Why everyone say Zen 5 is bad ?

[oxrufiioxo]

I never really liked them that much to begin with, I lost interest around 2004. AT, Xbitlabs and SPCR had lot more details in reviews, even if not having much in common.

I mostly used them and AnandTech.... My first diy pc was based on Toms tier list. I liked Techspot and Techpowerup also though in the late 00s

Tom's started going downhill in 2013 imho..... Although they took a slight hit to quality in 2007 when they were sold the first time.

 

[Mr_Engineer]

So the Windows 11 24H2 patch will improve performance by a few % points for not only Zen 5, but also Zen 4, Zen 3, and maybe Intel CPUs?

If that is the case, then the difference in performance improvement between 9000 series and 7000 series will still be the same, 5 to 10% on average.

My 5800x rig is due for a change, but I'll wait till X3D CPUs and Intel Arrow Lake are released to decide.

 

[Wolverine2349]

@

oxrufiioxo​

What a coincidence you say Toms Hardware was good back in 2008-2009 and started going down hill in 2013.

Getting off topic a bit but well 2013 is also when the housing market started to change unexpectedly and not for the better and for far worse for society lol. Insane inventory shortages and rapid rapid price increases which is a terrible thing for society. And never looked back and worst ever and ever today.

2008-2011 was the good old days of the housing market especially 2010-2011 when prices were fat at reasonable levels when the pain of them dropping form much inflated levels in 2008-2009 was over (much needed as they were so high before and never should have been).

It bogles my mind anyone could celebrate the fact that prices started to rise fast in 2013 or even late 2012 and never looked back. Flat home prices at 2010-2011 levels would have been best for most of society!!

The real housing crisis is not what happened in 2008-2009. Its high housing costs and terrible affordability from the long term perspective.

And I say this as a home owner since 2013. I pity future generations and feel bad for those who cant live with their parents and save and make progress on purchase price of a home because the price is so much higher years in future for those who hate mortgages. I am one of those and was very lucky to get in in nick of time while living with my parents to save for a house.

So the Windows 11 24H2 patch will improve performance by a few % points for not only Zen 5, but also Zen 4, Zen 3, and maybe Intel CPUs?

If that is the case, then the difference in performance improvement between 9000 series and 7000 series will still be the same, 5 to 10% on average.

My 5800x rig is due for a change, but I'll wait till X3D CPUs and Intel Arrow Lake are released to decide.

Yes good point which mean Zen 5 still sucks as it is 5% better at best on apples to apples comparison. Sounds like WIN11 gimped performance of Ryzen 5000 and 7000 and 9000 equally so this idea a patch will help not so since it also helps 5000 and 7000 just as much. Meaning point still stands that Zen 5 is hardly if any better than Zen 4 for gaming and most other consumer and even professional workloads that are not AVX512. AVX512 on other hand Zen 5 is a big big uplift. But few use AVX512 for anything.

 

[SL2]

So the Windows 11 24H2 patch will improve performance by a few % points for not only Zen 5, but also Zen 4, Zen 3, and maybe Intel CPUs?

If that is the case, then the difference in performance improvement between 9000 series and 7000 series will still be the same, 5 to 10% on average.

I'm not so sure, it seems to affect them differently, but not by much.

Given what's been said in L1's latest video, I don't think we've seen the end of it. Yes, the new Windows version shows improvements, but it's far too inconsistent to make me believe that this is it.

24H2 is only part of the solution, newer drives has been promised, and newer AGESA might also change things.

It's not that I refuse to believe that AMD can deliver a mediocre product, it's just that it needs more (beta) testing and improvements based on everything I've seen.

 

[phanbuey]

I'm not so sure, it seems to affect them differently, but not by much.

Given what's been said in L1's latest video, I don't think we've seen the end of it. Yes, the new Windows version shows improvements, but it's far too inconsistent to make me believe that this is it.

24H2 is only part of the solution, newer drives has been promised, and newer AGESA might also change things.

It's not that I refuse to believe that AMD can deliver a mediocre product, it's just that it needs more (beta) testing and improvements based on everything I've seen.

It's the exact same launch as zen4 - that took a good bit to mature.

The only launches that HAVEN'T been total sh*t shows by AMD have been the 2000 series and the 3000 series.

5000 had a ton of compatibility with x370 issues, WHEA issues - it was offset by the sheer performance gain. 7000 same thing with stuttering and latency issues on dual CCD parts. Most of it eventually seems to be fixed with updates.

 

[oxrufiioxo]

It's the exact same launch as zen4 - that took a good bit to mature.

The difference is even with it's teething issues Zen 4 performed way better at launch vs Zen 3.

