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

[DisEnchantment]

Speculate at will

 

[PJVol]

IIRC the decoder in GLC is power gated like 80% of the time.... though I wonder what applications they are using to get that figure lol

Not sure what glc is, but modern cpus have performance counters for almost everything, let alone dedicated registers used by firmware to control power or clock gating at the low level.
For example AMD has uProf for monitoring.

 

[JoeRambo]

IIRC the decoder in GLC is power gated like 80% of the time.... though I wonder what applications they are using to get that figure lol

How Zen 2’s Op Cache Affects Performance

Banner image credit goes to Fritzchens Fritz and his amazing die shots Recent AMD and Intel high performance CPUs implement an op cache that remembers decoder output and functions like a L0 instruc…

chipsandcheese.com

Not V-Ray i guess LOL. In fact Intel has probably found that their marketing cool aid percentages does not reflect actual uOP cache hit rates and is actively taking steps to improve front end in decode department.

 

[igor_kavinski]

Not sure what glc is

Golden cove

 

[Mopetar]

IIRC the decoder in GLC is power gated like 80% of the time.... though I wonder what applications they are using to get that figure lol

80% doesn't seem that surprising. Anything with a lot of loops in it will run over the same instructions again and again. There's no need to decode those instructions again if you've already done it recently. Commonly used functions/procedures might stick around in the cache as well.

I'd imagine AMD's design is similarly aggressive about turning off the decoder in situations where it's not actually needed. The circuitry for those things has to be pretty massive given the size and complexity of the x86 instruction set.

 

[Fjodor2001]

Methink that AM5 s 170W TDP/230W PPT was implemented with more cores in mind for future SKUs, it s really useless for 16C as demonstrated by benchmarks, they just pushed the 7950X for competitive reasons, but overall that ended being a discutable choice that only brought useless complications.

So then the question is when AMD will bump core count on AM5 to more than 16C? Zen5 or Zen6?

Is there any reason not do do it already with Zen5? Possibly they'll wait until Zen6 which is when the next process node jump will happen, so there is more transisors per die area available. But won't that be too late, when they are facing potential 16P+32E from Intel before Zen6 will be released?

 

[adroc_thurston]

So then the question is when AMD will bump core count on AM5 to more than 16C?

No.

Is there any reason not do do it already with Zen5

You don't need more cores.

 

[Glo.]

You don't need more cores.

WHAT DO YOU MEAN "YOU DON'T NEED MOAR COREZ"??!!!!

 

[Fjodor2001]

No.

You don't need more cores.

Oh no, he's back.

This is so funny.

I remember when Intel was doing 4C/8T forever. They claimed nobody need more than 4C.

Then AMD revolutionized the scene and introduced 8C Zen and soon after 16C.

Now AMD is stuck on 16C for several CPU generations. According to adroc_thurston because nobody need more than 16C.

Meanwhile Intel is moving along to 8P+16E and soon thereafter 8P+32E.

Such a hilarious plot twist! The tables are about to turn again (assuming AMD will stay on 16C for the upcoming AM5 desktop CPU generations, which I sincerely doubt!).

 

[adroc_thurston]

I remember when Intel was doing 4C/8T forever

That really wasn't the issue.
Intel was doing meagre 1t bumps.

 

[Fjodor2001]

That really wasn't the issue.
Intel was doing meagre 1t bumps.

There's no other way to substantially increase MT performance than bumping core count.

Going from 4C to 8C bumped MT performance by ~100%. Good luck achieving that by IPC of frequency increase in the same time frame.

 

[adroc_thurston]

There's no other way to substantially increase MT performance than bumping core count.

And bumping wattage.
ULV went from 42 to 282 and nT needle barely moved at 15-18W.

 

[Fjodor2001]

And bumping wattage.
ULV went from 42 to 282 and nT needle barely moved at 15-18W.

Not really substantial, w.r.t. to desktop CPUs. AMD stayed on 65W for 8C Zen 1700, and 105W for 16C Zen2 3950X. Very low power consumption increase when considering the core count bump and the competition from Intel at that time.

