OpenVPN on the BCM4912 (ZenWiFi Pro XT2 or GT-AX6000)

[maxbraketorque]

Yes, any WG implementation is incompatible with hardware NAT

Can this be remedied with changes to WG and/or the NAT acceleration code?

 

[N/A]

Not sure this is true, as my CPU idles at around 82C, pretty similar to my previous router.

What your experienced is a typical ASUS design choice/flaw. They put a decent sized aluminum heatsink on the chip but with a really thick and thermally non conductive thermal pad. IIRC they do this to reduce surface temperature on the exterior of the router and create an illusion of "good cooling" like some laptop manufactures. Simply replace the thermal pad will drop 20 or 30C.
If the chip can't dissipate heat then it's going to be hot regardless of the wattage.

 

[RMerlin]

Can this be remedied with changes to WG and/or the NAT acceleration code?

WG uses its own protocol, which flow cache and packet runner cannot handle.

The likelihood of Broadcom implementing support for that protocol is slim to none.

What your experienced is a typical ASUS design choice/flaw. They put a decent sized aluminum heatsink on the chip but with a really thick and thermally non conductive thermal pad. IIRC they do this to reduce surface temperature on the exterior of the router and create an illusion of "good cooling" like some laptop manufactures. Simply replace the thermal pad will drop 20 or 30C.
If the chip can't dissipate heat then it's going to be hot regardless of the wattage.

That would also be the case for my previous router. My point was that I doubt there was a significant power usage drop with the BCM4912 as reported temperatures are not really changed from a BCM4908-based model. A poor thermal interface would still report a similar delta.

 

[N/A]

That would also be the case for my previous router. My point was that I doubt there was a significant power usage drop with the BCM4912 as reported temperatures are not really changed from a BCM4908-based model. A poor thermal interface would still report a similar delta.

I think the best way to figure out the power is to get a data sheet/ask ASUS or measure with multimeter. From the spec available from Broadcom’s website, there isn’t much difference between 4908 and 4912. 0.2GHz bump of the same CPU, DDR4 memory controller, 2x10Gb PHY, one extra channel of PCIe won’t take too much more power but I might be wrong. The original quote from ARM says “TSMC’s 16FinFET process offers significant improvement over 28HPM for high end mobile computing and networking. Since designs could gain >40% faster speed at the same total power, or alternatively reduce >55% in total power at the same speed over 28HPM, it made sense to use this process to implement a more complex test chip with ARM’s Cortex-A57 and Cortex-A53.” Source. Both 4908 and 4912 use B53 which is a variant of A53 so the quote should be applicable in this case.

 

[Yota]

Most likely huge improvements on power and temperature. BCM4908 is built on (fairly sure it's TSMC because the other fab with 28nm is samsung and broadcom usually use TSMC) 28nm while BCM4912 is built on (only TSMC markets as 16nm so it must be them) 16nm. The density of the process almost doubled from 15 to 29. Given that BCM4912 is mostly a refresh of BCM4908 with only 0.2GHz frequency bump, you can expect BCM4912 to consume about 45% the power of BCM4908.

The BCM4912 looks very similar to the IPQ8074, also a 2.0 GHz quad-core Cortex A53, the difference is that Qualcomm uses 14 nm FinFET process. not to mention, qualcomm also has an IPQ8078 with a frequency of up to 2.2 GHz, and these chips were shipped in the Q4 of 2018 and the Q1 of 2019.

While we have to love Broadcom, but I must say that Qualcomm is the future.

Not sure this is true, as my CPU idles at around 82C, pretty similar to my previous router.

That's what I'm worried.

If the chip can't dissipate heat then it's going to be hot regardless of the wattage.

I have some very entry level Asus routers, they are DSL routers, they use MediaTek CPUs, once, I disassembled one and found nothing inside, no shield, no heatsink, nothing, but not hot. Although these CPUs are all less than 1 GHz, they are cool compared to 800 MHz/600 MHz CPUs like the BCM4708A0/BCM4706. I turned it on for 3 hours, and my fingers directly touched the CPU, but it still didn't get hot.

That would also be the case for my previous router. My point was that I doubt there was a significant power usage drop with the BCM4912 as reported temperatures are not really changed from a BCM4908-based model. A poor thermal interface would still report a similar delta.

Only use the watt meters to find out.

I think the best way to figure out the power is to get a data sheet/ask ASUS or measure with multimeter. From the spec available from Broadcom’s website, there isn’t much difference between 4908 and 4912. 0.2GHz bump of the same CPU, DDR4 memory controller, 2x10Gb PHY, one extra channel of PCIe won’t take too much more power but I might be wrong. The original quote from ARM says “TSMC’s 16FinFET process offers significant improvement over 28HPM for high end mobile computing and networking. Since designs could gain >40% faster speed at the same total power, or alternatively reduce >55% in total power at the same speed over 28HPM, it made sense to use this process to implement a more complex test chip with ARM’s Cortex-A57 and Cortex-A53.” Source. Both 4908 and 4912 use B53 which is a variant of A53 so the quote should be applicable in this case.

