The ODROID H-series is growing with three brand new models (ODROID H4/H4+/H4 Ultra)

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

domih wrote:
Mon Apr 22, 2024 2:01 am
Hardkernel has being using the same thermal paste for years on all its ARM and x86 models.
I just pulled off the Jetway's heatsink and found they used what looks to be a 0.5 mm thermal pad. Their heatsink has a central keel that's about 1mm thicker, which crosses the SoC dies.

I tried removing the rubber holdups and placing a passive copper LGA 775 heatsink on its existing thermal pad. It's quite a bit more massive and has much more fin area, being both about twice as tall and having about half the fin pitch. The difference isn't enough to overcome the lack of a fan, however, with idle temps being markedly higher (about 50 C instead of 30 C). I'll probably try replacing the thermal pad with some Artic MX-6 I have, just to see how much worse that thermal pad really is.
domih wrote:
Mon Apr 22, 2024 2:01 am
This is what I witnessed with the H3 and H3+. Given how Hardkernel designed the H4, I'm pretty sure it will have the same behavior.
Thanks for sharing.

I would note that the H3 uses a SoC that's made on Intel's 10 nm process (same as Ice Lake). The H4's SoC uses the first gen Intel 7 node that's used in the other Alder Lake variants. In spite of the name, the density should be similar. However, the main difference between the nodes is how much higher the Intel 7 node clocks (which explains why Intel cancelled the desktop version of Ice Lake and even retained 14 nm Skylake-derivatives as the high-end Wiskey Lake laptop tier).

Perhaps more importantly, Gracemont is a higher-performance microarchitecture. So, experiences with the H3 might have somewhat limited applicability to the H4.
Last edited by mattbot on Mon Apr 22, 2024 6:05 am, edited 1 time in total.

User avatar
mctom
Posts: 4330
Joined: Wed Nov 11, 2020 4:44 am
languages_spoken: english, polski
ODROIDs: XU4, M1, H3+, SP3, N2L, M1S, N2
Location: Gdańsk, Poland
Has thanked: 642 times
Been thanked: 875 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mctom »

I wish I wasn't bound by NDA not to discuss Intel nodes, or any others for that matter.

Just wanted to make sure my message was clear. Aluminum heat sinks aren't worse. They are designed differently with the material considered.
True, there are partially or fully copper heat sinks, especially when space is at premium and one needs a slim but tall heat sink, such as "tower" CPU coolers or in laptops.
In other words, you can replace every copper heat sink with aluminum equivalent, but will require more material and will be bigger. H-series heat sinks are huge and do their job really well.

As I expected, your issue with that other computer is package-to-heatsink resistance (or die-to-heatsink as you pointed out there is no package). I'm convinced this will not be the case for H4, alas I have no proof for that.
My H3+ with stress -c 4 and 60mm Noctua side fan settled at 60C just now. :)

By the way, smaller foundry nodes do not translate directly into certain speed gains or power / surface area savings. Microelectonic designs are ported automatically and then meticulously verified and fixed by a horde of cheap engineers from Poland, lol. There are no exact design copies from one node to another, so there are no rules as of what gains are to be expected.
I've heard of microelectronic projects where the customer wanted equally performing circuit in half-sized node, thinking this will let them save power and pack more chips on a wafer. In the end they got the same die dimensions, worse power losses and paid for a more expensive production node. Newer nodes aren't always better at everything, and "clock speed" is just one oversimplified node parameter.

I use thermal pads on ARM SoCs, because they are more convenient to apply and maintain, but I'd rather not use them for TDP > 5W. You may solve your problems by applying thermal paste instead.
Punk ain't no religious cult, punk means thinking for yourself!
OpenUPS
PiStackMon

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

mctom wrote:
Mon Apr 22, 2024 4:25 am
Aluminum heat sinks aren't worse.
The thermal conductivity of aluminum is 237 W/(m*K). For copper, this figure is about 400 W/(m*K). A higher thermal conductivity means heat can be drawn away faster, which reduces the severity of thermal hot spots.

For server CPUs and high-end GPUs, even the superior conductivity of pure copper is insufficient. Instead, they opt for vapor chambers which tend to range well into the thousands.
mctom wrote:
Mon Apr 22, 2024 4:25 am
you can replace every copper heat sink with aluminum equivalent, but will require more material and will be bigger.
I think you're assuming there's no bottleneck in transferring heat from the die to the heatsink. As I said, the kind of CPUs you probably have in mind use a copper heat spreader, in order to increase the interface area and reduce the problem of bottlenecks in transferring heat into the heatsink.
mctom wrote:
Mon Apr 22, 2024 4:25 am
My H3+ with stress -c 4 and 60mm Noctua side fan settled at 60C just now. :)
Please try with --cpu-method=matrixprod, as the different CPU methods vary greatly.
mctom wrote:
Mon Apr 22, 2024 4:25 am
smaller foundry nodes do not translate directly into certain speed gains or power / surface area savings.
The concern is W/m^2. Ryzen 7000 CPUs provide an interesting case study, as they tend to have thermal bottleneck issues, due to the thermal density resulting from their use of a 5 nm process node and relatively small cores (i.e. simpler than Intel's Golden Cove).
mctom wrote:
Mon Apr 22, 2024 4:25 am
There are no exact design copies from one node to another,
Nobody said otherwise. Jasper Lake used Tremont cores on Intel 10 nm, while Alder Lake-N uses Gracemont cores on Intel 7. Different microarchitectures and different nodes.
mctom wrote:
Mon Apr 22, 2024 4:25 am
so there are no rules as of what gains are to be expected.
LOL, TSMC would say otherwise. Each new process node announcement is accompanied by two numbers (assuming the same microarchitecture):
  • performance increase @ ISO power
  • power savings @ ISO performance
Nicely encapsulated in this info graphic, courtesy of WikiChip:

