards please? I’m sure the community would be interested to see…
From the launch of the A73 cores:
Tests carried out by BBench (Website loading benchmark): ARM Cortex-A73 claimed to be up to 10 percent faster than the A72 Tests carried out by SIMD/NEON (FFMEG Codec) workloads are supposed to improve by up to 5 percent Memory sees the largest gains with up to 15% improvement (JMC Stream Copy 64b) A 25 percent power reduction in integer workloads is expected
So basically slightly faster in benchmarks/vectorization, and more efficient. Core to core, barely the same, but it will be going from 2xA72 + 4xA53 to 4xA73 + 2xA53 cores, hence doubling the high performance cores and reducing the power efficient ones. I really don’t know about the GPU but obviously it’s expected to be better.
It’s all about use cases (as always). Both RK3399 and A311D are big.LITTLE implementations so contain both slow (efficient) and fast cores. As such asides synthetic benchmarks it really matters what your application is doing.
If you’re not doing number crunching (or running synthetic benchmarks in multi-threaded mode) then usually single-threaded performance is more important than multi-threaded.
Single-threaded means the application in question runs mostly one one single CPU core while multi-theaded means that all cores are (fully) involved.
Based on previous benchmarks the comparison between RK3399 and S922X (the similar Amlogic SoC on ODROID N2) looks as following:
- VIM3 will be +25% faster than Edge(-V) with multi-threaded workloads (probably +30%)
- VIM3 will be slightly faster than Edge(-V) with single-threaded workloads
Please note that these are only assumptions since ODROID N2 uses an S922X-A SoC with fast cores limited to 1.8GHz while Khadas now uses with the A311D an S922X-B variant able to clock up to 2.2 GHz. Also Khadas chose to allow Edge(-V) only to clock up to 1.8/1.4GHz while majority of other RK3399 boards in my list are allowed to use 2.0/1.5GHz.
It would be great if Khadas staff could give
sbc-bench -c a try. I guess some performance comparison numbers with ODROID N2 at launch day might even trigger CNX and others to give VIM3 greater coverage.
The above leaves both GPU and NPU aside and again: ‘use case matters’ whether you can take advantage of these separate engines or not.
Thanks for the excellent and detailed response. This is dissapointing for me though as I’ve just received my Edge-v… I ordered it before the VIM-3 was announced and it looks as though the VIM-3 may have 30% extra grunt… and it will drop straight into existing cases without the need for alterations …
Well, the lesson I learned is ‘buy IT stuff when you need it’. There will always be something faster, better, less expensive just next month so when waiting for the best opportunity you will have to wait indefinitely.
Asides that what’s your use case? If it’s for example ‘light Linux desktop with rootfs on SSD’ then this potential 30% higher multi-threaded performance of VIM3 is meaningless but way faster NVMe SSD access with Edge(-V) might result in an overall faster desktop experience. Also GPU/VPU acceleration do matter here a lot and as far as I know with RK3399 we’re already there but no idea about A311D (rule of thumb: the older the SoC the better the software support)
Staring at numbers or graphs made with kitchen-sink benchmarks IMO is rather meaningless unless use cases are considered. And also all the relevant stuff that happens aside ‘raw CPU horsepower’ (that’s a lot especially on SBC where drivers and settings play an important role).
A lot depends on the specific use of the devices. I didn’t compare EDGE to VIM3, but I compared it to N2. In a number of use cases, EDGE wins against S922. IMHO all artificial test, it’s a game marketers who show some conventional figures, which can vary from “small nuances” (conditions).
My use case is simple really, just as an Android box for entertainment purposes. I do usually multi-task though. Streaming, VPN and torrent running for a lot of the day. I’m a bit fussy, I really like to know I’ve got all the power I need and more & in a small form factor that doesn’t get too hot…
Then CPU performance benchmarks are close to irrelevant for you anyway. VPN might benefit from ARMv8 Crypto Extensions (which both RK3399 and A311D have) and torrent downloading has some light demands wrt random IO of the storage in question. Also no differentiation whatsoever.
GPU driver support with Android usually is no issue (since those SoCs are made for Android) but I’ve no idea about GPU performance differences and whether they’re important for anything besides gaming (I doubt it).
Nice. Thanks for your input.
I’m not a khadas staff, but I ran sbc-bench on my edge-v pro and got these results:
Memory performance (big.LITTLE cores measured individually): memcpy: 1437.0 MB/s (0.2%) memset: 4744.4 MB/s memcpy: 2132.2 MB/s memset: 4579.2 MB/s 7-zip total scores (3 consecutive runs): 3009,2970,2944 OpenSSL results (big.LITTLE cores measured individually): type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes 16384 bytes aes-128-cbc 38935.23k 123656.47k 263740.84k 385067.35k 444484.27k 449292.97k aes-128-cbc 92054.87k 241967.51k 371806.12k 419174.74k 447903.06k 448222.55k aes-192-cbc 37451.61k 111970.52k 220294.66k 300093.10k 335189.33k 335817.39k aes-192-cbc 105861.98k 238004.97k 320509.78k 376960.00k 394982.74k 395061.93k aes-256-cbc 36412.09k 104484.10k 193756.59k 252566.87k 277026.13k 278795.61k aes-256-cbc 103003.32k 218826.03k 299796.48k 325010.43k 339618.47k 340000.77k
Full result can be found here: http://ix.io/1MwC
The system currently runs from a sd card with Balbes armbian 5.88 latest image based on buster and kernel 5.2-rc4. The board is in the DIY case, with the new vim heatsink, but without fan so throttling must have occured.
