Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[coercitiv]

Oh well, the insults are already starting, it's going to be a long summer.

 

[Abwx]

Oh well, the insults are already starting, it's going to be a long summer.

To say that something that is taking 31% less time is 31% faster is an insult to intelligence, or even worse as stated by someone else, that 45% less time is 31% faster...

I would call this trolling by ignorance...

For those who dont understand let s say that a car that goes at 145 km/h take 31% less time for a given trip that a car that run at 100km/h.

Speed is 45% faster and time for the ride will be 31% lower..

 

[poke01]

Does the footnote state at what clocks did the Zen4 part run? No, there is your clue.

If we assume AMD ran the core at 5.5Ghz to get 15% improvement that is shockingly sad. Really sad!

Let's hope it was below was 5Ghz.

The closest we can compare on the same TSMC 5NM node is the M1 at 3.2Ghz it manages 1498 in Cinebench.
The 5950X at 3.4Ghz scores 1684. This AMD score for sure used more power than M1.
Apple still retains Pref/W crown. Seeing Apple can mange that score in an iPad Air/ Pro and Macbook air AMD and Intel need to up the game.

 

[uzzi38]

It's obvious I do, that is why I'm questioning your emphasis on perf/w when the current signs indicate that consumer oriented Zen4 SKUs are likely to clock high to compete.

I was replying to a comment talking about Apple, where it's the mobile parts that actually matter? The desktop parts aren't important for that discussion at all, and rather the perf/W for lower power targets absolutely does matter.

 

[inf64]

If we assume AMD ran the core at 5.5Ghz to get 15% improvement that is shockingly sad. Really sad!

Let's hope it was below was 5Ghz.

The closest we can compare on the same TSMC 5NM node is the M1 at 3.2Ghz it manages 1498.

Apple still retains Pref/W crown.

They labeled everything so vaguely on purpose. The parts are scheduled to launch in Sept/Oct, and they have competitor launching their refreshed lineup at that time.
No, the IPC is not just 5% higher on average, that would be a failure of monumental proportion. AMD knew what performance targets they had to hit in order to stay on top of Alder lake and Raptor Lake targets, and they knew this for more than a year:

From Computerbase latest summary:

ST:


MT:


Gaming:

 

[poke01]

To say that something that is taking 31% less time is 31% faster is an insult to intelligence, or even worse as stated by someone else, that 45% less time is 31% faster...

idk AMD said up to 31% faster in their slides. They did not say it takes 31% less time.

 

[exquisitechar]

To say that something that is taking 31% less time is 31% faster is an insult to intelligence, or even worse as stated by someone else, that 45% less time is 31% faster...

I would call this trolling by ignorance...

For those who dont understand let s say that a car that goes at 145 km/h take 31% less time for a given trip that a car that run at 100km/h.

Speed is 45% faster and time for the ride will be 31% lower..

Sorry, my brain stopped working...I know the basics. You’re right, I just looked at a report of it being 31% faster and wrote that instead of thinking.

 

[insertcarehere]

https://twitter.com/x/status/1528635957678854147

Robert Hallock is according to LinkedIn "director of technical marketing" at AMD.

 

[Abwx]

idk AMD said up to 31% faster in their slides. They did not say it takes 31% less time.

Dunno what comparison you re talking about, the only one posted about Blender is this one, they are talking of 31% less time elapsed...

They said 297 seconds for the 12900K and 204 seconds for the 7000.

That s 31% less time and 45% higher speed than ADL.

AMD Ryzen 7000: Raphael mit Zen 4 soll Single-Core „über 15 %“ zulegen

AMD hat Ryzen 7000 mit Zen 4 enthüllt: Die Serie bietet „mehr als 15 Prozent“ mehr Single-Thread-Leistung sowie immer eine RDNA-2-iGPU.

www.computerbase.de

 

[JoeRambo]

Does the footnote state at what clocks did the Zen4 part run? No, there is your clue.

Sorry to burst Your bubble, but AMD specifically told HUB in prebreafing that it is a mix of clock and IPC improvement. So it was not ISO clock in any way. But You are free to claim that it was 5051mhz for Z4 and 5050 for Z3.

