AMD Execution Thread [2024]

Well, I'd hold off on assuming none of this applies to Intel. It might very well be true, but I'd prefer someone actually test the 24H2 across a wider collection of CPUs from both major brands before making any hard statements. As I was watching the HardwareUnboxed video posted above, I was paying a lot of attention to the 1% low scores -- which I feel weren't even mentioned, and all their performance gains claims were specific to the average FPS.

As anyone who has done any benchmarking at any point in history could tell you, CPU gains in game benchmarks are most often evident as a gain in the lowest 1% of framerates. If you go back to watch the review, much of the Zen 5 gains against the Zen 4 are made in the lowest 1% framerates, even if the peaks aren't much different. Those gains aren't ubiquitous, and there are plenty of places where they're still kinda meh, or worse. Still, I'm not sure I would declare Zen 5 is "only" 5% faster given how much some of the 1% frametimes moved in a notable handful of games, no matter the Windows version being tested.

In summary: I'd like to see more CPUs in the mix for testing, and I'd love to see both 1% and 10% lows show up in the reporting. Remember when consistency in frametime was a thing we talked about?
 
This is the best speculation I've seen as to what the changes could be:


"Not working at AMD, but AFAIK: Older AMD CPUs have side channel vulnerability which requires extra BHT or BTB flushes to migitate. These are not needed on zen5. Current windows versions perrorm these flushes on all AMD chips, update removes these flushes on zen5."

Zen4 and intel also got some increases, depending on the game. Could be that Microsoft took a scorched earth approach to solving these security flaws with branch prediction, and now they're going back and streamlining the mitigations, so now different cpus are seeing difference performance gains depending how severe the mitigation requirements are.

I expect it'll take someone with an instrumenting profiler to do an investigation if Microsoft and AMD aren't forthcoming with the info, but if it's security related maybe they'll never fully reveal it.
 
This is the best speculation I've seen as to what the changes could be:


"Not working at AMD, but AFAIK: Older AMD CPUs have side channel vulnerability which requires extra BHT or BTB flushes to migitate. These are not needed on zen5. Current windows versions perrorm these flushes on all AMD chips, update removes these flushes on zen5."

Zen4 and intel also got some increases, depending on the game. Could be that Microsoft took a scorched earth approach to solving these security flaws with branch prediction, and now they're going back and streamlining the mitigations, so now different cpus are seeing difference performance gains depending how severe the mitigation requirements are.

I expect it'll take someone with an instrumenting profiler to do an investigation if Microsoft and AMD aren't forthcoming with the info, but if it's security related maybe they'll never fully reveal it.
Some of the gains are so large Steve said he's not sure if he believes his own results. And the 14600K gains 25% in Gears 5, but no change at all in other games. Could there be multiple things going on with this update?

I can never recall a new version of Windows or an update bringing such massive performance gains. 10% average uplift with plenty of games showing much more than that. It's basically a generational leap in some very popular titles (Fortnite, Hogwarts, COD etc.). If this really (mostly) doesn't benefit Intel, it changes the competitive landscape for gaming CPUs.
 
Last edited:
Some of the gains are so large Steve said he's not sure if he believes his own results. And the 14600K gains 25% in Gears 5, but no change at all in other games. Could there be multiple things going on with this update?
I think this is a reasonable hypothesis, yes. I agree with @Scott_Arm's posting regarding security mitigations which were required for both Intel and AMD in prior chip models are a solid hypothesis for a large portion of gains, however I also wager there are other fine-tunings (which very well might be related) to either the kernel process scheduler or to how process management and isolation might work, which would then result in measurable performance gains.

Or it could just be fairy dust, in code form ;)
 
Some of the gains are so large Steve said he's not sure if he believes his own results. And the 14600K gains 25% in Gears 5, but no change at all in other games. Could there be multiple things going on with this update?

I can never recall a new version of Windows or an update bringing such massive performance gains. 10% average uplift with plenty of games showing much more than that. It's basically a generational leap in some very popular titles (Fortnite, Hogwarts, COD etc.).