Amd even contradicts themselves with the gaming perfomance claims initially saying 9% then saying 5-8% in the same article lol also if you look at their performance results they make no sense for some of the games.

They need to hire better testers seems like the real problem.

 

[SL2]

It's the exact same launch as zen4 - that took a good bit to mature.

Maybe there were issues, but the reviews were a lot more positive in terms of performance alone, and they were consistent.

So no, not exactly the same.

 

[Assimilator]

They need to hire better testers seems like the real problem.

This has been AMD's problem for the entirety of its existence. QA departments can and do make or break products.

 

[oxrufiioxo]

This has been AMD's problem for the entirety of its existence. QA departments can and do make or break products.

While true this can be said for the other big two as well all there have had issues with their latest generations at the very least.

RDNA3/Zen4 are especially bad when it comes to marketing slides vs actual performance.

Zen 1-4 and RDNA 1-2 where at least ball park.

 

[SL2]

They need to hire better testers seems like the real problem.

Not their only problem though. I can't imagine AMD aiming for a 4% game uplift.

Now I understand that there might be 1000 reasons for why they won't get to their performance goals.

Bad testers, yes.

Half baked product, yes.

 

[oxrufiioxo]

Not their only problem though. I can't imagine AMD aiming for a 4% game uplift.

Now I understand that there might be 1000 reasons for why they won't get to their performance goals.

Bad testers, yes.

Half baked product, yes.

I meant so that they produce more realistic numbers not to make their products look better.

Their perfomance numbers this generation were straight out of dreamland.

 

[SL2]

I meant so that they produce more realistic numbers not to make their products look better.

Their perfomance numbers this generation were straight out of dreamland.

Yeah, they messed up for sure, just like HUB summarized all the bad AMD launches recently.

Maybe they've had too many good launches in a row and forgotten how much backlash they can get.

I mean Zambezi was incredibly hyped, but I don't remember AMD's own numbers before launch. Something tells me that it wasn't realistic lol

 

[oxrufiioxo]

Yeah, they messed up for sure, just like HUB summarized all the bad AMD launches recently.

Maybe they've had too many good launches in a row and forgotten how much backlash they can get.

I mean Zambezi was incredibly hyped, but I don't remember AMD's own numbers before launch. Something tells me that it wasn't realistic lol

The sad thing is even now if you looked at the new results some of them make zero sense.

No actual reputable reviewer would look at them and go yeah just publish the review. I'm pretty sure even a hobbyist on YouTube could produce more accurate results....

 

[SL2]

The sad thing is even now if you looked at the new results some of them make zero sense.

No actual reputable reviewer would look at them and go yeah just publish the review. I'm pretty sure even a hobbyist on YouTube could produce more accurate results....

Like I said, a half baked product.

PCWorld didn't publish for that very reason.

 

[A Computer Guy]

And when you think about a hylerscaler that likely overprovisions their machines with shared instances, caches are going to be thrashed and made useless.

Question. Wouldn't this cache thrashing still be an issue if the VM was overprovisioned across different CCD's for non-x3d cpus?

 

[RUSerious]

Like I said, a half baked product.

PCWorld didn't publish for that very reason.

This seems to be the only logical conclusion. One issue that will always affect AMD desktop CPUs is priority: Server>>Laptop>>Desktop (and DIY desktop is even lower). AMD needed DIY desktops to get Ryzen off the ground initially, but not anymore. The CCD we get for desktop is sort of the 'rejects' that were destined for Server CPUs. In this case, the perf/watt performance targets for Turin may have hurt granite ridge - in addition to the other issues.

 

[phanbuey]

Question. Wouldn't this cache thrashing still be an issue if the VM was overprovisioned across different CCD's for non-x3d cpus?

Azure HBv3 virtual machines for HPC, now up to 80 percent faster with AMD Milan-X CPUs | Microsoft Azure Blog

I think it depends on the what the VM is doing but looks like in some cases it definitely helps performance (according to the Azure team).

 

[A Computer Guy]

View attachment 360154
Azure HBv3 virtual machines for HPC, now up to 80 percent faster with AMD Milan-X CPUs | Microsoft Azure Blog

I think it depends on the what the VM is doing but looks like in some cases it definitely helps performance (according to the Azure team).

I'm still trying to wrap my head around the article above and how it applies to my question but I was in the middle of doing some tests in my VMWare Workstation with a 6 thread configuration and came across a surprising result. Note the Same CCD scenario scored relatively the same regardless of CCD1 vs CCD2 with CCD1 being slightly better likely because we already know CCD1 will have the stronger chiplet. The mixed CCD result was surprisingly much better. Take these results with a grain of salt because I'm sure this ad-hoc test has some flaw I haven't accounted for.

