We found the Missing Performance: Zen 5 Tested with SMT Disabled

[digitalfrost]

I wonder what you happen if you pinned the games to the physical cores via Process Lasso for example, but left SMT enabled. This way you could have the performance gain in gaming while still retaining the multicore performance for other tasks.

 

[user556]

That explains why the Linux tests were so much rosier. The scheduling will be a lot different there.

 

[Punkenjoy]

Zen 5 was designed to be able to milk even more the server market and keep margin as high as possible. Linux Server performance seems to be very good (altough it could just be a better scheduler) and many servers workload seems to benefits greatly from the uArch.

And its about the same size as Z4 CCD (slightly smaller by about 2%) and performance per watt in these workload seems higher. On a socket capable of very high power, it should be a killer for the datacenter.

I know some people will say that core density is all matter but that is not true at all as many server software use a per core license. Having more slower core is just a bad idea. Having the possibility to get more out of each existing core is what help those workload. The core density problem is well taken care of with the C variant.

Public Cloud Provider right now prefer the many core variant as they sell you badly sized server for many servers. The fact that you can't get a server with huge amount of ram and few cores is bringing many customer to get back to private cloud (on prem) or legacy On prem stack.

 

[mahoney]

That's some reaching now.

 

[Icon Charlie]

Thank you for the article W1zard.

 

[RandallFlagg]

I know some people will say that core density is all matter but that is not true at all as many server software use a per core license.

That is huge and largely ignored by the desktop crowd.

A per core license for a DB server can easily be over $5,000 - and sometimes more like $12,000 - so if you get 20% more performance from a core you just saved $1000-$2400++ per core, because you need 20% fewer cores hence 20% fewer licenses. Multiply that across say 64 cores.

And we haven't yet added per core OS licenses and so on - admittedly much lower cost, but it's still there.

That cost completely trumps about every other consideration.

There are some cases where density rules - when you're leasing out the cores (someone else pays for the low performance), custom app servers, many web servers are also low cost. But in those cases, you probably aren't looking for the latest and greatest anyway - old and paid for is fine.

 

[tfp]

SMT/Hyperthreading has always been a boost when there is hardware idle because of CPU internals not being use, call outs to ram, or disk. However software that uses up all of any one type of hardware resource across multiple program threads just does not scale and puts additional constraints on the CPU itself in the form of cache and Memory usage. We can see in the bench marks that the 78003D chips don't have the same fall off in any of the tests the new chips do and I believe it is because the threads aren't as cache starved and the 9XXX chips are slightly more efficient leaving less idle resources for another thread to use.

I have seen this at work as well. Run an Oracle database and do something very math and memory intensive across the same set of tables and it scales just fine until you run out of HW cores, if you keep adding threads that are now sharing cores the performance improvement stops and performance starts to degrade as you add more threads to the work.

In light of the results seen on AMDs processors, and Intel is still not throwing a lot of cache at it's nextgen desktop processors, it's likely not worthwhile to have another 8 software threads on the P cores. Without more cache, a better ring bus, and more ram bandwidth those (hopefully) more efficient cores will spend to much time in contention with each other and that waists power however minor at times. Considering how much power Intel's chips waist now, a 5 to10% savings means a lot more to them then AMD based on scale of usage alone.

 

[evernessince]

Missing Zen 5 performance wasn't found, there's a mere 2.5% between SMT off vs on in games and it's slower in applications.

That's only a slightly larger gain that the 7700X obtained in the charts and not nearly enough to make up for the difference between AMD's marketing claims and reality.

 

[phanbuey]

I wonder what you happen if you pinned the games to the physical cores via Process Lasso for example, but left SMT enabled. This way you could have the performance gain in gaming while still retaining the multicore performance for other tasks.

it works but not as well, and also it can crash the game (depending on the game) the next time you try to launch it. Setting high priority on process lasso does work though. So disabling SMT and then pumping priority can net some nice results I imagine.

 

[phints]

These gains with SMP off in gaming are still miniscule, +2.7% at 1080p (+5% at 1% lows is decent) and worse in productivity. Meh. So Intel has a point removing HT in their upcoming CPUs?

If you want faster fps in your games wait for 9800X3D it'll probably be 20% faster.

 

[Dr. Dro]

The Phoronix results suggest a lot of Windows software needs to be recompiled with a Zen 5 target. The most obvious are the AES and encryption results where Windows tests are showing Zen 5 to be slower than Zen 4, while in Linux Zen 5 is considerably faster, by more than 100% in some cases.

Zen 5 target = AVX-512 enabled? It's not a special processor. The thing about Zen 5 is that it's simply unremarkable, it doesn't have major breakthroughs like Zens 2 3 and 4 did. Its singular achievement is its absolute mastery of AVX-512. It'll do it without the drawbacks seen in Intel processors and earlier generations of Ryzen thus far.

