Here are some quotes by the Banana Pi simon about the power needed
1 Like
The clock speed of the R4's cpu is 200 MHz lower can the R3's speed. That maybe why in almost all benchmarks the R4 get beaten by the R3 but actually not by a big margin. There is one very prominent exception. AES based benchmarks are through the roof on the R4.
The CPU benchmark section seems very wrong. Perhaps something not right with your benchmark. R4 has four ARM Cortex A73 (so called performance core) while R3 has four ARM Cortex A53 (so called efficient core). The performance difference should be day & night. Such as what you observed in AES benchmark.
2 Likes
Agree, the technology advancement is big enough to make the difference, for example you won't expect an ancient Pentium 4 3GHz can run faster than a Celeron J4125 (2-2.7GHz), right?
And for R3, the switch chip on the 2.5GbE which means if you plug 3 x 1GbE together and all running 1Gbps full speed you will expect delay, and this won't happen on R4.
I have not published my graphs yet but I fear they are correct. I mean the clock speed is lower. I will repeat the benchmarks one more time and see if I made any errors but so far I'm quite sure that they're correct.
repeat the Banana Pi BPI-R3 test with OpenSSL 3.0.13
According to the table, the BPI-R3 also beats an A72 with 2.2GHz in SHA-1 + SHA265.
Technically not possible - can only be due to the SSL version used.
I have used OpenSSL 3.0.13 in the first place and I repeated the tests; conclusion: the values are correct.
I also updated my review page adding benchmark result graphs and some details.
1 Like
Or there is hardware acceleration involved?
What kind of hardware acceleration?
These are both ARMv8 processors with CE (Crypto Extension)
Banana Pi BPI-R3 = Cortex A53 2Ghz (little cores)
PUZZLE-M902 = Cortex A72 2.2GHz (big cores)
completely impossible that the A53 beats the A72 in any benchmark if the same software is used.>
OpenWRT 23.05 uses OpenSSL 3 and I experienced lower scores compared to OpenSSL 1 scores.
I'm talking about the results >
https://openwrt.org/docs/guide-user/perf_and_log/benchmark.openssl
OpenSSL 1.1.1s is specified for the Banana Pi BPI-R3
yeah it looks like I would need to redo all the compared benchmarks with 3.0.13 as it seems to boost AES stuff.
1 Like
I see some improvement comparing what is in the OpenWRT wiki page:
root@BPI-R4:~# ubus call system board
{
"kernel": "6.6.27",
"hostname": "BPI-R4",
"system": "ARMv8 Processor rev 0",
"model": "Bananapi BPI-R4",
"board_name": "bananapi,bpi-r4",
"rootfs_type": "squashfs",
"release": {
"distribution": "OpenWrt",
"version": "SNAPSHOT",
"revision": "r25933-cab2e1de0d",
"target": "mediatek/filogic",
"description": "OpenWrt SNAPSHOT r25933-cab2e1de0d"
}
}
| r25933 single-thread | 26.00 | 3.0.13 | 240527700 | 620693500 | 595024550 | 133290330 | 37625860 | 13402110 | 1428089510 | 1191378940 | 1022088530 | 195.2 | 7417.8 | 536.7
| 583.5 |
and
| r25933 -multi 4 | 26.00 | 3.0.13 | 957951320 | 2472094720 | 2375025320 | 532530040 | 150383270 | 53463720 | 5706645500 | 4759071400 | 4077639000 | 779.5 | 29545.7 | 2135.9 | 2303.2 |
indeed, that is better compared to what I achieve. You do not happen to have a heatsink installed or maybe even active cooling?
Here are my details. Why do you have kernel 6.6?
root@r4:/# ubus call system board
{
"kernel": "6.1.82",
"hostname": "r4",
"system": "ARMv8 Processor rev 0",
"model": "Bananapi BPI-R4",
"board_name": "bananapi,bpi-r4",
"rootfs_type": "squashfs",
"release": {
"distribution": "OpenWrt",
"version": "SNAPSHOT",
"revision": "r25876-923d7c5531",
"target": "mediatek/filogic",
"description": "OpenWrt SNAPSHOT r25876-923d7c5531"
}
}
I have a heat sink, but there is no difference during making the tests in the temperature.
Kernel 6.6 is marked as a TESTING_KERNEL in OpenWRT - I build my own image with included drivers/packages.
Is there any difference in the results comparing to kernel 6.1?
yeah. The results of the R4 in the OpenWrt Wiki are mine...
Edit:
I tested all devices again and updated my results in my review. Tested all devices with OpenSSL 3.0.13 and the overall conclusion keeps the same. The R4 is actually a bit slower than the R3, at least in most of the disciplines.
Yes, but the diagram looks more coherent this way and the marginal differences can be attributed to the clock frequency
and in MD5 SHA1 SHA2 the BPI-4 is now faster.
1 Like
Well if the R4 isn't ready to use the crypto extension in snapshot build then it could be possible?
such values cannot be achieved without CE
Whether the CPU features are activated can be seen in "cat /proc/cpuinfo"
Features : fp asimd aes pmull sha1 sha2 crc32 cpuid
OpenSSL uses CPU specific assembler code for CE
Comparative values from my 2.2GHz A72 are in the table
maybe there is s room for optimization - who knows
There are a couple of pitfalls in running openssl. Also, as you said kernel/software readiness. I'll take the result with a pinch of salt. Because A73 simply cannot just perform on par with A53. LOL
Hope I can get my hands on R4 soon.
If I do the math, the MT7988A is ~30% slower than the CN9130 at 20% less clock speed.
I would rather expect the MT7988A to be ~10-15% slower.
But who knows maybe bad silicon or missing some kernel optimizations.
The A73 cores are marginally faster than those of the A72, about ~15%
Maximum 1.2 DMIPS/MHz
The A73 is also not marketed as a performance successor to the A72 but as a power efficient successor that doesn't overheat and doesn't need as much cooling