Intel Ice Lake "0x78" Microcode Update Having Small But Measurable Performance Hits

[kcrudup]

Michael , et al. : Apparently there's also an important fix in this ucode as well- it fixes an issue with either the IOMMU or its related processing.

I used to have issues if I'd enabled IOMMU in the kernel on my Dell XPS 13 2-in-1 (i7-1065G7 CPU). When suspending, randomly the machine would lock up going into the S3 state and drain the battery, requiring a reset when I'd try to resume later. (If I booted without IOMMU enabled in the kernel, suspend/resumes were rock-solid, but I'd prefer to have IOMMU on as I have Thunderbolt). With every IOMMU commit in Linus' master I'd try turning it on again, no change. When I got the 0x78 ucode version I'd tried it again, and it works now, tested over many cycles.

I regressed back to 0x46 to verify, and sure enough a couple of suspend cycles in, the issue recurred. I'm not happy about the performance hit you'd found, but I much prefer the reliability.

[NotMine999]

I am waiting for a book which analyzes in detail this fall of Intel and rise of AMD.

I don't think Robert X. Cringely is up to writing another book.

[smitty3268]

Ice Lake core is 18% better in IPC. Say they lose at worst 5% with this microcode (to me it looks more like 1-2%) and they still have a 15% advantage over skylake, so they don't need to do anything to make Ice lake "look better" than previous gen.

That's only true if they can hit the same clockspeeds on each, and to date they definitely can't. Whether that's the Ice Lake architecture or the 10nm manufacturing process, it doesn't really matter.

18% faster than a 3.9Ghz turbo boost is only 4.6Ghz.

[caligula]

Ice Lake core is 18% better in IPC. Say they lose at worst 5% with this microcode (to me it looks more like 1-2%) and they still have a 15% advantage over skylake, so they don't need to do anything to make Ice lake "look better" than previous gen.

Skylake isn't the previous generation. Apparently you forgot not only Kaby Lake + refresh, but also Coffee and Whiskey Lake.

[hotaru]

IceLake requires the same workarounds as Zen 2. Stop being an a-hole.

from that link:

• Ice Lake requires mitigation for CVE-2018-3665, Zen 2 doesn't.
• Ice Lake requires mitigation for CVE-2019-1125, Zen 2 doesn't.

[carewolf]

IceLake requires the same workarounds as Zen 2. Stop being an a-hole.

I was talking about SkyLake, not IceLake. Hence the honest question. Skylake lost something like 20+% in IPC in its lifetime due to runtime bug-fixes. So if IceLake mostly fixed all of that, it should be a lot faster in total without even getting into architectural improvements.

[birdie]

18% better IPC including all the workarounds needed for Skylake or before?

IceLake requires the same workarounds as Zen 2. Stop being an a-hole.

[yeeeeman]

You should consult Anandtechs review on this. I think they had fixes for 9900K in place when they compared the IPC to 1065G7.

[carewolf]

Ice Lake core is 18% better in IPC. Say they lose at worst 5% with this microcode (to me it looks more like 1-2%) and they still have a 15% advantage over skylake, so they don't need to do anything to make Ice lake "look better" than previous gen.
They need to fix and patch absolutely every single vulnerabilities and loop hole, even with the enquired performance penalties. Then, in the next core revisions they need to come up with ways to fix them in hardware and not lose performance. The "vulnerability craze" needs to end...

18% better IPC including all the workarounds needed for Skylake or before?

[moilami]

I am waiting for a book which analyzes in detail this fall of Intel and rise of AMD.