Image

In their overview of Intel 4, Wikichip also repeats what I assume is an Intel claim about that node:

"The node takes full advantage of EUV and offers around a 20% performance/watt gain over Intel 7. At the SoC level, the node offers as much as a 40% reduction in power at iso-frequency or >20% frequency improvement at iso-power."
mctom wrote:
Mon Apr 22, 2024 4:25 am
I've heard of microelectronic projects where the customer wanted equally performing circuit in half-sized node, thinking this will let them save power and pack more chips on a wafer. In the end they got the same die dimensions, worse power losses and paid for a more expensive production node. Newer nodes aren't always better at anything, and "clock speed" is just one oversimplified node parameter.
Not all nodes are optimized the same way, however that project also sounds like it was horribly-managed. Also, I presume you're talking about two nodes from the same foundry. Comparing nodes between foundries has long been fraught with pitfalls. The above Intel 4 article takes quite some pains to try and compare that node to TSMC and Samsung's.
mctom wrote:
Mon Apr 22, 2024 4:25 am
I use thermal pads on ARM SoCs, because they are more convenient to apply and maintain, but I'd rather not use them for TDP > 5W. You may solve your problems by applying thermal paste instead.
Yeah, replacing it with Arctic MX-6 helped tremendously! Initially, it dropped temps by about 15 C. When I hit it with a single-core matrixprod load, it actually took a couple minutes before the heatsink becomes completely heat-soaked. Eventually, PkgWatt settled near 6 W, down from its initial value of 12 W.

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Surprise: Heatsink compound really matters!

Post by mattbot »

Just to follow up on the end of that last post, I re-mounted the original heatsink/fan on the Jetway board, but with Arctic MX-6. I also removed some plastic spacers that seemed to add about as much height as the thickness of the thermal pad. Of course, this is not without some risk of cracking the die, but I decided it wasn't fit for my purposes as it was.

The result was a similar 15 C drop in temperatures I saw with the passive, copper heat sink! Now, (updated) it only throttles on 2-core (among CPU-only) workloads. That said, keep in mind my ambient is 21 C and it's on an open bench. Also, the fan isn't exactly quiet under full load.

Interestingly, temperatures are actually lower with a 4-core workload than 1 or 2. Makes sense, if you think about it, because it's the same 12 W in all cases. So, there should be less of a hotspot if most of that is being distributed over 4 cores' area than just 1 or 2. (update) The 1-core case ends up being frequency-limited. It turns out the 2-core case can get enough frequency on each core to throttle, if I let it run long enough.

Something else I noticed is that even when I tweaked the PL1 & PL2 values, the CPU cores really wouldn't push package power figures above 13 W or so. I probably just need to spend some time experimenting with CPU frequency scaling governors to see if I can get more aggressive multi-core frequency curves. Anyway, if I added in a GPU workload, I could get package power to almost 18 W (assuming PL1 was at least that). Furthermore, that GPU workload would cause the CPU cores to throttle, if there wasn't enough power budget for both.

In conclusion, there definitely was a "thermal bottleneck", at the interface of the die and heatsink. However, I probably overestimated the impact of the heatsink material and underestimated quite how awful that thermal pad was. I still think I'd get lower temps (and therefore less fan noise), if they'd made that heatsink out of copper.

Anyway, it seems this is something to lookout for, with the H4 series. I hope someone takes the chance of testing with a high-performance thermal compound. Just beware that doing so is not without risk and will probably void your warranty.
Last edited by mattbot on Mon Apr 22, 2024 12:25 pm, edited 2 times in total.

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

mattbot wrote:
Mon Apr 22, 2024 1:29 am
I started out by asking how the H4's core temps look. That's the central question of my original post, in this series. It's right there in the first sentence.
I've just run stress-ng --cpu=1 --cpu-method=matrixprod and kept monitoring the CPU clock frequency and temperature for several minutes with the 92x92x15mm official cooling fan at default 1000 RPM on my H4.
The CPU frequency stably stay at 3.6GHz and the core temperature was hovering between 77~80°C while my ambient temperature was about 26°C. The 'Unlimited Performance' mode was enabled certainly.
Note that we've been using the same thermal paste for the H2, H3, and H4 series for the past 5 years and haven't seen any complaint reports about thermal throttling problem as far as I recall.