When the 6-core VM is not restricted by cores configuration the score is around 1300 but in that case Windows 10 is keeping all my activity of the VM on the first CCD.

 

[Wolverine2349]

This seems to be the only logical conclusion. One issue that will always affect AMD desktop CPUs is priority: Server>>Laptop>>Desktop (and DIY desktop is even lower). AMD needed DIY desktops to get Ryzen off the ground initially, but not anymore. The CCD we get for desktop is sort of the 'rejects' that were destined for Server CPUs. In this case, the perf/watt performance targets for Turin may have hurt granite ridge - in addition to the other issues.

Yeah and that sucks.

Though whats real funny about that, all the benchmark sites seem to disagree. The coverage of DIY desktop CPUs for Hardware Unboxed, CPU Monkey, Gamers Nexus, Toms Hardware, Anandtech, here at TechPowerup OC3D, Igors Lab, and the like are so much more for consumer DIY CPUs than for server and HEDT CPUs.

I mean maybe maybe in some way Sapphire Rapids would be a good solution to what we are looking for if we were willing to pay its price. Problem is there is almost 0 data on any of its CPUs regarding performance or if all cores on one die or core to core latency. You cannot even find CInebench score on CPU monkey for Sapphire rapids CPUs.

We know AMD has a max of 8 cores per CCX/CCD and their THreadripepr parts including Zen 4 high core counts is just a bunch of 8 core CCDs in one package.

Intel cannot find crap on SPR to know if maybe it could be an answer or not. I kind of think not, but the data is so limited I have no clue.

I mean if server and laptops were such priority why is it that sites like Anandtech and TechPowerUp ad Toms Hardware and Igors Lab show consumer DIY benchmarks more so. I understand why HUB and Gamers Nexus do as those channels are specifically for it. The other sites are more general hardware and not gamer centric unlike HUB and GN. SO if DIY was really so niche and unimportant to AMD and Intel, why are all sites prioritizing benchmarks for it rather than laptop and server CPU market??

I would love to see lots more data on SPY Xeon Workstation and overcocing cooling IPC, latency, are all cores in the 12 and 16 core parts one one die or no??? What stepping and how many they have?? How do they scale with RAM overclocking. There is 0 of that data to be found or extremely limited it is almost useless.

 

[R0H1T]

This seems to be the only logical conclusion. One issue that will always affect AMD desktop CPUs is priority: Server>>Laptop>>Desktop (and DIY desktop is even lower). AMD needed DIY desktops to get Ryzen off the ground initially, but not anymore. The CCD we get for desktop is sort of the 'rejects' that were destined for Server CPUs. In this case, the perf/watt performance targets for Turin may have hurt granite ridge - in addition to the other issues.

It's not that servers are a priority it's more like server workloads can actually make use of "moar" cores properly & the load is generally distributed properly even with some (latency/NUMA/memory) limitations! You think people at Google, Azure, Oracle, AWS just like to spy on you all the time

Why do you think Intel needed to clocks their chips to the crescent moon?

AMD, or Intel, in that regard's not responsible for bad coding nor developers/studios unable to use more than 6 cores!

 

[SL2]

The CCD we get for desktop is sort of the 'rejects' that were destined for Server CPUs. In this case, the perf/watt performance targets for Turin may have hurt granite ridge - in addition to the other issues.

This can only be true if low yields forces AMD to do so, otherwise they have no reason to. Of course they'll do it to some degree, but I'm not sure to what extent.

Since we don't know shit about yields, there's no point in speculating.

 

[igormp]

Question. Wouldn't this cache thrashing still be an issue if the VM was overprovisioned across different CCD's for non-x3d cpus?

Yup, which makes the current L3 (which is just a victim cache) really useless, no matter the size. So x3D or not, you'd be trashing it a lot.

View attachment 360154
Azure HBv3 virtual machines for HPC, now up to 80 percent faster with AMD Milan-X CPUs | Microsoft Azure Blog

I think it depends on the what the VM is doing but looks like in some cases it definitely helps performance (according to the Azure team).

That's not really related to the question at hand. The point we were discussing is a single VM host (with let's say 8 cores and 16GB of RAM) with 16 guests, each allocated 1 CPU and 2GB of RAM (so 16 CPUs and 32GB of ram in total, twice what the host has), which is the scenario of overprovisioning, and what's usually done by most cloud hosts.

Your example is more akin to single host with a single VM running HPC software (something that really loves the extra cache), and having multiple of those hosts running as a cluster (hence the whole infiniband propaganda in the actual link).