And anyway, you have previously claimed desktop PC users do not care about power consumption anyway (which I do not agree with). Have you changed your mind on that?

 

[Glo.]

Not really substantial, w.r.t. to desktop CPUs. AMD stayed on 65W for 8C Zen 1700, and 105W for 16C Zen2 3950X. Very low power consumption increase when considering the core count bump and the competition from Intel at that time.

And anyway, you have previously claimed desktop PC users do not care about power consumption anyway (which I do not agree with). Have you changed your mind on that?

Considering that majortity of mainstream platforms are Mobile - Adroc has a point.

We always think ONLY about desktop as mainstream, where it rapidly becomes nieche market.

 

[adroc_thurston]

Not really substantial, w.r.t. to desktop CPUs

816 RPL-S chugs 250W like its no one's business.
DT 4c was 91w and that's rather generously for 22/14mm parts.

 

[Mopetar]

WHAT DO YOU MEAN "YOU DON'T NEED MOAR COREZ"??!!!!

If you really did, you'd buy a Threadripper or an Epyc.

The problem isn't that AMD couldn't increase the core count, but that they'd need a new platform that has additional memory channels to keep it fed or a big increase in LLC to alleviate bandwidth bottlenecks.

I don't know if there's enough of a market to support AMD developing another socket to give you more cores, but not as many as Threadripper. If you had anything that could legitimately use ~32 cores, then it almost always means it would scale to 64 or even 128 and TR/Epyc is worth considering.

Now that Intel is starting to move to tiles, it means that they can actually compete in that space again. Maybe they'll offer those kinds of products that sit in the middle or more likely it means AMD will be more price competitive so the price jump from Ryzen to TR isn't as significant.

 

[itsmydamnation]

I recon we should go get posts from 2014 and replace bulldozer with e-core and sandy bridge with zen5, oh how the turn tables. . .

 

[Joe NYC]

Oh no, he's back.

This is so funny.

I remember when Intel was doing 4C/8T forever. They claimed nobody need more than 4C.

Then AMD revolutionized the scene and introduced 8C Zen and soon after 16C.

Now AMD is stuck on 16C for several CPU generations. According to adroc_thurston because nobody need more than 16C.

Low percentage of users benefits from > 8 cores.
Even lower percentage benefits from > 16 cores.

The market is just not there, because the software (which is even more immovable) than hardware) just doesn't benefit from more cores.

But what software always benefits from is higher single core performance.

Meanwhile Intel is moving along to 16P+8E and soon thereafter 16P+32E.

Such a hilarious plot twist! The tables are about to turn again (assuming AMD will stay on 16C for the upcoming AM5 desktop CPU generations, which I sincerely doubt!).

Where do you get 16P from? I haven't seen that.

16P tile would be interesting for Intel processors. They could finally retire the E-core nonsense...

 

[gruffi]

Well the problem here is that it's not 35% ipc uplift it's only 25% at 5.225 Ghz, at least according to the chart anyway.

Which problem? I don't see any problem yet. It's all just rumors and speculation. We don't have any facts about IPC or clock speed yet.

 

[gruffi]

So then the question is when AMD will bump core count on AM5 to more than 16C? Zen5 or Zen6?

My guess would be not before Zen 6.

Is there any reason not do do it already with Zen5? Possibly they'll wait until Zen6 which is when the next process node jump will happen, so there is more transisors per die area available. But won't that be too late, when they are facing potential 16P+32E from Intel before Zen6 will be released?

I don't think that Intel will go >8 p-cores anytime soon. They will likely push e-cores. And I think that even 15th gen won't have more than 8p+32e on the mainstream desktop platform. Which is expected when? Somewhere 2025, about one year after Zen 5? So, I don't see AMD under pressure to increase core counts before Zen 6. 13th gen looks to be launched like half a year before Zen 5. But it will be a joke compared to the latter one. Even 14th gen (if there will ever be a desktop family based on it) could be only similar or worse than Zen 5. Which is also some kind of sad. Because I would like to see a 12 or 16 core Zen CCD. But before AMD doesn't increase the total count of p-cores they likely won't change the number of p-cores in the CCD.

