Intel studies the performance impact of Meltdown fixes

Intel and Microsoft have insisted that the performance impact of the Kernel Page Table Isolation fix for the recently-revealed Meltdown vulnerability will be much smaller than people feared. Now, Intel's released some hard numbers, and perhaps to the consternation of the "everything is 30% slower" doomsayers, the actual impact for client users on modern machines appears to be minimal.

Intel's testing was done using Skylake, Kaby Lake, and Coffee Lake CPUs on Windows 10, as well as Skylake machines running Windows 7. All of the test rigs were equipped with SSDs, although Intel also tested the Skylake-and-Windows-7 combo on a hard-drive-based system. The company put the machines through SYSmark 2014 SE, PCMark 10, 3DMark Sky Diver, and WebXPRT 2015 both before and after the vulnerability mitigation patches. The chart that Intel published presents results as percentages of relative performance compared to the pre-patch systems.

Ultimately the results trend close to the 5% mark overall. Many tests, particularly 3DMark, show a 1-2% difference or none at all. The worst-case scenario appears to be SYSMark 2014's Responsiveness test which sees double-digit performance hits on every platform except the Windows 7 setup with a hard drive (which actually sees a performance gain). WebXPRT 2015 also sees slowdowns in the 5-10% range, but otherwise the differences are pretty minimal.

The published results summarize the speed hit that client users on recent platforms are likely to see. Intel says it plans to prepare a "representative" data set of results for hardware released within the last five years that should likely cover most active systems in the field today. Other reports have placed the performance impact for big servers and datacenters much higher, though. For now, you can click here and see Intel's chart for yourself.

Comments closed
    • watzupken
    • 2 years ago

    It make sense to call for independent reviews to determine the impact on performance post the application of the fixes. Netizens can overstate the performance regression, but the information that Intel provides is likely biased towards the “not going to matter much” side.

    • dme123
    • 2 years ago

    Considering how hard Intel screws your wallet for 5% extra performance I’d say that’s still pretty significant.

    • Rza79
    • 2 years ago

    Intel is hit hard where it matters the most, frametimes.

    Check out this and scroll down to frametimes:
    [url<]https://www.computerbase.de/2018-01/meltdown-spectre-amd-intel-benchmarks/#abschnitt_kaby_lake_unter_windows_10[/url<] Ryzen: [url<]https://www.computerbase.de/2018-01/meltdown-spectre-amd-intel-benchmarks/#abschnitt_ryzen_unter_windows_10[/url<] Intel, 8% loss on frametimes. Ryzen, 1%.

    • quaz0r
    • 2 years ago

    “Intel studies the performance impact of Meltdown fixes”

    Please. That’s like saying your heroin dealer studied the impact of cutting your heroin with rat poison.

    “perhaps to the consternation of the ‘everything is 30% slower’ doomsayers, the actual impact for client users on modern machines appears to be minimal.”

    Did someone from TR actually write this or is this copy-pasted from Intel? It’s hard to even count the number of offensive things in just this sentence alone, not the least of which the corporate talking point that classifies “client users” as grandmas who just check their email and possibly do some light web browsing. Yeah, I could replace my grandmother’s cpu with a 486 and she could still check her email. That obviously is not and should not be the point here. The point here is that we all have defective processors which cannot be fixed, and until we have the opportunity to buy new processors at some unknown point in the future which actually have this issue properly addressed in the design of the processor itself, we have to eat a kludgy software *workaround* (not “fix”).

    As for “perhaps to the consternation of doomsayers,” wow. I mean really, Wow. To try to turn such an epic cockup back around on the CUSTOMER, offering INSULTS and the insinuation that if they are upset about this, they must be a conspiracy theorist or doomsayer who would be CONSTERNATED at good news, just wow.

    Come on TechReport, we expect better from you.

      • Freon
      • 2 years ago

      The quote you pull out and have issue with is simply a statement of fact, not an opinion or editorial on how valid the fact may be.