The logic of it being security mitigation changes could make sense. I know some of the mitigations were supposed to have big impacts. I know you can manually disable a bunch of the cpu exploit fixes with registry changes. Things like Spectre, Meltdown, Downfall. I'd be curious to see someone take 23H2, or whatever, disable all of the cpu security fixes and compare that to 24H2. I know some gamers disable them all to get a performance boost if their system is basically just used for gaming.

But yah, there could be multiple things going on. It's one of the issues with being an administrator in Windows. You're never totally sure of everything that's being changed, unlike Linux where things are a lot more transparent.

There are a ton of angles to test here, but ultimately if other reviewers end up with similar findings, I guess the underlying cause doesn't necessarily matter. It'll just be the case that various generations of cpus get a performance boost.
 
My guess would be that this has something to do with HVCI but to test that we'd need to add Windows 10 in there.

AMD seemed to be telling hardware unboxed that this was a fix for branch prediction, which seemed odd because you wouldn't expect the OS to affect branch prediction behaviour. But you have side-channel attacks that try to exploit branch prediction hardware in the front-end of the cpu, and mitigations can do things like flush the branch prediction buffer. The fixes for all of these exploits are different, and complicated. All above my pay grade, but that speculation does seem to be in line with what AMD is telling hardware unboxed.
 
AMD seemed to be telling hardware unboxed that this was a fix for branch prediction, which seemed odd because you wouldn't expect the OS to affect branch prediction behaviour.
Yeah this has always seemed like a totally b.s. proposition to me. OS doesn't have anything to do with branch prediction of a CPU which is 100% in h/w.

But you have side-channel attacks that try to exploit branch prediction hardware in the front-end of the cpu, and mitigations can do things like flush the branch prediction buffer. The fixes for all of these exploits are different, and complicated.
These are fixed with AGESA microcode updates on the AMD platform (i.e. m/b BIOS updates). Can't really remember Windows doing m/c updates for Ryzen, ever, but my memory may fail me on that. In any case none of these should apply to CPUs which don't have the vulnerabilities so a suggestion that Windows somehow apply <=Zen 4 patches to Zen 5 (which likely have a totally different m/c to being with) sounds highly dubious.

Also this suggestion doesn't explain why Zen 4 gets about the same speed up as Zen 5 on 24H2. Worth remembering that even though that version was "launched" with Snapdragon laptops in July it is still very much in development for other h/w platforms. It is somewhat possible that this pre-release version runs without some security layers which would be turned on in the release one resulting in a performance drop down to 23H2 levels.

If not then the only thing which I can think of which might have gotten big improvements would be HVCI-related stuff (core isolation, memory integrity, etc). This was kinda hinted at in Wendel's video about a week ago here:


Edit: Also Intel gains in some titles. Definitely not a result of old Zen mitigations or "new branch prediction code" being applied to them.
 
Last edited:
@DegustatoR There's a long list of mitigations which can be disabled in the registry. I would guess it's a combination of microcode and software patching to implement the full mitigation. I know these mitigations have a performance cost. I don't think this is new "branch prediction code" either. Doesn't seem crazy to me that they could be misapplying software fixes to cpus that don't need them (zen4, zen5), or that they've reworked the fixes to lessen the performance impact.

 
Here's an interesting thread from David Huang explaining what the patch does (assuming they're correct.)


Basically the OS uses instructions to insert barriers that flush the branch prediction buffer. The penalty is bigger on zen5 because of the size of the buffer, so the patch changes the flushing behaviour, as I guess the CPU is not as vulnerable (hopefully).

I'm not familiar with this person, but they seem to have a fairly in-depth tech blog.
 
Been reading more. The mitigations for Spectre, especially variant 2, has led to changes in MSVC, and code generation. Really interesting how there are different software fixes for different generations of CPUs. These are heavy reading


And these are just Spectre related. There have been a number of side-channel attacks that have been discovered since.
 
If others have kept their Windows updated including optional updates, it would explains the lack of difference or at least part of it.
The branch prediction optimization has been released for 23H2 with KB5041587, which is available as optional update at this time.

This patch was out yesterday, I doubt that all the tests for these tweets were done yesterday evening.
 
This patch was out yesterday, I doubt that all the tests for these tweets were done yesterday evening.
Schilling (HardwareLuxx) has updated his test with the patch, results looking more like they should now apparently (haven't read the article before or after, just reporting from another forum)
 
Last edited:
Back
Top