AMD EPYC 9554 & EPYC 9654 Benchmarks - Outstanding Performance For Linux HPC/Servers

[Spacefish]

I think going below 35 W is probably going to be a bigger tradeoff, at least without dropping some PCIe lanes & speeds, as well as swapping out DDR5 for LPDRR5.

Guess this was the main source of idle power draw on the AM4 platform.. The "X570" Chipset did draw ~10W as PCIe 4.0 on 14nm just sucks..
Furthermore the I/O die on matisse CPUs (Ryzen 3xxx) was manufactured on in 12? or 14nm? by Global Foundries, it contains like 32 SerDes and the DDR4 memory Controllers.. Compared to Renoir (single die Zen 2) the Matisse CPUs had a much higher power draw, due to the "old process" I/O die..

In Idle most of the power draw is caused by the platform arround the zen chiplets but not by the cores themself..

The high-end AM5 platform is even more power hungry, 7W per PROM21 chip (two for X670E, one for B650 and A620), so 14W (X670E) instead of 10-11W (X570).
It seems like the they moved to 6nm TSMC for the chipset chip.

[Paradigm Shifter]

They don't. Server CPUs stick to their power & thermal envelope religiously. Going outside of that causes problems and additional operational costs that aren't likely to go unnoticed by their customers.

This is essentially why 14 nm Xeons would experience such heavy clock-throttling from AVX-512, while the desktop Rocket Lake CPUs would just guzzle more power while clocks barely dipped.

Yeah, I know they don't right now... but there's a first time for everything, and with Genoa appearing to stomp a new hole in the current Xeon range, we can only guess what Sapphire Rapids will bring. Hopefully not stupid stuff like exceeding TDP.

[t.s.]

Uh, well the other catch is that the AmLogic SoC powering N2+ is limited to 4 GB, which is a bit low for desktop usage. As an owner of a N2+, the few times I've used it with a GUI, I can say it feels very desktop-like, in its performance. However, it's not in the same league, performance-wise, as the newer RK3588-based SBCs.

The Orange Pi 5 supports more RAM, but says it requires a 20 W power supply. I don't see where their peak power dissipation is stated, however.

Driver support for the GPUs and hardware video decoders in these SBCs tends to be somewhat spotty. That's why I like going the Intel route - their Apollo/Gemini/Jasper/Elkhart Lake SoCs have iGPUs of the same lineage that we see on desktop and mobile. Their drivers are similarly reliable.

20 watt PSU still reasonable, as RockPi5B uses ~12W when full loaded (according to first hand tester). Add a couple of usb, 1x m.2 nvme, then it's around that number. Agree with the driver support. If you want good hardware accell, you can go android route (then use termux for linux). Not a smart move, yes, but if you had to..

As for Intel atom/ celeron level, the multithread score still below RK3588 perf (https://github.com/ThomasKaiser/sbc-...ter/Results.md). Their power/perf efficiency too. The pluses, of course, the hardware accell. I think, you'll better served with business mini-pc(s) like thinkcentre, elitedesk/prodesk, or optiplex than with these low-end Intels. I've a second hand elitedesk 800 G4 with i5-8500, 8GB/512GB that idle ~ 6 watt. Underclock + undervolt it, so as it won't go above 40 w.

Edit: and you could pre-order orangepi 5 4-8-16GB from aliexpress now. They'll send it in December. Quite cheap.

[coder]

Yeah, I know they don't right now... but there's a first time for everything, and with Genoa appearing to stomp a new hole in the current Xeon range, we can only guess what Sapphire Rapids will bring. Hopefully not stupid stuff like exceeding TDP.

You should be able to simply look at the specified TDPs, to see the elevated power consumption of this generation. These are customers who really don't like surprises, because more power consumption than specified would blow a hole in their TCO calculations and could exceed the capacity of their cooling & electrical infrastructure.

[coder]

As for Intel atom/ celeron level, the multithread score still below RK3588 perf (https://github.com/ThomasKaiser/sbc-...ter/Results.md).

Wow! Thanks for the link! That's some great data! I just wish the user interface were a bit more friendly. I see no way to sort by column value or export as a spreadsheet.

Their power/perf efficiency too.

I don't know where you're getting efficiency numbers, because neither that chart nor the full logs seem to have efficiency or power.

However, the N6005 indeed managed to win at single-threaded 7-zip. As for multi-threaded performance, I'd be curious to know how much the 4x A55's are adding to the Rock 5B's numbers. I know it's a little bit moot, but it would be enlightening to know how Cortex-A76 compares with Tremont, especially ISO-clock or ISO-power.

Edit: and you could pre-order orangepi 5 4-8-16GB from aliexpress now. They'll send it in December. Quite cheap.

