Saturday, March 4th 2023
Intel Releases Windows Workaround and Patch for Ethernet Stuttering and Disconnects
At the end of January, we reported on issues with Intel's i226-V 2.5 Gbps Ethernet controllers, which are used on a wide range of motherboards with both Intel and AMD CPUs, where users were having issues with stuttering and connections dropping. Intel has now released a workaround and a patch for the issue, not only for the i226, but also the i225 and the Killer E3100 2.5 Gbps network controller. What the workaround does, is disable the Energy-Efficient Ethernet mode, or EEE as it's also known as. EEE is only supposed to kick in when an Ethernet connection is idle and it's said to reduce power consumption by up to 50 percent.
However, in this case, it seems like EEE kicks in even when the Ethernet connection is active, which is causing the stuttering and connection dropouts. The patch disables EEE for all speeds above 100 Mbps, but it's also possible to disable EEE manually in the device settings in Windows. MSI is the first motherboard manufacturer to have released an updated driver on its site, but as this isn't a final solution to the problem, it's unlikely that this is the last we'll hear about the issue. Intel is apparently working on finding the root cause of the problem, but it's unclear if it's going to be possible to patch it in software or if a hardware revision is going to be required in the end.
Update Mar 4th: User @lovingbenji reports that on his system this new driver version does not fix the disconnect issue.
Sources:
Intel Community, via Tom's Hardware, MSI driver update
However, in this case, it seems like EEE kicks in even when the Ethernet connection is active, which is causing the stuttering and connection dropouts. The patch disables EEE for all speeds above 100 Mbps, but it's also possible to disable EEE manually in the device settings in Windows. MSI is the first motherboard manufacturer to have released an updated driver on its site, but as this isn't a final solution to the problem, it's unlikely that this is the last we'll hear about the issue. Intel is apparently working on finding the root cause of the problem, but it's unclear if it's going to be possible to patch it in software or if a hardware revision is going to be required in the end.
73 Comments on Intel Releases Windows Workaround and Patch for Ethernet Stuttering and Disconnects
Both Icelake's and Zen 4's AVX-512 implementations are done with the multiple 256 bit hardware designs.
Besides SIMD width expansion, AVX-512 introduced 32 registers programming model like on RISC based PowerPC.
Took a year but finally got it sorted.
The "fix" happened in 2020, but it doesn't matter anyway since whether the use cases fit you or not it's still a feature that was removed by Intel.
This fault is years old and you finally acknowledge and patch the issue, except it's apparently still broken!
Intel NICs used to be the gold standard. Now I actually look for Realtek!
Latest bios
This driver version
EEE disabled since I built the system.
problem still exists, I’ve purchased a 2.5gb pcie Realtek nic. I’ve given up.
Well, at least not a single disconnect today, but usually 2-4 times per day.
Give it a try
downloadmirror.intel.com/772070/Wired_driver_28.0_x64.zip
Well, zero issues here. Both NICs never gave me problems. And now I’m happy that asus cheaped out on my B550 mobo by going with a Realtek 1gbps vs other B550 with the intel 2.5.
I do agree though that it messes up software devs with the abandon mindset.
Edit: The ASRock X570 Steel Legend indeed has the i211.
Intel BIOS Updates Officially Disable Alder Lake's AVX-512 Capabilities for example
They tend to bring them back in later products, but it does come off as quite a risk to buy intel hardware without knowing what supported feature is going to get suddenly cut away 6 months later
Now, do i think disabling EEE if 2.5Gb is detected to be a major issue? Not really. This ones trivial, as a hardware revision will permanently fix the issue soon enough and percentage wise very few ethernet adaptors globally will be using extra power without it - and even then i'm pretty sure the wattage on ethernet chips is quite low
Zen 4 is missing Tiger Lake's VP2INTERSECT.
Zen 4 is missing Alder Lake's AVX-VNNI VEX encoded for AVX2. Zen 4 supports VNNI at AVX-512 EVEX encoded a.k.a AVX512-VNNI.
Intel made a mess after AVX2 extensions.
AM5's CPU support can evolve longer when compared to the dead-end LGA1700.
And before that, K6-2 didn't have an instruction that Pentium IIs had!
Uggh, looks like history repeating itself again!
AVX512-VNNI EVEX encoded has a 32-registers programming model and up to 512-bit register size. AVX512's EVEX encoding has support for AVX-128 EVEX, AVX-256 EVEX, and AVX-512 EVEX register sizes. Intel disabled the entire AVX-512 EVEX encoding feature on Raptor Lake while AMD's Zen 4 supports IceLake AVX-512 + BF16 extension support.
Both are Intel-originated AVX extensions that are incompatible with each other.
Zen 1.0 has quad 128-bit SIMD units.
Zen 1.5 has 12 nm refinements from 14 nm.
Zen 2.0 has quad 256-bit SIMD units.
Zen 3.0 has quad 256-bit SIMD and two extra FPU units expansion. Load-Store unit count expansion. Some AVX-512 instruction support e.g. AVX-512 Vector AES, AVX-512 Carry-Less Multiplication Quadword.
Zen 3 X3D cache. L3 cache expansion.K6-2 has 64-bit 3DNow SIMD (64-bit hardware implementation) which includes Pentium II's MMX SIMD. K6-2 doesn't support Pentium III's SSE (64-bit hardware implementation).
Pentium III's 128-bit SSE is implemented on 64-bit SSE hardware until Core 2's full 128-bit bit SSE hardware.
K8 has 128-bit SSE FADD and 64-bit FMUL SIMD hardware. Pentium IV has 64-bit SSE hardware implementation.
K10 has 128-bit SSE FADD and 128-bit FMUL SIMD hardware with K8's three instructions per cycle decoder limitations. Core 2's full 128-bit bit SSE hardware has four instructions per cycle decoders.
The instruction set may not follow actual hardware implementation.
PowerPC's Altivec 128-bit SIMD competition has 128-bit hardware implementation at PowerPC G4's introduction i.e. no bulldust 64-bit SIMD hardware with fake 128-bit SIMD instruction set.