BTW, when I ran stress-ng --cpu=4 --cpu-method=matrixprod on my H4, the 4-core multi-threaded sustained CPU frequency was 2.9Ghz as expected with similar temperature characteristic.
I mean there was no thermal throttling at all with the stock heatsink and the slim cooling fan inside the official H4 case Type-1.
These users thanked the author odroid for the post:
mattbot (Mon Apr 22, 2024 12:26 pm)

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

odroid wrote:
Mon Apr 22, 2024 11:49 am
I've just run stress-ng --cpu=1 --cpu-method=matrixprod and kept monitoring the CPU clock frequency and temperature for several minutes with the 92x92x15mm official cooling fan at default 1000 RPM on my H4.
...
when I ran stress-ng --cpu=4 --cpu-method=matrixprod on my H4, the 4-core multi-threaded sustained CPU frequency was 2.9Ghz as expected with similar temperature characteristic.
Excellent news! Thank you for checking!

BTW, I have yet to experiment with different frequency-scaling governors/settings, but I found that the 2-core case was actually most stressful, on my Jetway board. I updated my above post to indicate this: 2-core is the only CPU-only benchmark I can get to thermally throttle, now. I think that's because 1-core is frequency-limited and 3- & 4-core distribute the power over enough die area that the heatsink can keep up. However, in the 2-core case, the frequencies are still quite elevated, but the die area generating most of the heat is still fairly small.

As for GPU workloads, the most stressful one I've found was the glmark2 buffer benchmark, with the update-method=subdata parameter. For whatever reason, it's a lot more energy-intensive than the other buffer benchmarks. Also, using the --off-screen is much less energy-intensive than just letting it render to a window. I used a window resolution of 1280x1024.

Anyway, it sounds like you guys have another winner on your hands. Keep up the good work!
:-)

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

mattbot wrote:
Mon Apr 22, 2024 12:36 pm
BTW, I have yet to experiment with different frequency-scaling governors/settings, but I found that the 2-core case was actually most stressful, on my Jetway board. I updated my above post to indicate this: 2-core is the only CPU-only benchmark I can get to thermally throttle, now. I think that's because 1-core is frequency-limited and 3- & 4-core distribute the power over enough die area that the heatsink can keep up. However, in the 2-core case, the frequencies are still quite elevated, but the die area generating most of the heat is still fairly small.
For completeness,
With stress-ng --cpu=2 --cpu-method=matrixprod command on H4, the maximum sustained CPU frequency was 3.3~3.4Ghz and the CPU temperature was about 74~77°C.
With stress-ng --cpu=3 --cpu-method=matrixprod command on H4, the maximum sustained CPU frequency was 2.9~3.0Ghz and the CPU temperature was about 75~77°C.

For reference, we confirm that 3Ghz CPU clock frequency is sustained with the H4 Ultra model when all 8 CPUs are used with stress-ng --cpu=8 --cpu-method=matrixprod.
Compared to the H4/H4+ models, the H4 Ultra has a higher operating clock of about 100~200Mhz and the temperature was slightly higher at 78~82°C. But there was no thermal throttling phenomenon at all.
Since Intel has designed thermal throttling to start when the SoC temperature exceeds 95°C, we still have some room for GPU loads probably.
These users thanked the author odroid for the post (total 2):
mattbot (Mon Apr 22, 2024 3:09 pm) • xnd (Mon Apr 22, 2024 3:12 pm)

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

odroid wrote:
Mon Apr 22, 2024 2:49 pm
Compared to the H4/H4+ models, the H4 Ultra has a higher operating clock of about 100~200Mhz and the temperature was slightly higher at 78~82°C. But there was no thermal throttling phenomenon at all.
Since Intel has designed thermal throttling to start when the SoC temperature exceeds 95°C, we still have some room for GPU loads probably.
Yes, that's consistent with my Jetway board, where thermal clock-throttling only seems to kick in at about 95 C. I believe there's a BIOS option to adjust that threshold, but I have yet to touch any of those settings. I will probably see if I can configure a throttling temperature of about 85 C, just for the sake of hardware stability and longevity.

P.S. did you see my original post, where I asked about in-band ECC? I would love to have that capability!

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

mattbot wrote:
Mon Apr 22, 2024 3:15 pm
P.S. did you see my original post, where I asked about in-band ECC? I would love to have that capability!
I already replied to your post a few hours ago. I hope my answer was wrong though.

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

odroid wrote:
Mon Apr 22, 2024 3:26 pm
mattbot wrote:
Mon Apr 22, 2024 3:15 pm
P.S. did you see my original post, where I asked about in-band ECC? I would love to have that capability!
I already replied to your post a few hours ago. I hope my answer was wrong though.
Sorry, but I don't see any private messages and the only hits I'm getting on "ECC", in this thread, are in my own posts. Could you please repeat what you know about the potential for in-band ECC (IB-ECC) on Alder Lake-N boards? I linked to two other examples which claim to have it, but are (coincidentally?) also using soldered-down LPDDR5.