3D-cache doesn't help running VMs, but helps applications that like the 3D-cache and that may be running inside a VM (as long as you're not overprovisioned). Azure's H lineup is solely focused on HPC (they even have SMT disabled), and those VMs can't be overprovisioned. All you can do is rent an ENTIRE machine and either use all the cores, or constrain it to only use a subset of cores (for cheaper licensing fees), while keeping the remaining of the specs the same. Feel free to read more about in the actual docs for it:

HBv3-series virtual machine (VM) overview, architecture, topology - Azure Virtual Machines - Azure Virtual Machines

Learn about the HBv3-series VM size in Azure.

learn.microsoft.com

HBv3 size series - Azure Virtual Machines

Information on and specifications of the HBv3-series sizes

learn.microsoft.com

I'm still trying to wrap my head around the article above and how it applies to my question

Simple: it doesn't lol

The mixed CCD result was surprisingly much better. Take these results with a grain of salt because I'm sure this ad-hoc test has some flaw I haven't accounted for.

From a quick glance it seems like in your "same CCD" tests you may be using 3 cores, and the 2 logical threads from each core.
On your "mixed CCD" I believe you're actually making use of 4 different cores, 2 of each only have 1 logical thread active being loaded (so those threads can make use of an entire core's EUs instead of sharing those resources with the sibling thread).

 

[Solaris17]

You cannot even find CInebench score on CPU monkey for Sapphire rapids CPUs

I posted SPR cinebench scores in the cinebench threads I think earlier this week? But yes, It would be nice for more places to cover these systems or server parts in general I still need to a find a c741 board for my EMR ES.

Though none of that, even the VM discussion in general seem relevant at all to this thread or the release of Zen5.

 

[A Computer Guy]

Yup, which makes the current L3 (which is just a victim cache) really useless, no matter the size. So x3D or not, you'd be trashing it a lot.

Trashing the cache means poor performance correct? My ad-hoc test was attempting to confirm your statement so I was expecting worse performance in the mixed CCD scenario however my test may be flawed if Cinebench doesn't actually thrash cache due to the nature of it's work. Also stepping back to the original statement it was talking about overprovisioned cores. In my test clearly I'm not overprovisioning the cores so that is likely another flaw in my test not supplying enough contention on the cache to produce a meaningful test.

That's not really related to the question at hand. The point we were discussing is a single VM host (with let's say 8 cores and 16GB of RAM) with 16 guests, each allocated 1 CPU and 2GB of RAM (so 16 CPUs and 32GB of ram in total, twice what the host has), which is the scenario of overprovisioning, and what's usually done by most cloud hosts.

Your example is more akin to single host with a single VM running HPC software (something that really loves the extra cache), and having multiple of those hosts running as a cluster (hence the whole infiniband propaganda in the actual link).

3D-cache doesn't help running VMs, but helps applications that like the 3D-cache and that may be running inside a VM (as long as you're not overprovisioned). Azure's H lineup is solely focused on HPC (they even have SMT disabled), and those VMs can't be overprovisioned. All you can do is rent an ENTIRE machine and either use all the cores, or constrain it to only use a subset of cores (for cheaper licensing fees), while keeping the remaining of the specs the same. Feel free to read more about in the actual docs for it:

HBv3-series virtual machine (VM) overview, architecture, topology - Azure Virtual Machines - Azure Virtual Machines

Learn about the HBv3-series VM size in Azure.

learn.microsoft.com

HBv3 size series - Azure Virtual Machines

Information on and specifications of the HBv3-series sizes

learn.microsoft.com

Simple: it doesn't lol

That's what I thought.

From a quick glance it seems like in your "same CCD" tests you may be using 3 cores, and the 2 logical threads from each core.
On your "mixed CCD" I believe you're actually making use of 4 different cores, 2 of each only have 1 logical thread active being loaded (so those threads can make use of an entire core's EUs instead of sharing those resources with the sibling thread).

To clarify the VM I was testing was allocated 6 threads. In the Same CCD example those 6 threads are essentially mapped to 3 cores. In the mixed CCD setup I made it slightly more complex by only using 1.5 cores on each CCD. In reality, regardless of VM configured cores, it seems when unconstrained VMWare Workstation actually spreads the load out among more available cores given the chance however I noticed it was generally keeping my vm on the first CCD unless I directed it otherwise.

 

[Frick]

I meant so that they produce more realistic numbers not to make their products look better.

Their perfomance numbers this generation were straight out of dreamland.

Maybe it's my age talking, but the point of PR is to make the products look better (and it's a fundemental part of human nature to boot so it comes natural to us).

My first diy pc was based on Toms tier list.

Yeah, my age is talking.