Microsoft backports AMD branch prediction improvement to Windows 11 23H2 | Hacker News Reader

Microsoft backports AMD branch prediction improvement to Windows 11 23H2

tomshardware.com

41 pointsmariuz47 comments

Quite unbelievable how AMD has botched the Zen5 release, especially since there was no need in rushing these CPUs out the door at all. Why couldn't they wait until these improvements were commonly available? Also, at least from initial testing, it seems that these changes improve Zen4 performance almost the same way as Zen5, so at least when it comes to gaming performance, there's currently no reason to expect that Zen5 will be able to justify its increased price tag. I can safely say that buying a 5800X3D was one of the best purchases I ever made, I still see no reason to upgrade...

It doesn't seem to be even directly related to the Zen 5 launch, as the problem affects Zen 4 equally, and it looks like Windows has been leaving performance on the table for years, while AMD were completely unaware of the fact.

This only came to light as a result of investigation by Wendell from Level1Tech, who noticed the discrepancy between AMD-provided Zen 5 performance numbers with his own, as well as the performance variability between different BIOS settings. Meanwhile, AMD has been running their internal benchmarks using Admin account, and recommended the same in the official reviewers' guide. Not sure if such ignorance makes them look better or worse in this entire mess.

At least the initial reviewers' guide did not mention Admin account at all. And BTW, just to be clear: this does not mean "Account with administrator privileges", which would be normal for testing, but the (normally hidden) "administrator" account which has UAC completely disabled and hence should never be used for any real work. Why would you even test your hardware with such an obscure setup that no actual user will be using? Note that this account is not readily available but needs to be explicitly activated in an admin command prompt with

net user administrator /active:yes

It could be just easier for AMD's automated tests

Choosing to run specific applications with elevated privileges is enough to get the performance gains for those. Still not safe, but much safer than logging in with the Administrator account.

The 5800X3D isn't even 2,5 years old, of course an upgrade isn't worth its money. To put it into perspective: If you considered upgrading from a RTX3070 to a RTX4070, the RTX4070 wouldn't even have been released. (Time span between them was longer than 2.5 years.)

DDR4 to DDR5 is a pretty big jump, though. Up to 20-30% on Intel depending on what you're doing.

Of course it might not be that significant on AMD which (IIRC) struggles with higher frequencies anyway and at least for gaming memory speed doesn't matter as much cause of the extra cache.

Whether an upgrade is worth it depends on your workload and how much you depend on it. All I can say is that if Zen5 would have brought the same improvement over Zen4 as Zen4 did over Zen3, I would definitely have considered upgrading. But it seems it's nowhere near it. So yes, maybe AMD just made the mistake of performing too well in the past, and it probably was just a matter of time until they hit a wall. The problem really is how AMD has marketed Zen5, with numbers that simply couldn't be reproduced by reviewers, so it really leaves a bad taste.

In the exact same boat with the 5800X3D myself. It's starting to feel like a modern i5-2500k, a CPU that will just keep going and going. I built this machine not long after the 7000 series released and no regrets going for the older platform.

It's not unbelievable when you know AMD's drivers have always been hit or miss since the ATI days, with more misses than hits.

Fans call their driver development "fine wine" since they keep getting better over time, while realists call it "not being able to release good drivers from the start, unlike the competition, needing months to reach parity".

Yeah, their HW is good, but that's nothing without good SW and AMD still has loads of room for improvement there.

AMD being absolute crap on the software front is where Nvidia simply dominates and Intel still makes for a reasonable purchase.

I have a laptop with AMD and NPU. AMD still has no good easy to use out of the box examples on how to use that NPU for on-device AI applications. Nvidia has been delivering that for over 10 years now. SW seems like an afterthought at AMD.

If AMD expects their chips to take off for AI, they should be spamming us with ready made apps and code samples that just work and that we can then modify.

Much like AMD/ATI, Intel has had botched 3d graphics drivers since at least GMA 900 (probably earlier but that's as far as I remember). Try the new Intel Arc card and see how it works for you.

Except that AMD was never competing with Intel for graphics. Their competition was always Nvidia who's drivers and SW support was net superior. Intel was happy to just serve the ultra low end market with integrated GPUs for rendering MS Office back when AMD didn't even have integrated GPUs in it's chips.

Also, on this tangent since you brought up Intel, their Linux driver support is net superior to that of AMD. I have laptops with Intel CPUs (8th gen, 6th gen and 3rd gen) and Ryzen (5000 and 7000 series) and on the Intel ones linux is way more "Just Works" than on AMD. Sample zise of one person of course.

I guess the grass is always greener on the other side. My experience with Intel GPUs on Linux is features getting progressively disabled in newer kernel versions due to hardware bugs (FBC, PSR & DC6 C-state on my Goldmont laptop for example).

My favourite bit is how the intel xorg driver supported TearFree for years, then got sunset and everyone was told to migrate to the modesetting driver. Except there is no TearFree in the modesetting driver (just use a compositor!). It eventually landed in xorg master last year [1], but no one is making releases of it any more.

[1] https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests...

Just an anecdote, but the Arc A380 on my kids’ Fedora gaming PC has run everything on our Steam library just fine. A mix of older and more recent titles. I’m sure there are games that don’t work well, but it was cheap and it’s been rock solid.

Zen5 was released less than 2 weeks ago, so not a terrible delay.

If you already have Zen3, then yeah, probably not worth it. I like to wait 4-6 years so the upgrade feels more substantial. Every 2 years hasn't been necessary for at least a decade.

Apparently Zen4 also benefits from this patch - Hardware Unboxed had an excellent video where they did lots of benchmarking.

It's a terrible delay w.r.t. reviews, because all the reviews were done without these improvements. First impressions are very hard to fix, and Zen5 is already widely regarded as a flop, at least for desktop/gaming usage.