You've indeed been very helpful in addressing my concerns about temperatures and cooling. Thank you for that!

hominoid
Posts: 955
Joined: Tue Feb 28, 2017 3:55 am
languages_spoken: english
ODROIDs: C2, C4, XU4, MC1, N1, N2, N2L, N2+, HC4, M1, H2, H3+
Location: Lake Superior Basin, USA
Has thanked: 144 times
Been thanked: 464 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by hominoid »

It might be this.
viewtopic.php?f=171&t=48377
These users thanked the author hominoid for the post (total 2):
odroid (Tue Apr 23, 2024 9:27 am) • mattbot (Tue Apr 23, 2024 2:52 pm)

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

mattbot wrote:
Mon Apr 22, 2024 10:32 pm
Sorry, but I don't see any private messages and the only hits I'm getting on "ECC", in this thread, are in my own posts. Could you please repeat what you know about the potential for in-band ECC (IB-ECC) on Alder Lake-N boards? I linked to two other examples which claim to have it, but are (coincidentally?) also using soldered-down LPDDR5.
Sorry. I confused you with another user. :o

The same thread as @hominoid pointed.
viewtopic.php?f=171&t=48377
These users thanked the author odroid for the post:
mattbot (Tue Apr 23, 2024 2:52 pm)

TOMillr
Posts: 66
Joined: Fri Feb 23, 2018 3:41 am
languages_spoken: english
Has thanked: 1 time
Been thanked: 0
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by TOMillr »

Two quick questions regarding the Peripheral Expansion Header on the new boards:

1. Can you connect an external push button to power on the board?
2. Is there a way to internally connect USB devices?



Image

User avatar
domih
Posts: 812
Joined: Mon Feb 11, 2019 4:48 pm
languages_spoken: English, French
ODROIDs: UX4, HC2, N2, N2+, H2, H2+, C4, HC4, M1, M1S (with UPS, 4 x Relay) H3, H3+, H4+, H4 Ultra - 1GbE, 2.5GbE, 10GbE, 45+ GbE
Location: San Francisco Bay Area
Has thanked: 339 times
Been thanked: 247 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by domih »

TOMillr wrote:
Fri Apr 26, 2024 4:23 am
Two quick questions regarding the Peripheral Expansion Header on the new boards:

1. Can you connect an external push button to power on the board?
2. Is there a way to internally connect USB devices?
1. Yes. See https://wiki.odroid.com/odroid-h4/appli ... wer_switch.
2. Yes. 3 x USB 2 as indicated on the diagram. See USB2_P[5|6|7]_[DP|DN]. See also viewtopic.php?f=172&t=45594 for what users did with the H3. Same with the H4.
These users thanked the author domih for the post:
odroid (Fri Apr 26, 2024 9:00 am)

TOMillr
Posts: 66
Joined: Fri Feb 23, 2018 3:41 am
languages_spoken: english
Has thanked: 1 time
Been thanked: 0
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by TOMillr »

domih wrote:
Fri Apr 26, 2024 5:18 am
See also viewtopic.php?f=172&t=45594 for what users did with the H3. Same with the H4.
Sweet! Can you buy any of those boards pre-made from somewhere?

Also, how much max. power can those internal USB headers provide on the new H4 boards?

User avatar
mctom
Posts: 4330
Joined: Wed Nov 11, 2020 4:44 am
languages_spoken: english, polski
ODROIDs: XU4, M1, H3+, SP3, N2L, M1S, N2
Location: Gdańsk, Poland
Has thanked: 642 times
Been thanked: 875 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mctom »

TOMillr wrote:
Fri Apr 26, 2024 4:47 pm
Sweet! Can you buy any of those boards pre-made from somewhere?

Also, how much max. power can those internal USB headers provide on the new H4 boards?
Try asking in that thread maybe..

The current limit is not applicable to USB data lines.
Maximum current is mostly application dependent. My board has 1A limit per port, and 1.5A in total IIRC.
It's not meant for powering things. I use it every day for stuff like keyboard dongle, UART converters, audio adapters, pendrives and so on.
Punk ain't no religious cult, punk means thinking for yourself!
OpenUPS
PiStackMon

TOMillr
Posts: 66
Joined: Fri Feb 23, 2018 3:41 am
languages_spoken: english
Has thanked: 1 time
Been thanked: 0
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by TOMillr »

odroid wrote:
Wed Apr 17, 2024 5:46 pm
xnd wrote:
Wed Apr 17, 2024 5:12 pm
Regarding Intel I226-V, from another forum:
The I226-V is a problem. Had to switch off in my kernel ASPM, because otherwise the Ethernet connection will hang in after a few hours. Has been a familiar problem without real solution for years.
oh, is that true? :shock:

Could you please try to run H4 (Plus/Ultra) as headless server with ASPM enabled (and only LAN plugged) for couple of days and regularly monitor network somehow?
If you agree to configure the test unit as above and continue logging and check the log file one time per day for about a week, we will start testing tomorrow.
Any news regarding this issue? Could you replicate the problem?

xnd
Posts: 112
Joined: Sun Dec 04, 2022 7:48 pm
languages_spoken: english, slovak, czech
ODROIDs: H3, H4+
Location: Slovakia
Has thanked: 48 times
Been thanked: 34 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by xnd »

TOMillr wrote:
Sat Apr 27, 2024 11:46 pm

Any news regarding this issue? Could you replicate the problem?
Here is topic about that viewtopic.php?f=171&t=48358
proud owner of Odroid H3 ( + 48 GB RAM: 16GB+32GB Crucial CT32G4SFD832A & Samsung SSD 980 PRO 1TB M.2; OS: DietPi - Debian 12; idle power consumption: ~ 1.1—2W)

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

odroid wrote:
Tue Apr 23, 2024 9:27 am
mattbot wrote:
Mon Apr 22, 2024 10:32 pm
Sorry, but I don't see any private messages and the only hits I'm getting on "ECC", in this thread, are in my own posts. Could you please repeat what you know about the potential for in-band ECC (IB-ECC) on Alder Lake-N boards? I linked to two other examples which claim to have it, but are (coincidentally?) also using soldered-down LPDDR5.
Sorry. I confused you with another user. :o

The same thread as @hominoid pointed.
viewtopic.php?f=171&t=48377
BIOS 1.02 for H4 series with In-Band ECC function added has been released.
These users thanked the author odroid for the post (total 2):
mattbot (Tue Apr 30, 2024 3:15 pm) • domih (Tue Apr 30, 2024 7:43 pm)

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

odroid wrote:
Tue Apr 30, 2024 2:21 pm
BIOS 1.02 for H4 series with In-Band ECC function added has been released.
Such an understated post for a very cool feature! Thank you!
😎

I hope you update the H4/H4+/H4 Ultra product pages to reflect this, in due time.

mikefarmer01
Posts: 1
Joined: Thu Oct 15, 2020 6:12 am
languages_spoken: english, german, french
ODROIDs: N2
Has thanked: 0
Been thanked: 1 time
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mikefarmer01 »

Hi Hardkernel team,

consider adding the new boards to HackerBoards:
https://hackerboards.com/add

That way, the boards will receive better visibility.
Right now, only H2(+) are listed for x86_64 boards manufactured by Hardkernel.
These users thanked the author mikefarmer01 for the post:
mattbot (Thu May 23, 2024 7:31 am)

TOMillr
Posts: 66
Joined: Fri Feb 23, 2018 3:41 am
languages_spoken: english
Has thanked: 1 time
Been thanked: 0
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by TOMillr »

odroid wrote:
Tue Apr 30, 2024 2:21 pm
BIOS 1.02 for H4 series with In-Band ECC function added has been released.
Can somebody maybe elaborate on how enabling this impacts performance? Does the OS or the software running on it have to support this feature as well?

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

TOMillr wrote:
Mon May 20, 2024 8:24 pm
odroid wrote:
Tue Apr 30, 2024 2:21 pm
BIOS 1.02 for H4 series with In-Band ECC function added has been released.
Can somebody maybe elaborate on how enabling this impacts performance? Does the OS or the software running on it have to support this feature as well?
Synthetic memory benchmarking results show a memory read/write overhead of about 10-20%. But typical computing performance will see less than a 5% degradation.
To monitor ECC status, use Linux kernel 6.5 or higher.
viewtopic.php?p=384722#p384722

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

TOMillr wrote:
Mon May 20, 2024 8:24 pm
Can somebody maybe elaborate on how enabling this impacts performance? Does the OS or the software running on it have to support this feature as well?
Anandtech benchmarked a system with + without IB-ECC, however it was using a Raptor Lake P-series CPU. It should give you at least some idea of the impact.

https://www.anandtech.com/show/18732/as ... akep-ecc/2

For instance, 2.3% less performance on Cinebench R23 single-threaded, and 1.5% less performance on multi-threaded. On x264 encoding, IB-ECC caused a 5% deficit. 7-zip compression was down a whopping 30%! Although, decompression performance wasn't hit at all.

Hardware transcoding and GPU workloads, in general, were also significantly impacted. qsv_h264 took a 10.6% hit. GFXBench 5.0 took a 29.9% hit. However, 3DMark: Time Spy only took a 14.8% hit. Keep in mind that the i7-1360P's iGPU is 3-4x as big as the one in Alder Lake-N and a little higher-clocked.

Another thing to consider is that the i7-1360P system they tested has a lot more compute, relative to its memory bandwidth. It's has 4 P-cores + 8 E-cores, for a total of 16 threads. So, at least twice the compute performance of the N305, yet the memory configuration of that machine is only 2x DDR4-3200, which is only 33.3% higher than the 1x DDR5-4800 supported by the H4 series. Therefore the effect of reducing memory bandwidth is going to be more exaggerated, on the workloads where it has a significant impact.