I wonder what you happen if you pinned the games to the physical cores via Process Lasso for example, but left SMT enabled. This way you could have the performance gain in gaming while still retaining the multicore performance for other tasks.

No, an SMT thread is a side channel to execute instructions from anywhere within a pipeline while the main pipe is stalled, the goal is to hide memory latency and increase throughput achieved by filling in "blanks" in the execution. So you have no option but to disable SMT to achieve the desired result here.

 

[RandallFlagg]

It's always entertaining to see these sites rediscover something that's been known for a decade or more.

HT / SMT has always affected scores by a few percent, largely depends on what games or game mix you're using.

Gaming benchmarks: Core i7 6700K hyperthreading test

Greetings all! According to my previous benchmark Intel Ivy Bridge Core i7 processor offers no hyper-threading (HT) performance in gaming. On the contrary, HT turned on on the Core i7 3770 did hurt the performance slightly when compared to just the 4 physical cores of Core i7 3770. This time...

www.techpowerup.com

Does Intel Hyper Threading Hurt Gaming Performance? I9-9900K Analysis – RedGamingTech

www.redgamingtech.com

(2017 thread):

Theories on why the SMT hurts the performance of gaming...

www.overclock.net

 

[ShogoXT]

Thanks for the hard work. It's clear Microsoft doesn't give a damn about Windows anymore. These issues have gone unresolved for years.

 

[tfp]

It's always entertaining to see these sites rediscover something that's been known for a decade or more.

HT / SMT has always affected scores by a few percent, largely depends on what games or game mix you're using.

Gaming benchmarks: Core i7 6700K hyperthreading test

Greetings all! According to my previous benchmark Intel Ivy Bridge Core i7 processor offers no hyper-threading (HT) performance in gaming. On the contrary, HT turned on on the Core i7 3770 did hurt the performance slightly when compared to just the 4 physical cores of Core i7 3770. This time...

www.techpowerup.com

Does Intel Hyper Threading Hurt Gaming Performance? I9-9900K Analysis – RedGamingTech

www.redgamingtech.com

(2017 thread):

Theories on why the SMT hurts the performance of gaming...

www.overclock.net

Yep this was talked about when the P4 came out and it had hyperthreading.

 

[Makaveli]

Should bin gaming tests below 1440p.

No one with a 4090 and Zen 5 will use 1080p or 720p. So what's the point of testing them ?

 

[rodrigorras]

Zen 5 was designed to be able to milk even more the server market and keep margin as high as possible. Linux Server performance seems to be very good (altough it could just be a better scheduler) and many servers workload seems to benefits greatly from the uArch.

And its about the same size as Z4 CCD (slightly smaller by about 2%) and performance per watt in these workload seems higher. On a socket capable of very high power, it should be a killer for the datacenter.

I know some people will say that core density is all matter but that is not true at all as many server software use a per core license. Having more slower core is just a bad idea. Having the possibility to get more out of each existing core is what help those workload. The core density problem is well taken care of with the C variant.

Public Cloud Provider right now prefer the many core variant as they sell you badly sized server for many servers. The fact that you can't get a server with huge amount of ram and few cores is bringing many customer to get back to private cloud (on prem) or legacy On prem stack.

I agree, I am disappointed with Hardware Unboxed, they were saying a lot of crap about AMD products,
They didn't even wait for 9900, or 9950 to conclude that Zen5 is shit. They never read the Tomshardware or Phronix review...

 

[dgianstefani]

Missing Zen 5 performance wasn't found, there's a mere 2.5% between SMT off vs on in games and it's slower in applications.

That's only a slightly larger gain that the 7700X gained in the charts and not nearly enough of a difference to make up for the difference between AMD's marketing claims and reality.

Significant, when there's a "mere" 2.5% average gain between 7700X and 9700"X", otherwise. So this small tweak effectively doubles the performance difference, and brings gaming much closer to expectations.

It's also only slower in some applications.

 

[HD64G]

So Intel was on to something for Arrow Lake, ditching HT...

Indeed! They forced to disable HT for their server-data center chips to cover the security holes by the vulnerabilities which used that. Google it of you don't believe what I posted.

 

[Hattu]

I don't know why this new am5 platform would be D.O.A. At least in Linux, the performance is there.

All I see is better performance. If I had the money, I would change my system in a heart beat.

 

[Vayra86]

Oh no, don't you dare!

Would be the most hilarious BIOS update. Adds option ' Disable AMD' and then the whole GUI also turns blue and the system always tells you it runs cool and quiet even when overheating and actively degrading your chip. It also enables Extended RMA as an auto default then.