Just ordered 5700X3D... So I will be going on with that for a while. Maybe I see the 11700X3D or 11800X3D... Not feeling current gent GPUs either mostly due to pricing...

This doesn't make a lot of sense off the bat, are there more details to the story that explain how CPU branch prediction performance depends on Windows behaviour somehow?

Dave of Dave's Garage explained it here: https://www.youtube.com/watch?v=OBn3FbKusZI

Basically (according to him) they disabled the spectre/meltdown mitigation if you were an admin - because you were an admin and have full rights to the machine anyway. So now they are doing that for all users?

He wasn't clear - or at least it wasn't clear to me. So the mitigations are no longer needed? I think I have more questions than before after listening to that.

Dave lies. A lot. I would take anything he says with a bucket of salt.

Interesting. I just found his channel. What does he lie about? Factual stuff or MS history?

Just the latest one.

https://x.com/davepl1968/status/1812559735616577565

He claims he wrote realtime rendering of vertical text in start menu.

Turns out he didn't. No released or unreleased version or source code show that.

That logic doesn't pass a sniff test at all.

Linux has different performance schedulers you can choose that can dramatically change performance dependant on workloads. It's very possible the CPU is being too smart for its own good and the way Windows is lining stuff up it's not executing branches or something like that.

This is just shooting from the hip, but there is a ton of different stuff going on in both the OS and the CPU around this.

The mitigations for Spectre/Meltdown involved deliberately breaking branch prediction across the user/kernel boundary, so it may be as simple as them disabling some mitigations which turned out to not be needed anymore on newer CPU architectures.

Did Intel's stock drop as a result of these AMD performance improvements?

Honestly, they are huge improvements and really put the nail in Intel, now it has an even higher road to climb to get competitive again.

But Intel CPUs are more human than ever, just the other day I read they are killing themselves due to being overworked.

Does anyone know what the technical change in this patch is?

I think of branch prediction as being something the CPU does internally; where does the operating system come into play?

I still haven't seen anywhere reporting what, exactly, was the issue, and why would using the "true Administrator" account on Windows (or using Linux) avoid the issue.

My own uninformed guess, given that the "true Administrator" account avoids it, would be some sort of security mitigation, for instance flushing the branch predictor state when entering or exiting the kernel unless the current user is SYSTEM or Administrator.

If you have a hot path with a lot of branch mispredictions that can impact performance. Could be as simple as Microsoft found a dumb if inside a tight loop.

In what update is this packaged for Windows 10 and Ryzen 7800x3d?

i am curious about ryzen 9000 performance on windows 10 but nobody seems to be testing that for some reason

Windows 10 is EOL next year. The world is moving on.

Right off a cliff. I refuse to downgrade to windows 11 and when push comes to shove I will be using Linux, I already use it a lot for my projects so ditching windows won’t be hard at all.

Same. My main laptop came with 11 and won't regrade.

My desktop at home and “spare” laptop (the one that goes with me where it might get (more) damaged or lost, so I don't want to take the newer/expensive one) are both still running 11. Laptop2 keeps nagging me to “upgrade” but the desktop isn't even compatible (wrong sort of TPM?). They'll be getting the Linux treatment next year, as might the one that is currently on 11.

Though I suspect the life will get extended, like happened to XP when small-factor low-spec laptops (and to a lesser extent tablets, though few of those ran Windows or pure Linux, usually Android instead) were popular, there was no way Vista would run on them, and 8 wasn't ready enough for release. With the number of Win10 installs out there that might move to Linux, or be binned and replaced by an Apple device, it might not make sense for them to allow it to EOL without a sudden uptick in Win11 adoption. They'll need to keep supporting 10 for some large commercial and government clients anyway.

Im not so opposed to W11.

From a privacy standpoint, W10 has seen most of the crap backported already so its a bit of a wash.

Windows 11 is a bit different and does have some wonk still. But there are things integrated that are nice. I moved to an ultrawide recently and the W11 native options for things like window sizing or taskbar management are much better and don't need 3p apps like you do on w10 with the likes of PowerToys or Taskbarx.

I still use Linux day to day and may even move there for work to be honest now that I don't do a lot of IC work in windows now.

Well, there's no time like the present. What's stopping you from switching right now? If I didn't want to not have to fuss around with games I'd have already made the switch myself.

Inertia and I have some very Windows specific programs that I’m using for my current project and using the alternatives will just be slower. I’m using windows LTSM so I’m good for another few years.

That is the analogy unlike to use when people are blinded by “progress”. Sometimes you need to turn around or step on the breaks.

I have no issues with Linux for productivity, but gaming is still an issue. :-/

I really wonder if this is cueing up another situation similar to what happened with winxp, and wannacry on the many machines still using it years after support ended. There will be some that move over to win11 with varying degrees of enthusiasm or reluctance, but win10 is going to linger.

With gaming on linux, there's been a "it just works" attitude with enthusiasts for years but that rings hollow to me in a similar way that "it's just like lego" is for self-building your own PC - it ignores the small mountain of paper cut issues and gotchas it's almost inevitable someone will run into especially if you try to do something off the beaten path. Not all of gaming is neatly within the steam short-walled garden or well-behaved even on windows. It's fine if you have the attitude and time to be prepared to tackle those issues, but nowhere near where I'd want to recommend it for even a small proportion of the tens of millions of PC gamers still using win10.

I'll have you know my pirated Windows 10 LSTC Enterprise for Internet connected Japanese e-toilets is supported until 2032.

As long as my Japanese e-toilet can run Doom, we're good to go.

>Japanese e-toilets

The scientific name for them is Toto.

except it didn't? most people are still on windows 10

Reply on news.ycombinator.com