AMD Execution Thread [2024]

Yeah, that's a reasonable hypothesis IMO. The wider cores (more execution units along with higher L1/L2 bandwidth) provide a lot more opportunity to cram multiple disparate instructions into a singular core. Current X86 cores really only expose two logical threads per physical core (2-way SMT) however there are other chips out there in the world which expose 4-way and even 8-way SMT, such as IBM's Power series they use in AIX boxes.

If you've donned your thick glasses, pocket protector, and highwater paints with tall white socks and patent leather shoes today ( ;) ) you may be interested in reading Anand's writeup of the Power 8 SMT capabilities here: https://www.anandtech.com/show/10435/assessing-ibms-power8-part-1/4
 
The fact that they aren't patching/mitigating Ryzen 1000, 2000, or 3000 series is pretty insane considering they're spending the time and effort to patch every generation of Epyc going back to Zen1, which use the same dies, and are also affected. There are an awful lot of Ryzen 3000 chips still out in the wild in active use.

 
Disagree, AVX-512 is very niche (not in the least, because of Intel's abysmall adoption of their own instructions) . For non-worstation users, doubly so.

As you said (that it positively impacts per/watt), it seems to be aptly intended for mobile where perf/watt is more important

Going through the Zen 5 arch writeup, it seems to have been a rebuild very much focused on perf per watt with as little area cost increase as possible. The arch goes wide and other trade offs all over the place, attempting to claw perf/w anywhere it while minimizing area increase, and it looks to have worked out.

That Zen 5 doesn't increase perf per area is only a concern for gamers at this point, datacenter is running into major power issues from exponential growth, and mobile always wants more perf/w. This is a great arch for AMD's finances. But they probably should've skipped any gaming claims at all and nixxed the review units to a lot of the youtuber channels in favor of telling them to wait for X3D parts.
 
Going through the Zen 5 arch writeup, it seems to have been a rebuild very much focused on perf per watt with as little area cost increase as possible. The arch goes wide and other trade offs all over the place, attempting to claw perf/w anywhere it while minimizing area increase, and it looks to have worked out.

That Zen 5 doesn't increase perf per area is only a concern for gamers at this point, datacenter is running into major power issues from exponential growth, and mobile always wants more perf/w. This is a great arch for AMD's finances. But they probably should've skipped any gaming claims at all and nixxed the review units to a lot of the youtuber channels in favor of telling them to wait for X3D parts.
Which write up did you read?
 
The agreement was reportedly that if/when the contract ended, the ZLUDA code could be open-sourced. That's what happened back in February. But now that code has been retracted from the official public GitHub repository. It's not from legal threats from NVIDIA as one might imagine given its working to support CUDA on non-NVIDIA hardware, but rather from AMD itself.

Andrzej Janik updated the GitHub repository a few minutes ago with the message:
IMPORTANT

What happened

The code that was previously here has been taken down at AMD's request. The code was released with AMD's approval through an email. AMD's legal department now says it's not legally binding, hence the rollback. Before anyone asks: I have received no legal threats or any communication from NVIDIA.

What now

At this point, one more hostile corporation does not make much difference. I plan to rebuild ZLUDA starting from the pre-AMD codebase. Funding for the project is coming along and I hope to be able to share the details in the coming weeks. It will have a different scope and certain features will not come back. I wanted it to be a surprise, but one of those features was support for NVIDIA GameWorks. I got it working in Batman: Arkham Knight, but I never finished it, and now that code will never see the light of the day:

So six months after the code was made public as open-source, at the request of AMD's legal department, that ZLUDA code has now been removed. Though given it's Git and may have been cloned, the open-source code likely exists elsewhere by those that were intrigued by this effort.
 
Which write up did you read?

Based on current sales numbers of Zen 5 desktop I question whether desktop sells to anyone but gamers anymore. Should AMD do X3D only for desktop?
 

Based on current sales numbers of Zen 5 desktop I question whether desktop sells to anyone but gamers anymore. Should AMD do X3D only for desktop?

Gamers making up the majority of desktop sales wouldn't be shocking. I think the majority of people using computers for productivity get them from work, and they're laptops. All of the private stuff would be done on their phones. I'd guess most students would have laptops as well. I'd be curious to know how many family households have a kind of shared desktop computer anymore. Maybe it's more common than I think, but if it's not for gaming I'd expect a lot of them to be purchased and kept for a very long time.
 
Gamers making up the majority of desktop sales wouldn't be shocking. I think the majority of people using computers for productivity get them from work, and they're laptops. All of the private stuff would be done on their phones. I'd guess most students would have laptops as well. I'd be curious to know how many family households have a kind of shared desktop computer anymore. Maybe it's more common than I think, but if it's not for gaming I'd expect a lot of them to be purchased and kept for a very long time.
I think laptops are making a come back. My nieces have all gotten laptops over the last 2 years. they have chromebooks from school and iphones and some had ipads but they all have stopped using the ipads and moved to laptops. I have no idea why. my nephews are all into gaming pcs which is great because we built them together and it reminded me of my uncle helping me build my first pc back in the 80s. My father in law has asked me to build him a new pc for his CPA business. So i still think its pretty broadly used. When I need to do something that isn't browsing the web I grab my surface pro and my wife uses it also.