But who knows. Maybe we see something like 8p+16/24/32c with Zen 6. Made of one p-core CCD and one c-core CCD. Which might look similar to what Intel is doing. But AMD's c-core is much stronger than Intel's e-core. So, even an 8p+16c Zen 6 could be a quite decent improvement over 16p Zen 5.

 

[adroc_thurston]

Somewhere 2025

Q4'24.

 

[Timmah!]

You don't need more cores.

I would be a judge of what i need, thank you.

If you really did, you'd buy a Threadripper or an Epyc.

The problem isn't that AMD couldn't increase the core count, but that they'd need a new platform that has additional memory channels to keep it fed or a big increase in LLC to alleviate bandwidth bottlenecks.

I don't know if there's enough of a market to support AMD developing another socket to give you more cores, but not as many as Threadripper. If you had anything that could legitimately use ~32 cores, then it almost always means it would scale to 64 or even 128 and TR/Epyc is worth considering.

Now that Intel is starting to move to tiles, it means that they can actually compete in that space again. Maybe they'll offer those kinds of products that sit in the middle or more likely it means AMD will be more price competitive so the price jump from Ryzen to TR isn't as significant.

I think AMD could still get away with at least 24C on mainstream platform and 2 memory channels. I may be wrong though. Additionally, nothing prevented them from moving to 4 channels on mainstream platform anyway. Stuff used to trickle down from server to desktop in the past, now the server adds more channels like with each new generation, currently on what, 12?...while desktop gets stuck in stone-age at 2.

Regarding TR, i would buy it. If only it was not 3x more expensive, while providing 1,5x more cores, which are one gen older than the desktop parts. So ultimately, even in MT, 24C TR is not really faster than 16C Ryzen.

 

[coercitiv]

I think AMD could still get away with at least 24C on mainstream platform and 2 memory channels. I may be wrong though. Additionally, nothing prevented them from moving to 4 channels on mainstream platform anyway. Stuff used to trickle down from server to desktop in the past, now the server adds more channels like with each new generation, currently on what, 12?...while desktop gets stuck in stone-age at 2.

You're advocating for morphing the mainstream consumer platform into HEDT. First ask for more cores, then more memory channels, then more PCIe lanes. This would only lead to HEDT prices, for both motherboards and CPUs. (from both BoM increase and market positioning, since they would be workstation/server material).

Like I said before, we would not be having this conversation if HEDT lineups from both companies were alive and reasonably priced. The reality is you have a specific need, and these two companies want you to show them your wallet. More cores on the mainstream consumer platform is just a band-aid, in the end either you or AMD/Intel will blink first and you will get your true workstation. We'll just have to see at what price.

 

[Timmah!]

I think its matter of semantics, what one wants to call it. 4 channels was a feature of X99 or maybe even X79 more than 10 years ago. Its not to far fetched to ask to come to regular desktop already by now. I mean, if its the reason why we are stuck at 16C.
I already paid HEDT price for my current motherboard btw. 700 EUROs for Asus x670e Hero.

 

[yuri69]

It's funny how AMD states "Can't do >16c it's too much for the 2ch bandwidth". Yet, Intel got a 8+16 setup and plans a 8+32 cores - all that using 2ch.

AMD's IF bandwidth limitations are apparent but RAM?

 

[leoneazzurro]

It's not that AMD "can't do" but more "does not want to" and it is true that for most of the users more than 16 core on desktop is more than enough. While having higher 1T performance is never enough for most of the desktop applications, including the highest market driving one, that is, gaming. Intel got a 8+16 which is on the same level of the 16 with AMD so the comparison is a bit pointless. And we don't know if and how much a 8+32 would be limited by RAM bandwidth (and maybe a new socket is required?).
All in all, I understand those who would like more cores at lower price point. Unfortunately, it would be also true that in that case the platform costs would have been probably higher, which is not good for a platform like AM5 which goes from entry level to "almost HEDT" and it's built to last for a few CPU generations.