I never figured out why Broadcom was so hot I'm not worried that the CPU will burn itself, I just think it's a waste of energy. I know that BCM4708/BCM4709 have no power management, so their CPU always runs at 100% frequency. BCM4906/BCM4908 seems to introduce CPU wait and some switch power saving strategy. But I think the CPU still maintains a high frequency when idle. that's why they are still hot. When they get hot, that means electricity is being converted into useless heat, which to me is a big waste of energy considering the CPU is sitting idle a lot of the time. I don't use a fan to cool the CPU in the summer, because that's more energy wasted.

 

[N/A]

The BCM4912 looks very similar to the IPQ8074, also a 2.0 GHz quad-core Cortex A53, the difference is that Qualcomm uses 14 nm FinFET process. not to mention, qualcomm also has an IPQ8078 with a frequency of up to 2.2 GHz, and these chips were shipped in the Q4 of 2018 and the Q1 of 2019.

While we have to love Broadcom, but I must say that Qualcomm is the future.


That's what I'm worried.


I have some very entry level Asus routers, they are DSL routers, they use MediaTek CPUs, once, I disassembled one and found nothing inside, no shield, no heatsink, nothing, but not hot. Although these CPUs are all less than 1 GHz, they are cool compared to 800 MHz/600 MHz CPUs like the BCM4708A0/BCM4706. I turned it on for 3 hours, and my fingers directly touched the CPU, but it still didn't get hot.


Only use the watt meters to find out.


I never figured out why Broadcom was so hot I'm not worried that the CPU will burn itself, I just think it's a waste of energy.

I usually prefer Qualcomm over Broadcom but in this case I wouldn't say Qualcomm is better. Both BCM4912 and IPQ8078A are 4 core A53 at about 2GHz with DDR4 and PCIe 3.0. Broadcom uses TSMC 16nm while Qualcomm uses Samsung 14nm. These two processes have roughly same transistor density. However we have a perfect canadiate for comparasion which is Apple A9. It was manufactured on these two processes simultaneously and the TSMC one was better. Marketing number of the process does not equal to reality since many years ago. TSMC has been the industry leader for the last few process nodes against Intel and Samsung.

Just because the frequency is the same does not imply the power is same. The MediaTek CPUs you referred are most likely these low power MIPS ones with much lower performance/power/transistors compared to the MIPS BCM4706/ ARM BCM4708.

Watt meters are not very useful in this case because there are other chips beside the CPU that we can't disable.

 

[Yota]

I usually prefer Qualcomm over Broadcom but in this case I wouldn't say Qualcomm is better. Both BCM4912 and IPQ8078A are 4 core A53 at about 2GHz with DDR4 and PCIe 3.0. Broadcom uses TSMC 16nm while Qualcomm uses Samsung 14nm. These two processes have roughly same transistor density. However we have a perfect canadiate for comparasion which is Apple A9. It was manufactured on these two processes simultaneously and the TSMC one was better. Marketing number of the process does not equal to reality since many years ago. TSMC has been the industry leader for the last few process nodes against Intel and Samsung.

Just because the frequency is the same does not imply the power is same. The MediaTek CPUs you referred are most likely these low power MIPS ones with much lower performance/power/transistors compared to the MIPS BCM4706/ ARM BCM4708.

I wouldn't use Apple chips as a good comparison because they are designed for completely different purposes, and Apple usually doesn't provide the public with the technical details of their CPUs, they just say "up to 100000000000% faster CPU performance than previous"

I know TSMC is ahead in this area, and I think the overheating of Broadcom chips has more to do with their design and power management strategies, not the process.

Watt meters are not very useful in this case because there are other chips beside the CPU that we can't disable.

Yes, but when the CPU adopts new generation technology, other chips are usually also new generation technology. So I would use a wattmeter to check the power consumption of the router, after all I need to care how much energy is consumed by router, not just how much energy is consumed by that chip.

 

[L&LD]

The number used in the process node is just marketing.

Intel is the leader here. Even if they took a while to get to the top again.

There is not one number that can fully describe the process node of a chip, that is why Intel has changed their naming designations recently. Still not directly comparable to others, but still the better technology, overall.

Many articles about this online.

 

[Tech9]

They put a decent sized aluminum heatsink on the chip but with a really thick and thermally non conductive thermal pad.

There are few possible and related reasons why is Asus doing it. Most Asus routers have shared heatsink between the CPU and RF ICs. Single part is easier to manufacture and cheaper to assemble. They prefer to cook the CPU, because the temperature is still in specs, but the heatsink is at about 50C with no fast changes up or down making life easier for more sensitive to temperature RF ICs. The better design is separate heatsinks and proper thermal transfer material, but it makes the product more expensive to manufacture. This design makes the PCB more thermally stressed and the RoHS directives additionally lower the overall reliability, but the product life expectancy is just few years anyway and the warranty period rarely exceeds 2 years.

 

[Yota]

The number used in the process node is just marketing.

Intel is the leader here. Even if they took a while to get to the top again.

There is not one number that can fully describe the process node of a chip, that is why Intel has changed their naming designations recently. Still not directly comparable to others, but still the better technology, overall.

Many articles about this online.

