Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Heartbreaker]

Seems like either windows 11 or nvidia drivers needs to update the thread scheduling for Zen5 (?)
That's at least my take on this data..

That graph shows improvement of only 1.5% with SMT disabled... Which isn't exactly massive.

Depending on the set of games, sometimes SMT improves the result, sometimes it makes it worse. I doubt this will lead to a rewrite of Windows scheduler.

 

[Det0x]

That graph shows improvement of only 1.5% with SMT disabled... Which isn't exactly massive.

Depending on the set of games, sometimes SMT improves the result, sometimes it makes it worse. I doubt this will lead to a rewrite of Windows scheduler.

Updated screenshots
Its 1.5% at 4k res, higher at lower res

Some more screens

 

[fastandfurious6]

superb .1% lows, sign of a superior CPU

needs a David Huang level analysis

 

[LightningZ71]

You might do better with something like process lasso to direct other processes away from the second thread on the two prime cores. It's too bad that you can't just designate some cores as SMT off.

 

[Tuna-Fish]

I think I read in some online posts that higher tREFI can result in hotter DIMMs. Is it possible to do 65536 tREFI without special cooling on DIMMs?

No, it reduces heat. But it increases sensitivity to heat, which means you have to worry more about cooling.

tREFI is the refresh interval, or how long a single row go between refreshes. The caps that hold charge in DRAM are leaky, and tREFI needs to be short enough that if a cap has charge, there is still sufficient charge in the cap to sense as 1 just before the refresh triggers. Because of fairly basic physical reasons, this time is extremely sensitive to temperature. The hotter your modules get, the shorter it needs to be.

Because a bank that is refreshing cannot be accessed, this has a fairly direct effect on average bandwidth and worst-case latency, but also because the memory manufacturers are scared of lots of ram being faulty and returned, the default is set quite low.

The correct way to adjust tREFI is to set it as high as you dare, very exhaustively test memory while also running some torture test on GPU to get your machine hot, and then if the machine passes, pull it down ~25% just to be sure it still passes when your case has more dust and/or the ambient temperature is higher. The common advice of just setting it to 64k scares me.

 

[Heartbreaker]

Updated screenshots
Its 1.5% at 4k res, higher at lower res

And at lower res, you can see the same kind of thing for Zen 4 as well. This has been the reality of HT/SMT since it was first introduced decades ago. Sometimes it helps, sometimes it harms, but the overall consesnus is just turn it on and forget it.

A completely reworked core, is going to shift SMT behavior somewhat, but there isn't anything that significant going on here, except a whole lot of cope grasping at straws.

 

[Hail The Brain Slug]

The common advice of just setting it to 64k scares me.

Works on my machine!









I also have active cooling on my dimms.

 

[StefanR5R]

Two areas in which Zen 5 notably improves over Zen 4 have been pointed out here:

• vector arithmetic,
• web browsers/ JITs and the likes.

Glancing over the TPU review, it appears there is another area:

• databases.

MySQL TPC-C test:

9700X ......... 15,200 TPS (126 %)
7700X ......... 12,900 TPS (107 %)
7700 ........... 12,100 TPS (100 %)
7800X3D .... 11,700 TPS (97 %)​

MongoDB 6, time for 10M requests:

9700X ......... 67.5 s (141 %)
7700X ......... 90.0 s (106 %)
7700 ........... 95.4 s (100 %)
7800X3D .... 98.3 s (97 %)​

edit: added relative performance, Ryzen 7700 as baseline​

 

[yuri69]

Seems like either windows 11 or nvidia drivers needs to update the thread scheduling for Zen5 (?)

Don't think i've observed this behavior on my 16 core Zen5

Issues like this *might* be caused by a power-saver or bugged power plan/scheduling scheme. "Filling" all physical cores first should be the energy efficient strategy.

Two areas in which Zen 5 notably improves over Zen 4 have been pointed out here:

• vector arithmetic,
• web browsers/ JITs and the likes.

Glancing over the TPU review, it appears there is another area:

• databases.

MySQL TPC-C test:

9700X ........ 15,200 TPS​

7700X ........ 12,900 TPS​

7700 .......... 12,100 TPS​

7800X3D .... 11,700 TPS​

MongoDB 6, time for 10M requests:

9700X ........ 67.5 s​

7700X ........ 90.0 s​

7700 .......... 95.4 s​

7800X3D .... 98.3 s​

Check Phoronix's DB section, the DB gains are nice.

 

[Fjodor2001]

How many? And which of these are you running once a month or more often? Thanks in advance.

Video transcoding, source code compilation, ...

 

[Det0x]

And at lower res, you can see the same kind of thing for Zen 4 as well. This has been the reality of HT/SMT since it was first introduced decades ago. Sometimes it helps, sometimes it harms, but the overall consesnus is just turn it on and forget it.

A completely reworked core, is going to shift SMT behavior somewhat, but there isn't anything that significant going on here, except a whole lot of cope grasping at straws.

Did you miss the following quote ?

During the course of our testing, we observed that Windows 11 was scheduling workloads on the 9700X in a manner that would try to saturate a single core first, by placing workloads on each of its logical threads. Additionally, the placement would put load on the CPPC2 "best" or "second-best" core (gold and silver in Ryzen Master)—which makes sense. However, if a highly demanding single threaded workload runs on one core, scheduling another demanding workload on the second thread of that core will result in lower overall performance. It would be better to place them on two separate cores, where they each have access to the full resources of that core.

This is not how SMT scheduling have worked in the past, nor how it should work
(for reference, check the numbers for 7700X how it behaves with SMT ON/OFF as a comparison)

Kinda seems more like its someone else that's "coping and grasping at straws" as you put it.. Why is that ?

 

[StefanR5R]

Seems like either windows 11 or nvidia drivers needs to update the thread scheduling for Zen5 (?)
That's at least my take on this data..

During the course of our testing, we observed that Windows 11 was scheduling workloads on the 9700X in a manner that would try to saturate a single core first, by placing workloads on each of its logical threads.

Whoa, I didn't realize that Windows is still that bad. I am forced to use it (Win 10) as application launcher at work, and ignore it as best as I can…

 

[igor_kavinski]

Some more screens

I love how the situation totally changes with SMT off in Excel and Outlook, arguably two of the most used applications in offices all around the world:



7700X with SMT disabled is closer to 9700X SMT disabled in Outlook but loses any chance whatsoever in Excel, which really fills me with delight!

Now just gotta wait for Excel 2024 with AVX-512!

 

[DaaQ]

Did you miss the following quote ?

This is not how SMT scheduling have worked in the past, nor how it should work
(for reference, check the numbers for 7700X how it behaves with SMT ON/OFF as a comparison)

Kinda seems more like its someone else that's "coping and grasping at straws" as you put it.. Why is that ?

He didn't read the article.. hopefully he will. Probably won't.

 

[CouncilorIrissa]

Updated screenshots
Its 1.5% at 4k res, higher at lower res

Some more screens
View attachment 105032

View attachment 105033

View attachment 105034

View attachment 105035

View attachment 105036

That's a...non-insignificant gain in front-end bound workloads like browsing. I expected less.
2 decode clusters doing work?

 

[igor_kavinski]

Clearly something was very lacking in Zen 4.

9700X SMT disabled performance uplift @1080p:

Baldur's Gate 3 +6.78%

Remnant II +6.67%

Spiderman Remastered +17.17% (!!!!) and RT +15.82% (!!!!)

9700X SMT disabled performance uplift @1440p:

Baldur's Gate 3 +6.69%

Spiderman Remastered +17.22% (!!!!) and RT +18.73% (!!!!)

9700X SMT disabled performance uplift @ 2160p:

Baldur's Gate 3 +7.23%

Spiderman Remastered +5.16% and RT +8.10% (!!!!)

AND

+11.61% CS 1080p minimum fps

+12.63% Remnant II 1080p minimum fps

+60.27% (!!!!) Spiderman Remastered 1080p minimum fps

+22.50% (!!!!) Last of Us 1080p minimum fps

+8.30% Baldur's Gate 3 2160p minimum fps

+31.91% (!!!!) Spiderman Remastered 2160p minimum fps


And people are disappointed.

SMH

 

[CouncilorIrissa]

Clearly something was very lacking in Zen 4.

9700X SMT disabled performance uplift @1080p:

Baldur's Gate 3 +6.78%

Remnant II +6.67%

Spiderman Remastered +17.17% (!!!!) and RT +15.82% (!!!!)

9700X SMT disabled performance uplift @1440p:

Baldur's Gate 3 +6.69%

Spiderman Remastered +17.22% (!!!!) and RT +18.73% (!!!!)

9700X SMT disabled performance uplift @ 2160p:

Baldur's Gate 3 +7.23%

Spiderman Remastered +5.16% and RT +8.10% (!!!!)

AND

+11.61% CS 1080p minimum fps

+12.63% Remnant II 1080p minimum fps

+60.27% (!!!!) Spiderman Remastered 1080p minimum fps

+22.50% (!!!!) Last of Us 1080p minimum fps

+8.30% Baldur's Gate 3 2160p minimum fps

+31.91% (!!!!) Spiderman Remastered 2160p minimum fps


And people are disappointed.

SMH

If these figures are from TYC review, I wouldn't put too much stock into them yet. I'd wait for someone else to verify that.

 

[coercitiv]

And people are disappointed.

So you're buying today, right?

 

[igor_kavinski]

So you're buying today, right?

9800X3D for me

I don't like changing CPUs too often. I wish I could but installing CPUs is not fun, much to my dismay

 

[Hail The Brain Slug]

So you're buying today, right?

This is the very thing I've been railing against chatting with others. Your opinion of the launch doesn't matter if you never intended to buy.

Not to devalue or dismiss everyone's opinions, but if there was no circumstance under which you would have bought these new products, your opinion doesn't matter to AMDs bottom line.

Only people who would have bought but don't, or wouldn't but do make a difference to their bottom line.

 

[igor_kavinski]

If these figures are from TYC review, I wouldn't put too much stock into them yet. I'd wait for someone else to verify that.

@B-Riz already posted one graph of Baldur's Gate 3 minimum fps improving with 9600X. At this point, at least I don't need further proof that there's some VERY good and VERY practical improvements in Zen 5, enough to warrant an upgrade for some, if not all, Zen 4 users.

 

[Heartbreaker]

Some more screens

Now you are cherry picking the screens where SMT does worse, ignoring the ones where it does better.

If they would have included Intel CPUs with HT in their you would likely see a similar thing. SMT is never an exclusive win.

If 9700X is behaving worse with SMT on than previous cores, that's Zen 5 problem not that HT/SMT scheduling is wrong after it's been on the market for decades.

 

[Tuna-Fish]

That's a...non-insignificant gain in front-end bound workloads like browsing. I expected less.
2 decode clusters doing work?

I don't know if the two decode clusters can actually serve a single thread in practice. However, the µop cache does seem to work, in that the machine can do two taken branches per clock from µop cache, which is very useful on it's own when running interpreters.

 

[igor_kavinski]

This is the very thing I've been railing against chatting with others. Your opinion of the launch doesn't matter if you never intended to buy.

Zen 5 has launched at an unfortunate time. If we had data to see how many users upgraded in the past 6 months from their aging platform to 7800X3D or any other Zen 4 CPU, we would understand why the Zen 5 reception is lukewarm. AMD should've launched Zen 5 in January. This is a market where waiting just hurts you more because a lot of people with the upgrade itch don't wait for impending launch of new CPUs. There's a reason why it's called an "itch". People just want something new and they want it NOW.

 

[CouncilorIrissa]

I don't know if the two decode clusters can actually serve a single thread in practice. However, the µop cache does seem to work, in that the machine can do two taken branches per clock from µop cache, which is very useful on it's own when running interpreters.

Gotta wait for C&C article to finally get some clarity on this.