      While it doesn’t mean we can’t question Intel’s motives, you’re jumping on TR for just reporting the news. Until TR actually runs some tests I don’t think they should be editorializing anyway.

      Calm down.

      • Rapster
      • 2 years ago

      These folks are the most honest ones in the biz. Really, it’s not all a conspiracy and you likely won’t notice any performance difference unless you’ve got an unusual or corporate workload. Chillax bro.

      • egon
      • 2 years ago

      Fair point about the [i<]"everything is 30% slower" doomsayers[/i<] label. One of the enemies of honest discourse is to misrepresent the opposing argument by focusing on the weakest or most extreme part. But responding with strident rhetoric isn't helping to set the bar where you think you want it.

      • Voldenuit
      • 2 years ago

      And still no incoming security fixes for systems running older than Skylake processors from the major motherboard makers. That’s potentially tens, if not hundreds of millions of PCs out there.

      What about PoS, ATM and embedded systems? Will those ever see a fix?

      The performance penalties are a necessary evil to address the vulnerability(ies), but as far as I’m concerned they are a bare minimum standard that is not being met by OEMs and vendors at this point in time.

        • thecoldanddarkone
        • 2 years ago

        It’s almost certain many oem pc vendors will patch as far back as Haswell.

        Custom motherboard manufacturer’s, not so much.

    • tfp
    • 2 years ago

    This is all well and good but what is the impact to the Core2 Quads? When will they fix this the right way in HW vs a micro code update the slows things down? No reason to upgrade until they do.

    • MetricT
    • 2 years ago

    It is completely believable that consumer-level hardware won’t be too affected by this bug. The only apps that really push performance are games (which avoid syscalls) and video compression. If the penalty is 5%, users won’t even notice, as CPU’s are insanely fast for most other real-world purposes (office apps, email, browser)

    It is completely *unbelievable* that server-grade hardware won’t be affected by this, especially in IO-rich areas (storage, high-performance networking), and Intel isn’t fooling anyone here. When my 2000 node compute cluster loses 10% of its performance and my storage array really does lose 30%, that’s completely unacceptable.

    • Bauxite
    • 2 years ago

    I smell some big cow manure:

    Sky, Kaby and Coffee cores showing varying results? The actual x86 cores for all those are identical. Intel wouldn’t be playing some shenanigans with clock speeds, timings or similar dumb tricks now would it?

    :thinking: Hmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm

    Meanwhile other places show massive suffering when [i<]fully[/i<] mitigating every process on the system: [url<]http://lists.dragonflybsd.org/pipermail/users/2018-January/335643.html[/url<]

      • jihadjoe
      • 2 years ago

      Not all workloads are the same.

      • willmore
      • 2 years ago

      FireflyBSD has very few optimizations. It also goes to excess when trying to remediate problems like this. They’re not using the suggested settings for their mitigation–they turned the knobs up another notch for extra paranoia. So, don’t expect their results to be indicitave of *anything*.

      • Freon
      • 2 years ago

      There’s a lot more to the CPU than “the actual x86 cores” even if they are identical. There’s margin of error on all testing. Your questioning the variance within Intel’s own testing doesn’t make much sense.

      I’m happy to look at third party testing, of course.

      • psuedonymous
      • 2 years ago

      [quote<]The actual x86 cores for all those are identical. [/quote<]A popular meme, but not actually true.

    • barich
    • 2 years ago

    My current desktop is a Sandy Bridge-E i7 on an Intel DX79SR desktop board. I don’t think a firmware update is forthcoming, and I’m pretty irritated because it’s more than powerful enough for what I do with it these days.

    • TwoEars
    • 2 years ago

    I heard that the hit on older machines such as the i7-920 was nasty.

      • Legend
      • 2 years ago

      Will be watching this with a W3680.