It's not my major, but @RMerlin definitely has a lot to say about chips because he's a pro.

There are few possible and related reasons why is Asus doing it. Most Asus routers have shared heatsink between the CPU and RF ICs. Single part is easier to manufacture and cheaper to assemble. They prefer to cook the CPU, because the temperature is still in specs, but the heatsink is at about 50C with no fast changes up or down making life easier for more sensitive to temperature RF ICs. The better design is separate heatsinks and proper thermal transfer material, but it makes the product more expensive to manufacture. This design makes the PCB more thermally stressed and the RoHS directives additionally lower the overall reliability, but the product life expectancy is just few years anyway and the warranty period rarely exceeds 2 years.

Before the start of summer, I had purchased copper heatsinks a few weeks ago and replaced them for my three routers.

I would like to thank you especially because the information you provided here was very helpful, even though I had disassembled my router before buying the heatsink, I was unable to confirm the thickness. Just wanted to remind you that you are confusing the thickness of the CPU and RF chips.

 

[Tech9]

Just wanted to remind you that you are confusing the thickness of the CPU and RF chips.

I believe it is correct. As far as I remember the CPU gap was the larger one on this specific AC86U. The copper shim I used was a piece cut off AMD desktop CPU heat spreader and it's about 2.4mm thick. It was about 1.5y ago though and I have no intention to open it and double check.

 

[Yota]

I believe it is correct. As far as I remember the CPU gap was the larger one on this specific AC86U. The copper shim I used was a piece cut off AMD desktop CPU heat spreader and it's about 2.4mm thick. It was about 1.5y ago though and I have no intention to open it and double check.

Haha, never mind, I just replaced them a few days ago, so I'm pretty sure that the RT-AC86U WIFI chips is 11.5*12*2.5mm and the CPU is 15*15*1.5mm.

Anyway, just thanks, I may share some pics after this year's threads discussing CPU temperature comes out, I'm sure these threads are coming lol.

 

[L&LD]

Here's what Intel has been working on for a few years already. I believe they're the first customer to acquire this level of manufacturing equipment.

Intel Opens D1X-Mod3 Fab Expansion; Moves Up Intel 18A Manufacturing to H2’2024

www.anandtech.com

 

[RMerlin]

Here's what Intel has been working on for a few years already. I believe they're the first customer to acquire this level of manufacturing equipment.

Intel Opens D1X-Mod3 Fab Expansion; Moves Up Intel 18A Manufacturing to H2’2024

www.anandtech.com

Intel is notorious for missing deadlines when it comes to new node processes. See how often 10 nm (AKA Intel 7) got delayed for example.

At this time, TSMC has superior capabilities, delivering 5nm right now, and having planned 3nm as soon as 1Q23:

TSMC Roadmap Update: 3nm in Q1 2023, 3nm Enhanced in 2024, 2nm in 2025

www.anandtech.com

If things go as planned, I don't expect Intel to catch up until the second half of 2023, or 2024. Personally I expect delays, as it was recently announced that there currently is a shortage of manufacturing equipment in that sector, which may delay many recent expansion plans announced by Intel and TSMC.

 

[L&LD]

I believe Intel is well past those issues with their 10nm designs.

TSMC, Samsung, and others are not using actual/accurate node process sizes for a very long time now. Again, marketing. It is not just the node process for the transistors that matters, the interconnecting circuitry needs to be scaled similarly too. Intel is (and I believe, always) was the leader for this on a total package scale.

Intel started using 'Intel 7, Intel 4, Intel 3, Intel 20A, Intel 18A' to better indicate their position vs. the competitors. See the second sentence above.

Intel's Process Roadmap to 2025: with 4nm, 3nm, 20A, and 18A?! (anandtech.com)

Intel has more than caught up, today. In the near future (i.e. before this year is out), Intel will be firmly in the lead once more.

Either way, we can each interpret the facts as we wish, but comparing like to like is key to better conclusions.

 

[RMerlin]

In the near future (i.e. before this year is out), Intel will be firmly in the lead once more.

That's not what the TSMC short term oadmap indicates...

 

[JGrana]

WG uses its own protocol, which flow cache and packet runner cannot handle.

The likelihood of Broadcom implementing support for that protocol is slim to none.

Likely true, maybe ChaCha20 engine in the future though?

 

[RMerlin]

Likely true, maybe ChaCha20 engine in the future though?

Chacha20 is just a cipher that is applied on the packet content, it`s not tied to the protocol itself, and has no relation to the flow cache compatibility issues.

 

[um hombre]

Not sure this is true, as my CPU idles at around 82C, pretty similar to my previous router.

Seems pretty high to me, mine hovers around 65 degrees idle with ambient temperature of 20-22 degrees Celsius.

It stands on a separate shelf, though, and no obstructions so plenty of breathing space. Transmitter power set to "Fair" for both radios, don't need more power for my needs.

 

[RMerlin]

Seems pretty high to me, mine hovers around 65 degrees idle with ambient temperature of 20-22 degrees Celsius.

Mine might have a poorly seated heatsink then. It wasn't a retail unit, so it might not have gone through the same QA as a retail box product.