Latest Realtek NIC driver offers 10% CPU reduction on high bandwidth loads

mrk

mrk

Man of Honour
Joined
18 Oct 2002
Posts
102,784
Location
South Coast
Last updated the Realtek LAN drive back in April but never paid any attention to actual CPU usage when maxing out the bandwidth, only really noticed the higher than expected CPU use after monitoring task manager doing speedtests since going FTTP and doing a bit of research leading to info showing that the Realtek 8125 family of onboard NICs on all mobos use CPU resources. It's nothing that results in noticeable performance drop in anything, but it's still CPU usage nonetheless and a 10% saving here and there is nice to have I guess.

HWINFO64 was reset for each speedtest run, 947Mbps up and 947Mbps down, the max CPU usage at the end was recorded.

April driver:
INNisPB.png


July driver:
wEcBtcC.png
 
It's a shame some mobos come with such poor LAN chips these days.

I believe that board has a 2.5GbE capable Realtek chip but even after that driver update it could be using over half your CPU delivering that, whereas a better chip with hardware offload would have very little CPU use. For those of us using local network file servers that is a bad limitation but sadly common these days. I try to get boards with a spare slot that's still usable once a modern huge GPU is added and has enough lanes for a decent single or multiport NIC.
 
Basically all boards I have looked at in recent years come with Realtek's NICs, so this would apply to all these boards inc higher end "gaming" boards and the like. Intel's 2.5GbE NICs have their own issues as well so it seems unless you buy enterprise orientated NICs which cost £££, then you're stuck with buggy Intel cards, or CPU intensive Realtek ones.

I'm tempted to actually buy a USB 3 to Ethernet adapter to see how much CPU utilisation they have vs the onboard Realtek. The USB chipsets do dedicated processing amd adapters like the cableMatters/Ugreen etc cost less than £30 and appear to have dedicated chips for processing.

Actually I think I will order one off Amazon to put this to the test lol.

LAN transfers don't matter to me as I don't do LAN file sharing, but I have FTTP which is basically a gig up and down symmetrical so I'd still see the same CPU utilisation up/downing large files etc.
 
Last edited:
No performance is fine, there's just higher than expected CPU using when large downloads or uploads start it seems

Edit*
There is a setting in device manager called Large send offload (tcpip v4) - That's on by default, turning it off resulted in a 2% reduction in CPU use during the speedtest run. Speeds remained the exact same. Some of the other settings may reduce CPU use further too but have not played around with them much yet.
 
Last edited:
You’ll have to forgive my ignorance but from what I’m reading here, using the onboard NIC on my PC is stuffing its performance? That’s utterly scary.

If you read some of the comments of developers that have had to implement support for RealTek chips there is some useful functionality missing and/or done in software instead of hardware. I don't know if this applies to all Realtek chips but motherboard integrated network chips are probably the cheapest variants.

Modern desktop/laptop PCs generally have plenty of CPU cycles to spare but its when better NICs do more of the work in hardware that's not ideal, and for lower power devices for routing/firewalls this can make a noticeable difference. Sadly in the move to 2.5GbE even Intel has had issues with hardware and drivers (random disconnects) but I still stick to motherboards with Intel NIC chipsets generally for performance, driver reliability and support in virtualisation, or as per my comment above add a standalone quality NIC for my high network use machines.

Another issue is when I started re-purposing my old software dev PCs to home server roles I had several onboard network failures even with Intel, and I suspect its the high continual network load (CCTV/all home devices nightly backups) and the resulting temps of those chips. I've yet to have a quality standalone NIC card fail in the same PCs but then they have good size heatsinks and or fans. I doubt Realtek vs Intel and others makes much difference in the 1GbE range but as with the move to 2.5 and beyond the difference in quality/design will start to be more important.

(Note when testing - anti-virus software can add a lot of overhead so always disable)
 
Don’t get me wrong, I wasn’t complaining about the performance of our PCs, more just that I’d rather have as much of the available performance of the CPU as possible. We use large numbers of Intel Celeron appliance PCs to run our security software, Untangle. Some of those appliance PCs have as many as 12 onboard NICs (generally Intel 226V these days, but also Intel 225V3 on older appliances) and I wouldn’t want to be wasting a single CPU cycle if I could help it.
 
Back
Top Bottom