For me, it's a no-brainer. At this price point and power level, if you want/need ECC, then the performance hit of enabling IB-ECC is very acceptable for most people and applications. Speaking of power, that i5-1360P system idles at 15.2 W and used up to 95.0 W under load. To get out-of-band ECC, I think you'd have to go to a full desktop die (or HX series CPUs, if a laptop).

fvolk
Posts: 945
Joined: Sun Jun 05, 2016 11:04 pm
languages_spoken: english
ODROIDs: C4, H3, M1S
Has thanked: 0
Been thanked: 155 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by fvolk »

Maybe I'm dumb, but nowhere is written what the size (outer dimensions) of each case is?
It is not on the product page, e.g. https://www.hardkernel.com/shop/odroid-h4-case-type-1/
?

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

fvolk wrote:
Tue May 28, 2024 6:08 am
Maybe I'm dumb, but nowhere is written what the size (outer dimensions) of each case is?
We've updated the product pages.
Type 1 : 130mm x 132mm x 66mm (height)
Type 2 : 130mm x 132mm x 76mm (height)
Type 3 : 150mm x 132mm x 127mm (height)
Type 4 : 225mm x 125mm x 200mm (height)
These users thanked the author odroid for the post (total 2):
mattbot (Sat Jun 01, 2024 6:42 am) • domih (Sun Jun 02, 2024 8:36 am)

fvolk
Posts: 945
Joined: Sun Jun 05, 2016 11:04 pm
languages_spoken: english
ODROIDs: C4, H3, M1S
Has thanked: 0
Been thanked: 155 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by fvolk »

odroid wrote:
Tue May 28, 2024 11:33 am

We've updated the product pages.
Type 1 : 130mm x 132mm x 66mm (height)
Type 2 : 130mm x 132mm x 76mm (height)
Type 3 : 150mm x 132mm x 127mm (height)
Type 4 : 225mm x 125mm x 200mm (height)
Much appreciated, thank you!

fvolk
Posts: 945
Joined: Sun Jun 05, 2016 11:04 pm
languages_spoken: english
ODROIDs: C4, H3, M1S
Has thanked: 0
Been thanked: 155 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by fvolk »

Looking at the videos of the cases again,
if there is only one 2.5" SSD, still use case 2 with the long cable and place it at the bottom where the netcard should be.
Would that work?

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by odroid »

fvolk wrote:
Fri May 31, 2024 4:48 am
if there is only one 2.5" SSD, still use case 2 with the long cable and place it at the bottom where the netcard should be.
Would that work?
If you don't use a Net-Card, there seems to be enough space for a 7mm thick 2.5" SSD.
However, there is no gap between the four side walls and the H4 PCB, so the SATA data and power cables cannot reach to the 2.5" SSD.

fvolk
Posts: 945
Joined: Sun Jun 05, 2016 11:04 pm
languages_spoken: english
ODROIDs: C4, H3, M1S
Has thanked: 0
Been thanked: 155 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by fvolk »

odroid wrote:
Fri May 31, 2024 9:16 am
If you don't use a Net-Card, there seems to be enough space for a 7mm thick 2.5" SSD.
Ok, great! :-)
odroid wrote:
Fri May 31, 2024 9:16 am
However, there is no gap between the four side walls and the H4 PCB, so the SATA data and power cables cannot reach to the 2.5" SSD.
Oh, no :-(

Then... mmmhh.... Case 1... put case vertical, no fan, place SSD in fan position.
Unfortunately, the heat sink would radiate/warm the SSD constantly - probably not a good idea.

If I only had access to a laser cutter - build my own mini-tower case...

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

odroid wrote:
Fri May 31, 2024 9:16 am
However, there is no gap between the four side walls and the H4 PCB, so the SATA data and power cables cannot reach to the 2.5" SSD.
That's a potentially fixable problem. For the SATA power, you can get extension cables. For the data, can't you just plug a longer data cable into the board?
fvolk wrote:
Fri May 31, 2024 2:07 pm
Then... mmmhh.... Case 1... put case vertical, no fan, place SSD in fan position.
Unfortunately, the heat sink would radiate/warm the SSD constantly - probably not a good idea.
One idea, though I don't know if it'll help, would be to remove the SATA SSD from its case and attach the bare board to the computer's case via thermal tape. The PCBs are usually a lot smaller than the 2.5" form factor cases and often just have a thermal pad between one side and their case.
fvolk wrote:
Fri May 31, 2024 2:07 pm
If I only had access to a laser cutter - build my own mini-tower case...
There are places online that you can send a CAD drawing and they'll laser cut it for you. Some even have no minimum order size.

fvolk
Posts: 945
Joined: Sun Jun 05, 2016 11:04 pm
languages_spoken: english
ODROIDs: C4, H3, M1S
Has thanked: 0
Been thanked: 155 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by fvolk »

mattbot wrote:
Sat Jun 01, 2024 6:47 am
That's a potentially fixable problem. For the SATA power, you can get extension cables. For the data, can't you just plug a longer data cable into the board?
I think odroid meant the cables need to be routed outside and back in, or I would have to cut a hole at the side.
mattbot wrote:
Sat Jun 01, 2024 6:47 am
would be to remove the SATA SSD from its case and attach the bare board to the computer's case via thermal tape.
Hmm... that I hadn't considered.
Although a 870 EVO 4TB should be quite large.

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models

Post by mattbot »

fvolk wrote:
Sat Jun 01, 2024 1:28 pm
Although a 870 EVO 4TB should be quite large.
Instead of just guessing, you'll find most reviews actually open them up so you can see for yourself.

Anandtech's review shows it right in the headline banner!

Image

User avatar
mad_ady
Posts: 12102
Joined: Wed Jul 15, 2015 5:00 pm
languages_spoken: english
ODROIDs: XU4 (HC1, HC2), C1+, C2, C4 (HC4), N1, N2, N2L, H2, H3+, Go, Go Advance, M1, M1S
Location: Bucharest, Romania
Has thanked: 663 times
Been thanked: 1323 times
Contact:

Re: The ODROID H-series is growing with three brand new models (ODROID H4/H4+/H4 Ultra)

Post by mad_ady »

Man, they're tiny! I had opened an old, broken 50GB SSD from a decade ago, and it took up the whole case.

User avatar
mctom
Posts: 4330
Joined: Wed Nov 11, 2020 4:44 am
languages_spoken: english, polski
ODROIDs: XU4, M1, H3+, SP3, N2L, M1S, N2
Location: Gdańsk, Poland
Has thanked: 642 times
Been thanked: 875 times
Contact:

Re: The ODROID H-series is growing with three brand new models (ODROID H4/H4+/H4 Ultra)

Post by mctom »

Years ago there were SATA SSDs without cases available on AliExpress for cheap.
Punk ain't no religious cult, punk means thinking for yourself!
OpenUPS
PiStackMon

mattbot
Posts: 125
Joined: Fri Jan 08, 2016 10:35 am
languages_spoken: english
ODROIDs: N2+
Has thanked: 64 times
Been thanked: 12 times
Contact:

Re: The ODROID H-series is growing with three brand new models (ODROID H4/H4+/H4 Ultra)

Post by mattbot »

The product listing page says:
We used our SmartPower3 (see https://www.hardkernel.com/shop/smartpower-iii/) to test and measure the ODROID-H4 Ultra
power consumption while performing specific activities.
However, the Smart Power 3 page says its output voltage is "DC 3.0 V ~ Input voltage -1 V". I initially read just the "DC 3.0 V" part, but now I wonder if you're saying it handles a range from 3.0 VDC, up to the input voltage -1?

Also, what's the precision of its measurements? Note that I'm not asking about accuracy, as don't require it to be calibrated, but I rather just care about the size of the differences it can measure.

Thanks!

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models (ODROID H4/H4+/H4 Ultra)

Post by odroid »

If the input voltage of the SmartPower3 is 19Volt, the maximum output voltage is 18Volt.
We can see two digits after the decimal point, and when measuring remotely, up to three digits can be read. The error is approximately 2-3%.
For more detailed specifications, please refer to the power sensor IC datasheet. https://www.microchip.com/en-us/product/pac1933

This thread is locked.
If you have any questions or comments, please create a new topic in the appropriate sub-forum.
These users thanked the author odroid for the post:
mattbot (Thu Jun 27, 2024 4:02 pm)

User avatar
odroid
Site Admin
Posts: 42659
Joined: Fri Feb 22, 2013 11:14 pm
languages_spoken: English, Korean
ODROIDs: ODROID
Has thanked: 3758 times
Been thanked: 2123 times
Contact:

Re: The ODROID H-series is growing with three brand new models (ODROID H4/H4+/H4 Ultra)

Post by odroid »

Two Major New M.2 cards for the ODROID-H4 and ODROID H3 Series

Introduction
After the commercial success of the NetCard 2.5 GbE for the ODROID-H series, we looked into how we could provide new add-on cards to extend further the versatility of your ODROID-H3 or ODROID-H4.

The result is two new very low cost M.2 cards:
M.2 2x2 card (Compatible with ODROID-H3 and H4 series)
M.2 4x1 card (Compatible with ODROID-H4 series only)

Note: These two new M.2 cards are not compatible with the ODROID-H2 series.

The first card, the M.2 2x2 card, provides two M.2 PCIe Gen 3 x2 lanes connectors, from the PCIe Gen 3 x4 lanes of the mainboard. It allows you to use two M.2 devices: any combination of NVMe drives, Network adapters, WiFi adapters, 5G adapters, etc that leverage the bandwidth of PCIe Gen 3 x2 lanes.

The second card, the M.2 4x1 card, provides four M.2 PCIe Gen 3 x1 lanes connectors instead. It allows you to use four M.2 devices: any combination of NVMe drives, Network adapters, WiFi adapters, 5G adapters, etc that leverage the bandwidth of PCIe Gen 3 x1 lane.

Because the PCIe Gen 3 configuration (bifurcation) is embedded in Intel microcode that is merged into the BIOS bin file at build time, you need to flash a different version of the BIOS to use one of these two new cards.

The M.2 4x1 card requires the same BIOS version as the NetCard 2.5 GbE. You find this BIOS on the Wiki pages for the H3 and H4.