Thank you but as you can see the performance of your install is inferior caused obviously by no cpufreq support with kernel 5.2. Your scores are lower than 50% of what usual RK3399 builds with either RK 4.4 BSP or older mainline kernel variants provide. According to 7-zip’s own calibration routine both the big and little CPU cores are limited to 600 MHz.
I was more interested in VIM3 benchmarks since with RK3399 everything is already known and result variation as in this case is ‘settings’
Definitely, but I think his point was: why release something new and better so close after you just sent the product that comes before. More like they probably cannibalize themselves except for users that need the edge (no pun) that Edge gives when it comes to display/camera and M2 cappabilities (here the differences). The thing is to see which one is worth for you, but in general if displays/M2 doesn’t matter to you and they end up having the same price, there’s no point in picking the Edge. The Edge is already comming, and I will probably also pick the VIM3 to decide which one fits better in my case.
That was precisely my point. With a negligible difference in performance and the fact that I’ve got to mess around with a case now, I would have just gone with the VIM-3 had I known. I regret my Edge-V purchase now and I haven’t even been able to use it yet, still got to doctor that damn case…
So you know the final VIM3 prices already?
The final price was given in another thread:
So the basic version should be 99$ and the pro 139$.
I always wondered which use case justifies spending those huge amounts of money for a tiny ARM thing (where you need to add an additional 50 bucks for enclosure, heatsink and fan). Especially now that the RPi 4 is available and still sells at below 40 bucks…
Great discussion topic! Don’t want to derail your performance thread, but I think the point is that SBCs/SoMs are used for all sorts of things, not just file servers and routers.
Your VIMS proposal would clearly be a market success for NAS at the price point you propose, whereas it’s difficult to justify a fuller-featured and more expensive board like the VIM3 or Edge V in that role. Most of a full-featured board’s strengths would be unused for a server.
But for some applications, price is less important than oodles of GPIO, quality construction, and a shape that works for integration.
An example: I’m designing a livestock feed system full of sensors, heating elements, motors and solenoid valves. I’m planning to use an Edge as controller and to drive the interface. These machines cost $15-20k+. A $100 SoM is still cheaper than just one of the solenoid pinch valves; even a $250 SoM would be cheaper than each precision gear pump.
What I like about the Edge:
- plenty of CPU and RAM for future software expansion
- loads of GPIO & SPI, not constrained by trying to squeeze onto a 40-pin DIP header
- built-in eMMC so I don’t have to deal with MMC on my own board, or use SD cards
- built-in wifi so I don’t have to plonk an external wifi module on my own board
- RSDB is really handy for infrastructure internet + ad-hoc net between parts of machine
- module form factor: separate RPi style board and 1.27mm ribbon cable looks amateurish
- support for decent resolution eDP touchscreens
Build quality is where some of the competitors really let themselves down. For example, I went through three Rock Pi 4 boards before I got one that was tidily soldered and not a mess. Users on the forum complain about connectors falling off, and certainly on the board I have here, there’s hot glue reinforcing the headphone jack from the factory because the attachment points aren’t up to the job. (I wondered if I’d be disappointed by Edge too, but actually I’m very impressed with the quality — which is quite unusual for me!)
Proper mainline kernel support really, really matters too, at least to me. I’d happily pay double not to be stuck in some board-specific kernel cul-de-sac, and similarly for u-boot. This horrible habit of so many ARM vendors turns long-term development/maintenance of a product into a complete nightmare!
Honestly doing this now is unfair, since nobody knew until today that Raspberry Pi 4 was to be released now or that it would have a 4xA72. Until today if you wanted a good SoC you had to go with +100bucks boards, with a few exceptions. Not that I’m going against RPi4, on the contrary I love that we finally got a powerful SoC for so cheap and will get the 4GB one once it’s in stock again. But it’s basically what ChrisW said. Khadas make good form factor boards that can be suited for commertial or even industrial products, whereas an Raspberry Pi is a general purpose computer with a jack of all trades form factor to please a big ammount of users. They also include eMMC. And if you were already complaining about the cooling system Khadas has, now imagine the nightmare it will be to properly cool a RPi4 having all that tiny space for general heatsinks/fans.
One of the more disappointing things from the Raspberry Pi 4 announcement from my personal perspective (wanting to integrate boards into low-volume but high-end products) is that it still has that horrible form factor.
Having all manner of random connectors plonked higgledy piggledy along two perpendicular edges is practically guaranteed to make it a nuisance and untidy to integrate… and other designers then copy the same mess in order to be compatible with cases, add-on boards and the rest of the RPi ecosystem. Argh!
I’m glad at least some boards are designed with better taste.
Honestly if they released a RPi Compute Stick 4 with the 4xA72 that would be perfect for me since it would still cost 30-35 bucks and then you can design your own board to fit your product, but the normal RPi form factor doesn’t fit my needs and has many connectors that I won’t be using (not a problem in general, but they would lie in the outside part) and the Compute Stick 3 isn’t powerful enough.
I was also looking the Rock960 which was really similar to the Edge-V but I didn’t even got a response from them, so I prefered to go for the one that has an active community instead of finding myself in a dead end without any help.