 

[Macro Device]

Adds option ' Disable AMD' and then

...your brand new RX 7900 XTX explodes.

 

[Vayra86]

...your brand new RX 7900 XTX explodes.

No it'll magically accelerate RT content now and matches a 4090

 

[Panther_Seraphin]

So Intel was on to something for Arrow Lake, ditching HT...

HT has been a massive security issue for a very long time, however HT was an easy way to increase performance for very little die space and minimal extra power. (Look at things like Cinebench etc)

AMD have already removed HT from their chips this with some of their Bergamo EPYC designs with the HT fused off for this particular reason.

That is huge and largely ignored by the desktop crowd.

A per core license for a DB server can easily be over $5,000 - and sometimes more like $12,000 - so if you get 20% more performance from a core you just saved $1000-$2400++ per core, because you need 20% fewer cores hence 20% fewer licenses. Multiply that across say 64 cores.

And we haven't yet added per core OS licenses and so on - admittedly much lower cost, but it's still there.

That cost completely trumps about every other consideration.

There are some cases where density rules - when you're leasing out the cores (someone else pays for the low performance), custom app servers, many web servers are also low cost. But in those cases, you probably aren't looking for the latest and greatest anyway - old and paid for is fine.

It is a VERY BIG consideration for people like Googke/AWS/Azure etc, ya know the ones who are going to order the high end Xeon/Epyc chips in the thousands/tens of thousands.

In Intel Emeral Rapids I can get 64 core per CPU or I can get 128 Core from AMD in a single socket. So in a 2U design it will be common to see 128 Intel vs 256 AMD cores. Now lets buld out a data center on their scale. How many Us are now required for say 15k cores.

Now every extra Intel system requires more power connections,, more rack space, more networking etc etc etc and an not inconsequental increase or cooling requirementsof CPU power draw as its the heat from the PSUs, Mainboards, add in cards etc extra per machine.

This is where desiging a server is a LOT more complicated of "MOAR CORE MOAR BETTA" as in AMD chips you have frequency focused or cache focused designs so you really have to benchmark your workloads and decide if the extra frequency is worth the core count loss etc.

I don't know why this new am5 platform would be D.O.A. At least in Linux, the performance is there.

All I see is better performance. If I had the money, I would change my system in a heart beat.

From my perspective people seemed to have gone 7700x was equal too or better than the 5800x3d before it (marginally) so the 9700x should easily beat the 7800x3d.

People have missed the trees for the forest in that regards as AMD have released a part that is
A it is better than preceeeding parts,
B fixes all the issues with "power draw" and "temperatures"
C is a drop in replacement/compatible with hardware already out
D pricing hasnt magically gone up cause number is bigger

Now has it been a perfect release? No, far from it! Its plagued by the same problems the X3D parts had when they released. They cherry picked the weaker value parts to be released first before releasing the parts people really want of that family. In the normal family people are really wanting to see the higher core parts (9950x) and I am happy to eat my hat if they dont try the reverse when the X3D parts get their release in the coming months where they release the 9950X3D first when people really want to see the 9800x3d.

 

[Darmok N Jalad]

I’m going to chalk this up to the Windows Scheduler, which, IMO, is garbage. To be a broken record, I witness this pretty much every single day at work when my 2P-8E core i7 Adler Lake laptop fails to use all the cores in a multithreaded task. It often maxes the 2 real cores, but leaves the HT cores and E cores underutilized, resulting in a crawl when doing a basic task like copy/paste in excel. Doesn’t Mike at Phoronix get much better results on Linux? That tells you a lot right there. Kinda makes me wonder if the Windows scheduler is already prepped for Arrow Lake and the lack of SMT.

AMD Ryzen 5 9600X & Ryzen 7 9700X Offer Excellent Linux Performance Review - Phoronix

www.phoronix.com

When taking the geometric mean of those nearly 400 raw benchmark results, it sums up the greatness of Zen 5 with the Ryzen 5 9600X and Ryzen 7 9700X processors. The Ryzen 7 9700X delivered 1.195x the performance of the Core i5 14600K competition or 1.15x the performance of the prior generation Ryzen 7 7700X. The Ryzen 5 9600X came in at 1.35x the performance of the Core i5 14500 and 1.25x the performance of the Ryzen 5 7600X. Or if still on Zen 3 for comparison, the Ryzen 5 9600X was 1.82x the performance of the Ryzen 5 5600X.

 

[SalemTheKit]

Would be interesting if the x3D part will keep the same clocks as 9700x considering its just an 65W part , maybe they will release an 9800x too @ 105W or something.

The 7800x3d only uses 88w on all-core workloads via GN. Same as the non-x 7000 chips and the 9000X chips we have so far.