The M.2 2x2 card requires a new BIOS that provides a 2x2 bifurcation. You also find this BIOS on the Wiki pages for the H3 and H4.

Important Note: Make sure to select the right BIOS, for the H4 or for the H3.

Like the NetCard 2.5 GbE, these two new cards are installed under the mainboard, they slide into the mainboard M.2 connector and then are screwed to fix them solidly.

Thanks to this integration, these two new cards are compatible with all the ODROID-H3 and ODROID-H4 series. Note however that if you use tall M.2 devices, you may have to select a case that provides enough bottom vertical space.


M.2 2x2 card (Compatible with ODROID-H3 and H4 series)

You can install two 2280 size M.2 PCIe devices, and each slot has two lanes of PCIe 3.0 (2 x 8GT/s)

Image


Multiple M.2 SSDs can be installed using the JBOD (Just Bunch of Disks) concept. Since it only uses two PCIe lanes per slot, you can still access files at up to 1600~1800MB/s, albeit at half the original PCIe speed.

Image


Or you can install PCIe devices such as WiFi 6E, Ethernet, 4G/5G modem, AI accelerator TPU/NPU, etc. in M.2 form factor.
For example, you can also use a Google Coral Dual Edge TPU and NVMe simultaneously, which each use two PCIe lanes.

Image


As stated earlier, you have to flash another ESF BIOS for 2 bifurcated x2 lanes, since the default BIOS provides non-bifurcated 1 x4 lanes.




M.2 4x1 card (Compatible with ODROID-H4 series only)

You can install four 2280-sized M.2 PCIe devices, and each slot has one PCIe 3.0 lane (1 x 8GT/s)

Image


Multiple SSDs can be installed using the JBOD (Just Bunch of Disks) concept. Since it only uses one PCIe lane per slot, the original PCIe speed is reduced to a quarteris half, but you can still access files up to 800~900MB/s.

Image


Or you can install PCIe devices such as WiFi 6E, 5GbE Ethernet, 4G/5G modem, AI accelerator TPU/NPU, etc. in M.2 form factor.

Image


As stated earlier, you have to flash another ESF BIOS for 4 bifurcated x1 lanes, since the default BIOS provides non-bifurcated x4 lanes.




Typical Use Cases
These two new M.2 cards are intended for any ODROID-H4 or ODROIR-H3 who want to split off the PCIe Gen 3 x4 lanes for use with two x2 lanes or four x1 lane devices, at the expense of the max theoretical speed provided to each device.

Use Cases

* You want as much SSD space as possible:
-- The M.2 2x2 card with 2 x 4TB NVMe SSD provides 8TB of SSD space.
-- The M.2 4x1 card with 4 x 4TB NVMe SSD provides 16TB of SSD space.

* While keeping an SSD at PCIe Gen 3 x2 or x1 speed, you want to use one of several extra network adapters, a PCI WiFi adapter, a 5G adapter, a PCIe AI adapter, etc.

If your goal is to use an M.2 NVMe SSD at maximum speed (4 x 8GT/sec), keep using the PCIe Gen 3 x4 lanes fully dedicated to the NVMe SSD.


Informational Benchmark

Using a generic NVMe SSD we measured and compared the throughput you gain in each configuration, we also measured the throughput of a 128GB eMMC and a SATA SSD, all using iozone 3.

We obtained the following results:

Image


The obvious first remark is that reducing the PCIe throughput by a factor of 2 or 4 does not mean the speed of your SSD will also be divided by the factors. The reason is that your NVMe SSD also has its own limits no matter how fast is the PCIe channel(s).

The second remark is that in both cases, M.2 2x2 or M.2 4x1 card, the NVMe SSD still provides faster speeds than the SATA SSD and eMMC.


Product Compatibility
Screenshot from 2024-07-10 17-36-29.png
Screenshot from 2024-07-10 17-36-29.png (25.31 KiB) Viewed 2122 times
(1) The H2 CPU (Celeron J4115) does not support the 2x2 PCIe bifurcation.

(2) The M.2 4x1 card has a larger footprint and does not physically fit under the H2 or H3 boards. Therefore, the card cannot be fixed directly to the board like with a H4. HOWEVER, if you can find an M.2 male/female extension cable plus perform some additional DIY for mounting the card in your box or case, you can use the M.2 4x1 card with the H2 or H3.


Availability and Price

You can purchase it right now.
M.2 2x2 card (Compatible with ODROID-H3 and H4 series) : $10 https://www.hardkernel.com/shop/m-2-2x2-card/
M.2 4x1 card (Compatible with ODROID-H4 series only) : $15 https://www.hardkernel.com/shop/m-2-4x1-card/
These users thanked the author odroid for the post (total 3):
mctom (Wed Jul 10, 2024 6:23 pm) • domih (Wed Jul 10, 2024 9:36 pm) • xnd (Fri Jul 12, 2024 1:43 am)

Locked

Return to “General Topics”

Who is online

Users browsing this forum: No registered users and 10 guests