Router idle power consumption: Xeon D 1521 vs. i7 8700k

[autoturk]

Hi all,

I'm trying to build a 10 GB capable router appliance running either TNSR or pfsense. I will probably start with the latter to get familiar. I am, however, a bit confused at power consumption differences between the two boards, so thought to ask and see if there were any suggestions as to which board to go with for this use case

I have the following two machines, both ITX form factor. All are measured without anything plugged into NIC ports and in Ubuntu 22.04:

Consumer

• Intel i7 8700k
• EVGA h370 ITX motherboard - factory default bios settings
• Intel X520-DA2 SFP+ NIC
• Idle power: 16 Watts, 9 Watts without NIC.

Server

• Supermicro X10SDV-4C-TLN2F with Xeon D 1521 - factory default bios settings
• Built in 10 GB Intel X552 RJ45 NICs.
• Idle power: 30 W - 8 Watts IPMI = 22 "Net" watts
• IPMI Cannot be turned off from what I can tell.

Everything else between them is the same (8 GB non-ECC UDIMM RAM, same SSD set up, Consumer SFX power supply, etc.)

I'm a bit surprised with the increased power consumption of the server board. I know that copper RJ45 connections consume more power, so can that be attributed to the difference in consumption? I assumed that the server board would be way lower power given that it's a Xeon D and has very few bells and whistles (e.g. no audio), so I am a bit surprised.

Whichever board doesn't get used as a router will end up being a Proxmox host running a NAS, some docker images, and a Plex server. Any suggestions on which way I should go here?

 

[Sean Ho]

That sounds reasonable; Xeon-D (especially older gen) are not all that low-power compared with modern processors, only in comparison to E5 of same gen. And the 10GbaseT ports do draw power even when not in use.

As a side note, in TNSR if any port is polling, the cores pinned to the receive queue will always be at 100%.

 

[autoturk]

That sounds reasonable; Xeon-D (especially older gen) are not all that low-power compared with modern processors, only in comparison to E5 of same gen. And the 10GbaseT ports do draw power even when not in use.

As a side note, in TNSR if any port is polling, the cores pinned to the receive queue will always be at 100%.

Thanks. that's helpful, though a bit disappointing (due entirely to my fault) that the Xeon D board is not as efficient as I thought. Any thoughts on using it as the Proxmox server? I imagine it'll be good at that since I can passthrough devices easily vs. the consumer board.

As for TNSR polling: do you mean that if TNSR is looking for a connection on a port, the cores will be running at 100%? E.g. an unconnected port will cause a core to go to 100%?

 

[pngu]

As for TNSR polling: do you mean that if TNSR is looking for a connection on a port, the cores will be running at 100%? E.g. an unconnected port will cause a core to go to 100%?

That's due to the way VPP (what TNSR is built on) works. It basically "steals" an entire NIC from the kernel and then polls it as fast as it can, 24/7 (as opposed to letting the card raise an interrupt when it receives data, which would then prompt the kernel to start processing). This youtube video explains the core concepts pretty well: