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

[DisEnchantment]

Speculate at will

 

[MarkPost]

https://twitter.com/x/status/1848096683034628344

Still wondering what this setting actually does.. I can see higher gaming performance with disabling CCD1/SMT, but higher memory bandwidth ?!
Normally the only way this happen is with faster FCLK 🤔

well on my 9950X this option disables SMT and CCD1 all together 🧐

scared when opened task manager and only eight threads were there 😅

dont know what other evil things it does 🤣

 

[Hitman928]

well on my 9950X this option disables SMT and CCD1 all together 🧐

scared when opened task manager and only eight threads were there 😅

dont know what other evil things it does 🤣

It turns on these options as well, but I have no idea what they do.

 

[arcky]

A fourth GB6 score emerges: https://browser.geekbench.com/v6/cpu/8477557

Boost clock 5.247

So we have 2 x870e crosshairs on gb6 6.2.2 boosting to 5.3 - could be the same person testing different ram configs (one run is 48gb the other is 32)

Then we have one Taichi and one Valkyrie, on gb6 6.3, boosting to 5.25

 

[MarkPost]

It turns on these options as well, but I have no idea what they do.

View attachment 110274

Didnt realize the first one because I had it enabled already. But the second doesnt appear here

 

[tsamolotoff]

It turns on these options as well, but I have no idea what they do.

Afaik it sets REFI to some high value like 50000 and probably changes some other timings (so if you already have them tuned, it'd do nothing most probably)

 

[igor_kavinski]

dont know what other evil things it does 🤣

So as per the clue given by @tsamolotoff , give us screenshots of ZenTimings before and after.

Please

 

[ju-rek]

(cb23 mt is 28.9% higher than my best, cheesed 7800x3d score of 19594 at -40 CO)

CO -40

 

[sl0519]

Isn't there supposed to be an announcement today regarding X3D?

 

[Saylick]

Isn't there supposed to be an announcement today regarding X3D?

Maybe AMD saw the pitiful gaming performance of Arrowlake and decided to scrap their plans to announce the 9800X3D?

Nah, I’m just joking. There should be an announcement today, if not relatively soon.

 

[igor_kavinski]

Isn't there supposed to be an announcement today regarding X3D?

Maybe they postponed it. They really must've been scared of the imaginary leprechaun they expected Intel to pull out of their behinds

Lisa Su in an internal memo: "Crisis averted. Everyone go back to sleep. Or working. Or doing whatever. And please, get us a better mole inside Intel so we don't have to take emergency preventive measures at such short notice in future."

 

[StinkyPinky]

I'm more interested in the 9900X3D and 9950X3D. Any word to the rumor that both chiplets will have the 3D cache this time around? This makes them hugely more interesting, especially the 9900X3D which would actually have a reason to exist now.

 

[gdansk]

I'm more interested in the 9900X3D and 9950X3D. Any word to the rumor that both chiplets will have the 3D cache this time around? This makes them hugely more interesting, especially the 9900X3D which would actually have a reason to exist now.

Except as a mitigation to scheduling problems, what makes dual 3D cache more interesting? Doesn't it come with a clock speed reduction?

 

[StinkyPinky]

It will, but I think it's a good compromise for those of us that like high core counts but also like to game.

 

[MarkPost]

So as per the clue given by @tsamolotoff , give us screenshots of ZenTimings before and after.

Please

sure

before (EXPO enabled, all timmings auto):



After (EXPO enabled, all timmings auto):



So as @tsamolotoff said, it sets tREFI to a higher value and tRFC to a lower value.. and thats it

btw @Hitman928, I set "EXPO/XMP high bandwidth support" to disabled, and it was not enabled after enabling this x3d turbo mode thing (and restarting) 🧐

 

[Thunder 57]

I'm more interested in the 9900X3D and 9950X3D. Any word to the rumor that both chiplets will have the 3D cache this time around? This makes them hugely more interesting, especially the 9900X3D which would actually have a reason to exist now.

I very much doubt that. This feels very akin to the Zen 5 "32% IPC" rumor. People get excited only to be let down.

 

[gdansk]

It will, but I think it's a good compromise for those of us that like high core counts but also like to game.

I guess, I'd be more interested if AMD found a way to have 3D cache variants dynamically disable 3D cache and clock higher automatically without requiring a restart. But I'm not sure that's possible.

 

[Jan Olšan]

I'm more interested in the 9900X3D and 9950X3D. Any word to the rumor that both chiplets will have the 3D cache this time around? This makes them hugely more interesting, especially the 9900X3D which would actually have a reason to exist now.

Well, it's probably a rumor or somebody's speculative guess too somewhere, but if you are talking about the news techPowerUp published based on BenchLife's articel based on some twitter post with a screenshot from bilibili, then that was purely TPU misreading the source.

The original table was quite confusing but only TPU got caught in that. It did not mean that 9950X3D and 9900X3D will have 2x V-Cache, actually it was the opposite.

TL;DR not true, unless some surprise happens.

 

[arcky]

CO -40

View attachment 110279

Nice. Only source of the 83 point (.4%) difference would be Win 22H2 vs 24H2, and your ram having slightly lower latency being 32gb thus being able to run lower trfc. Interesting that 2000 fclk also appears not to be a hindrance vs 2200

 

[Josh128]

sure

before (EXPO enabled, all timmings auto):

View attachment 110282

After (EXPO enabled, all timmings auto):

View attachment 110283

So as @tsamolotoff said, it sets tREFI to a higher value and tRFC to a lower value.. and thats it

btw @Hitman928, I set "EXPO/XMP high bandwidth support" to disabled, and it was not enabled after enabling this x3d turbo mode thing (and restarting) 🧐

What differences, if any, do these make to your mem read/write bandwidth and latency?

 

[yottabit]

I very much doubt that. This feels very akin to the Zen 5 "32% IPC" rumor. People get excited only to be let down.

Well AMD did hint that the 2 CCD V cache parts were “special” no?

What else could make them special?

 

[igor_kavinski]

What else could make them special?

Almost similar boost clocks to vanilla parts?

Both higher quality and top binned CCDs?

Better thermal dissipation and even lower temps than vanilla parts?

6 GHz fmax (venturing into fantasyland here)?

Premium quality iOD allowing easy DDR5-9000 EXPO set and forget?

One V-cache die but covering both CCDs so both get access to the same frequently re-used data?

3nm CCDs?

One regular Zen 5 CCD with V-cache die and one Zen 5c CCD with 16 cores without V-cache?

World's first 10 core 3nm CCDs???

 

[Timorous]

Almost similar boost clocks to vanilla parts?

Both higher quality and top binned CCDs?

Better thermal dissipation and even lower temps than vanilla parts?

6 GHz fmax (venturing into fantasyland here)?

Premium quality iOD allowing easy DDR5-9000 EXPO set and forget?

One V-cache die but covering both CCDs so both get access to the same frequently re-used data?

3nm CCDs?

One regular Zen 5 CCD with V-cache die and one Zen 5c CCD with 16 cores without V-cache?

World's first 10 core 3nm CCDs???

They will have red stripes because red stripes makes things faster.

 

[Det0x]

It turns on these options as well, but I have no idea what they do.

View attachment 110274

I'm pretty sure "high bandwidth support enabled" just rises tREFI, would explain the higher bandwidth

"low latency support" is GDM disable ?
"Low latency support" is changing nitro values from lets say 2-3-1 to 1-2-1 ?

If thats the case, this new X3D turbo option does nothing for "us tweakers" that already run a optimized system, good for the rookies tho i guess
Would also help the reviewers, so thats a pluss

*edit*
Can see we already have it figured out 👍

 

[MarkPost]

What differences, if any, do these make to your mem read/write bandwidth and latency?

lower latency, higher mem write but lower read and copy

"X3D Turbo Mode" disabled (EXPO profile loaded with all the rest mem settings set to auto):


"X3D Turbo Mode" enabled (EXPO profile loaded with all the rest mem settings set to auto):

 

[arcky]

lower latency, higher mem write but lower read and copy

"X3D Turbo Mode" disabled (EXPO profile loaded with all the rest mem settings set to auto):
View attachment 110319

"X3D Turbo Mode" enabled (EXPO profile loaded with all the rest mem settings set to auto):
View attachment 110320

So turbo increased trefi and lowered trfc to get you marginally better latency, and shut off a whole ccd crippling your read/copy bandwidth lol