    • moose17145
    • 2 years ago

    Want to see Broadwell-E numbers…

    • fyo
    • 2 years ago

    Could TR do some of their awesome frametime plots pre- and post-patch? Preferably on a variety of systems, but even on just a single system it would be valuable. There are a lot of results posted on various forums that indicate that while average framerates are pretty much unaffected, minimum framerates have taken a considerable nosedive.

    It would be great to see some reliable data from TR to either confirm or dispel this.

      • Prestige Worldwide
      • 2 years ago

      A million times this, please. We would be eternally grateful to our benevolent TR overlords!

      • Firestarter
      • 2 years ago

      computerbase.de did some benchmarks: [url<]https://i.imgur.com/lAOCME7.png[/url<] [url<]https://www.computerbase.de/2018-01/meltdown-spectre-amd-intel-benchmarks/[/url<] they also found that the 99% percentile framerates are more affected than the averages

        • JustAnEngineer
        • 2 years ago

        Ugh. The micro-code patch has murdered the performance of my pricey 960Pro SSD. 🙁

          • Bauxite
          • 2 years ago

          I’m afraid to see what it did to optane, they are amazing for speeding up zfs arrays.

    • ronch
    • 2 years ago

    Independent tests, please. And what about AMD? How much impact do they suffer? (Yeah, suffer.)

    • exilon
    • 2 years ago

    These are Spectre and Meltdown fixes benchmarked, not just Meltdown.

    • hansmuff
    • 2 years ago

    Did any of those machines have the updated CPU microcode applied?

    • gmskking
    • 2 years ago

    I turned off Windows Updates a while ago on all my Windows 7 machines. I don’t see the point of taking anymore updates for Windows 7. I definitely wouldn’t be taking this update either way.

      • K-L-Waster
      • 2 years ago

      Yeahhh, it’s not like they’re building new exploits or security flaws for an OS that old anymore…

    • willmore
    • 2 years ago

    Them not testing Win 7 with an SSD makes me want to start folding some tin foil….

      • Redocbew
      • 2 years ago

      Meh, it’s just more PR. We already knew Meltdown probably wasn’t going to be felt heavily on the desktop, but like Mr. Funk just said if you move into the server world it can be quite a different story.

        • willmore
        • 2 years ago

        It’s not that. It’s that Win7 was supposed to have been hit harder than the newer versions and SSD based systems are supposed to be effected more as well. So, it’s almost as if they didn’t use an SSD in the Win7 system so that it wouldn’t look so bad.

          • Redocbew
          • 2 years ago

          The Win7 kernel doesn’t support the PCID and INVPCID instructions, so that part makes sense. I hadn’t heard that about SSD based systems. Perhaps there’s just less latency to hide the penalty incurred on each system call? If there was any added shenanigans going on aside from the usual stuff we often see with vendor supplied tests, then I’d expect it would be the opposite.

        • tacitust
        • 2 years ago

        Sure, but when your client computing chip business brings in almost double the revenue than your server chip business, you do everything you can to protect your biggest source of income first. The fact it tells a better story is a bonus.

      • RAGEPRO
      • 2 years ago

      They did test Windows 7 with an SSD. [quote<]Intel's testing was done using Skylake, Kaby Lake, and Coffee Lake CPUs on Windows 10, as well as Skylake machines running Windows 7. All of the test rigs were equipped with SSDs, although Intel also tested the Skylake-and-Windows-7 combo on a hard-drive-based system.[/quote<]Suppose I should have written "additionally tested" to be clearer.

        • egon
        • 2 years ago

        Also worth noting it’s not a straight 7-vs-10 comparison, as they used different SSDs for some reason – the Skylake Windows 10 test setup had a 600p while the Windows 7 one had a 540s.

          • jihadjoe
          • 2 years ago

          Either way, looks like 7 did better than 10 all around.

    • derFunkenstein
    • 2 years ago

    So glad I’m not on any of our various corporate datacenter IT teams and don’t have to worry about this in a production environment. Those guys are pulling out their hair, though. I don’t envy those guys (or lots of TR folks) at all.