Current preorder price is $75 for the 8 GB model. I have offered to donate one to the site, if Michael is interested. He just needs to PM me the pre-checkout total and I'll paypal the $.

https://www.aliexpress.us/item/32568...andl_shipto=US

However, from what I can tell, it has a fan header but they offer no heatsink or fan specifically for it. My best guess is that the SoC package is about 17 mm on a side, with the dotted line around it measuring 19 mm on a side. I think you could fit a 20 mm x 20 mm copper heatsink, like this one:

https://www.aliexpress.us/item/32568...andl_shipto=US

[coder]

Here's an excerpt of that table, including some of the more notable entries:

Device / details	Clockspeed	Kernel	Distro	7-zip multi	7-zip single	AES	memcpy	memset	kH/s
Radxa ROCK 5B	2350/1830 MHz	5.10	Focal arm64	16450	3146	1337540	10830	29220	25.31
Pentium N6005	3300/2000 MHz	6.0	Jammy amd64	10810	3485	922000	9600	11300	20.15
ODROID-N2+	2400/2015 MHz	5.14	Hirsute arm64	9780	2386	1366090	4030	7120	-
Raspberry Pi 4 B	1800 MHz	5.15	Jammy armhf	6300	1844	82750	1190	3110	-
Raspberry Pi 3 B+	normal	4.14	Raspbian Stretch	3040	856	36600	1050	1500	-

I was trying to find a good way to copy-paste tables into a post, but without success. Finally, it occurred to me to save a spreadsheet as HTML and then open it with VIM and run a few substitutions. Works, but annoying. If I had to do it a lot, I'd write an XSLT script.

[Michael]

Current preorder price is $75 for the 8 GB model. I have offered to donate one to the site, if Michael is interested. He just needs to PM me the pre-checkout total and I'll paypal the $.

Thanks, currently though waiting to hear back from OrangePi... They messaged me on Twitter asking if I'd write a news about it, I asked them then about any review sample.

[t.s.]

Wow! Thanks for the link! That's some great data! I just wish the user interface were a bit more friendly. I see no way to sort by column value or export as a spreadsheet.

I don't know where you're getting efficiency numbers, because neither that chart nor the full logs seem to have efficiency or power.

However, the N6005 indeed managed to win at single-threaded 7-zip. As for multi-threaded performance, I'd be curious to know how much the 4x A55's are adding to the Rock 5B's numbers. I know it's a little bit moot, but it would be enlightening to know how Cortex-A76 compares with Tremont, especially ISO-clock or ISO-power.

You're welcome.

For the efficiency number, it's not included there. But you can find the estimated number here for Rock5B (RK3588) (https://forum.radxa.com/t/rock-5b-de...tion/10483/400). For other number, sorry, just pulled out from my memory, but IIRC, I read it at one of the RK3588 SBC articles from cnx-software.com.

N6005 win because their turbo boost, I think. If the Cortex A-76 clocked to ~2.8GHz, I think it'll tied. Here https://forum.radxa.com/t/rock-5b-de...ation/10483/56 for the A55 perf.

[t.s.]

Here's an excerpt of that table, including some of the more notable entries:

Device / details	Clockspeed	Kernel	Distro	7-zip multi	7-zip single	AES	memcpy	memset	kH/s
Radxa ROCK 5B	2350/1830 MHz	5.10	Focal arm64	16450	3146	1337540	10830	29220	25.31
Pentium N6005	3300/2000 MHz	6.0	Jammy amd64	10810	3485	922000	9600	11300	20.15
ODROID-N2+	2400/2015 MHz	5.14	Hirsute arm64	9780	2386	1366090	4030	7120	-
Raspberry Pi 4 B	1800 MHz	5.15	Jammy armhf	6300	1844	82750	1190	3110	-
Raspberry Pi 3 B+	normal	4.14	Raspbian Stretch	3040	856	36600	1050	1500	-

I was trying to find a good way to copy-paste tables into a post, but without success. Finally, it occurred to me to save a spreadsheet as HTML and then open it with VIM and run a few substitutions. Works, but annoying. If I had to do it a lot, I'd write an XSLT script.

I've tried RPi3+, RPi4B, N2. RPi4B (idle ~4w) and N2 (idle ~2w) surprisingly capable and quite quick opening apps, despite using microSD. The problem, especially N2 is the 4GB RAM. Now, with something like OPi5 and Rock5B that have up to 16GB, my dream to have a ~2 watt desktop will come true soon-ish. Well, must wait for the GPU driver, ~1-2 years.

[coder]

I've tried ... N2 (idle ~2w) surprisingly capable and quite quick opening apps, despite using microSD.

100% agree. I got the eMMC module, and the N2+ gives the most desktop-like user experience I've had from a SBC.

The problem, especially N2 is the 4GB RAM. Now, with something like OPi5 and Rock5B that have up to 16GB, my dream to have a ~2 watt desktop will come true soon-ish.

Indeed. However, I'm in no hurry. I'm curious to see if HardKernel will release an ODROID based on the RK3588, or similar. Maybe they'll use their H-series form factor, since those have a bigger passive heatsink and can accommodate a M.2 SSD.

And although I really like what Firefly did, it's just too expensive for what it is. I feel like they over-engineered it, with that whole daughter board aspect.

https://en.t-firefly.com/product/industry/itx3588j

BTW, I really wish we could get some momentum behind one of the standard form factors, like nano-ITX or 5x5.