But i certainly do agree that laptops are being used for 5-7 years. I am gonig through this with my younger cousin now. Her 6 year old laptop died so we are picking a new one for her and i'm giong to transfer all her data over for her.
 
So, AMD designed this 'core parking' control driver for dual-CCD 7000-series X3D processors - and now enabled even for regular 9000-series non-X3D models, BUT... unlike Intel Speed Shift implemented with on-chip hardware, it uses a proprietary kernel-mode driver that can't be uninstalled or disabled without reinstalling Windows from scratch, even though it's known to be buggy enough to misidentify X3D-enabled cores and interfere with single-CCD models! What the...


The core parking tech effectively shuts down one CCD during gaming to boost cache hit rates, reduce cross-CCD traffic, and keep the workload pinned to the fastest CCD, all of which boosts gaming performance. The dual-CCD provisioning approach works exceedingly well for X3D processors because only one CCD has the vast L3 cache enabled by a vertically stacked chiplet. Still, it should also benefit the dual-CCD Ryzen 9000 chips without X3D cache because it keeps latency-sensitive game data close to the execution cores.

The feature requires four components: a new chipset driver, updated BIOS, Windows Game Mode, and the Xbox Game Bar (make sure to update it through the Microsoft store). The Xbox Game Bar contains a KGL (known good list) of games that it detects when active, thus triggering Game Mode (you can also instruct the game bar to recognize unknown games and/or other applications as games). The driver communicates with Windows Game Mode to park the cores on a single CCD, thus constraining latency-sensitive workloads (like games) to the higher-performance chiplet.

The core parking feature has a major problem: it can’t be uninstalled from the operating system. As such, if you later install another processor but use the same operating system, the feature will persist and can continue to park cores (potentially unbeknownst to the user), thus hamstringing performance with processors that aren’t designed to use the feature.

This also creates problems for reviewers who test multiple processors on the same motherboard. (It’s even conceivable that this issue contributed to the inconsistent results we have seen with the first wave of Ryzen 9000 reviews, as 7000X3D models can also trigger the issue and cause all other chips tested on the platform to not operate as intended.)

We also noticed that the core parking feature isn’t bulletproof — we observed cores slipping in and out of a parked state during gaming on several occasions. This would obviously have a negative performance impact, but the feature also worked perfectly fine at other times.


Shouldn't this actually be a feature of the Windows process scheduler, rather than some 3rd-party patch with a proprietary filter driver and a kernel-mode service controlled by the Xbox Game Bar?


PS. Well, it looks AMD 'core parking' control drivers are not as woefully underengineered as assumed by Tom's Hardware review of Ryzen 9 9950X above.

If you move from single-CCD to dual-CCD processors, you can simply uninstall then reinstall the latest AMD Chipset Drivers (currently 6.07.22.37) to enable/update the "AMD PPM Provisioning File Driver" / " AMD 3D V-Cache Performance Optimizer Driver" and clear the previous settings; no clean-state reinstall of Windows should be required.

You may need to change the default UEFI BIOS setting for core preference in Advanced -> AMD CBS -> CMU Common Options -> CPPC Dynamic Preferred Cores = Driver (instead of Auto, Frequency, or Cache)
- though 'Auto' is recommended by AMD in the Community thread above and it should default to 'Driver' on double-CCD processors, sometimes it's not working as intended until you set it manually...




PPS. FYI 'core parking' is essentially a manufacturer-defined processor power state, like the C6 'Deep Power Down' first introduced by Intel Penryn-M; the 'Core Parking Engine' is already a part of Windows Processor Power Management (PPM), though it powers down the cores way too aggressively - so it was customary to use PowerCfg commands and 3rd party tools like Park Control, as well as hidden 'Ultimate Performance' power plan, to prevent CPU cores from entering these additional processor states...
 
Last edited:
Benchmarks all over the place. At higher power envelopes there appears to be 0 improvement over Zen 4 (averaged out) other than AVX512

I've seen claims that Linux runs some games faster than Windows, which is... odd.
 
Last edited:

I don't know what to make of this (have to use Windows hidden admin account or lose performance?) but it also affects Zen 4. He didn't test Zen 3 or below.

Did this affect Ryzen for at least an entire generation and nobody noticed? Whoops.

Whatever it is they say it will be fixed in a Windows update.
 
@homerdog curious if it affects Intel. Windows sucks so I’m not surprised there are issues like this. Hoping for the day something more streamlined displaces windows as the main gaming operating system.
 
Back
Top