      • Redocbew
      • 2 years ago

      I was just talking with a friend and saying exactly the same thing. Most everyone who has been in the business for long enough ends up taking at least one professional ass-kicking on a bad project for one reason or another. Glad to not be there myself on this one.

        • cygnus1
        • 2 years ago

        It’s the modern cost of doing business. I voted for any additional required capacity expansions (for any apps that are run that close during peak times) to come out of the security budget since it’s their fault everything is getting slowed, lol. They didn’t think that was funny…

      • Chrispy_
      • 2 years ago

      I’m not pulling out my hair; My annual budget already has a contingency quota that should cover this whole fiasco.

      It wasn’t hard to negotiate that with our financial director either. About 8 years ago I simply said “This is what I can predict I’ll need to spend this year, and I’m not a psychic so you’d better add 20% just to be safe” and that arrangement has continued ever since. The fact that I rarely go over-budget just means that if I had to exceed even the contingency budget, I’d probably get a pass, and the FD and MD both know that when I say something isn’t optional it isn’t optional.

      As long as the worst-case is a 30% performance loss, I think I can just plug the gap by retiring old hardware slightly faster than planned to get rid of any outstanding Ivy and Haswell-based hardware. Chances are it’s already nearing the end of its cost-effective maintenance plans anyway….

        • Voldenuit
        • 2 years ago

        [quote<]As long as the worst-case is a 30% performance loss, I think I can just plug the gap by retiring old hardware slightly faster than planned to get rid of any outstanding Ivy and Haswell-based hardware. Chances are it's already nearing the end of its cost-effective maintenance plans anyway....[/quote<] Dragonfly BSD guys measuring up to 64% performance loss on Haswell systems vs 34% on Skylake and 31% on Kaby in compile test: [url<]http://lists.dragonflybsd.org/pipermail/users/2018-January/335643.html[/url<] Mind you they point out that this is a very artificial scenario (with IBPB on and IBRS=2). On the recommended mitigation setting of IBPB=0 and IBRS=1, performance loss is 16% on Haswell and ~4% on Skylake and newer, which is much more acceptable.

        • derFunkenstein
        • 2 years ago

        We’re basically 100% reliant on a combination of SQL Server, IIS, and SSRS. Basically everything is IO-bound. To my knowledge we weren’t really stressing the hardware before but it’s mission-critical software for schools so if we’re stressing the hardware, we’re keeping teachers from doing their primary job.

        • Whispre
        • 2 years ago

        No hair pulling here either… we build our systems with a +25% performance headroom at a minimum… when we start getting within 20-25% of performance max in any workload, we start looking at upgrades.

        We have patched 100% of our systems as of last Saturday at midnight… our worst performance hit so far have been Hyper-V hosts and SQL VM’s… and we’re not seeing anything more than 7-10% hit which has been absorbed by our reserves.

        Nothing to panic about.

    • willmore
    • 2 years ago

    [quote<](which actually gains sees a performance gain)[/quote<] Change horses in mid stream there, did 'ya?

      • morphine
      • 2 years ago

      Fixed, thanks for the heads-up.

        • willmore
        • 2 years ago

        No worries. I only recognize that mistake because I make it so damn often. *sigh*

    • Duct Tape Dude
    • 2 years ago

    Do Haswell and all the Bridges just get a score of “[i<]now[/i<] you should upgrade?"

      • cygnus1
      • 2 years ago

      I haven’t noticed any performance hit on my 4770k. Still nothing out there *better enough* to justify an upgrade that would also require MB, RAM, and NVMe SSD.

      • ermo
      • 2 years ago

      Not pulling the trigger until Zen+ lands because I want to see where the 8c/16t SKUs sit compared to the 8700K.

      Until then, my 3770K will have to do.

      • Srsly_Bro
      • 2 years ago

      The bridge to Ryzen+

    • chuckula
    • 2 years ago

    Here’s the official [url=https://youtu.be/uQJ8WrKnLUs<]video response from Intel management.[/url<]

Pin It on Pinterest

Share This