 

[inf64]

Yeah, AMD worded it oddly in the blender slide - Zen4 finished the task in 31% less time than ADL (ADL is base, a 100%; Zen4' score is 0.68 or 68% of the base). So Zen4 is 45% faster than ADL as it takes 45% more seconds than Zen4's base score of 204 (204x1.45) to get to ADL's benchmark score (in seconds).

Sorry to burst Your bubble, but AMD specifically told HUB in prebreafing that it is a mix of clock and IPC improvement. So it was not ISO clock in any way. But You are free to claim that it was 5051mhz for Z4 and 5050 for Z3.

No need to be sorry, you are not bursting anything. If you think that Zen4 has 5% higher IPC om average than Zen3, that's fine. Time will tell who was right.

I expect that the average IPC is no less than 15%, using the same methodology AMD used when they calculated the Zen2->Zen3 uplift. It's possible that R23 is just on the lower end of the scale for ST IPC uplift at ISO clocks.
The clocks are more or less known by now.

So in summary this is what I expect for Zen4:
- ST performance ~25% higher on average for the top 16-core Zen4 part vs 5950X
- MT performance ~30% higher on average for the top 16-core Zen4 part vs 5950X
- gaming performance ~20% higher on average for the top 16-core Zen4 part vs 5950X

 

[Karnak]

According to Ian (and he was talking with Hallock about it) the 170W is refering to PPT. So going by AMD's TDP -> PPT ratio of ~1.35x we're looking at a 125W TDP max. for the CPU's.

 

[JoeRambo]

No need to be sorry, you are not bursting anything. If you think that Zen4 has 5% higher IPC om average than Zen3, that's fine. Time will tell who was right.

My claim was and is 5-10% based on GB leak with no other knowleage. While Yours is running around forums, stating "massive ipc increase" whatever that means in numbers.
Let's just hope this Zen4 IPC prediction does not end up in Your signature as 99% correct as it is now plainly obviuos it is not.

That is great news right there: being able to run tight DDR5 6000CL30 bodes well for platform, isn't official speed rumoured as 5400?

 

[inf64]

My claim was and is 5-10% based on GB leak with no other knowleage. While Yours is running around forums, stating "massive ipc increase" whatever that means in numbers.
Let's just hope this Zen4 IPC prediction does not end up in Your signature as 99% correct as it is now plainly obviuos it is not.



That is great news right there: being able to run tight DDR5 6000CL30 bodes well for platform, isn't official speed rumoured as 5400?

"Running around stating "massive ipc increase"" is 2 posts in 3 months on this one forum. Are you Ok? You seem a bit emotional for some reason. My track record is nearly perfect btw, proof is in my signature. Feel free to check it out. I haven't really posted much this year, so I don't know what you are referring by your statement.

Here is my performance projection :
- ST performance ~25% higher on average for the top 16-core Zen4 part vs 5950X
- MT performance ~30% higher on average for the top 16-core Zen4 part vs 5950X
- gaming performance ~20% higher on average for the top 16-core Zen4 part vs 5950X

 

[mikk]

My claim was and is 5-10% based on GB leak with no other knowleage. While Yours is running around forums, stating "massive ipc increase" whatever that means in numbers.
Let's just hope this Zen4 IPC prediction does not end up in Your signature as 99% correct as it is now plainly obviuos it is not.

Do you have a link to this GB leak?

 

[Abwx]

That is great news right there: being able to run tight DDR5 6000CL30 bodes well for platform, isn't official speed rumoured as 5400?

DDR6000-CL30 is for the 12900K, the 7000 use DDR6400-CL32.

 

[JoeRambo]

My track record is nearly perfect btw, proof is in my signature. Feel free to check it out. I haven't really posted much this year, so I don't know what you are referring by your statement.

This is the reason i specifically singled out Your performance projections, due to Your previuos good estimates. And it collides with what is known from GB leaks, as "massive IPC increase" requires equally massive architectural changes.
On topic of "emotional" -> i am simply non native english speaker and fighting autocorrect sometimes too. Just wanted to point out that peoples hopes of majour IPC gains were completely unfounded. That does not preclude nice performance advances due to clock increases in ST/MT and when combined with modest IPC change of "up to 10%" that will still result in solid performance gains.

 

[inf64]

This is the reason i specifically singled out Your performance projections, due to Your previuos good estimates. And it collides with what is known from GB leaks, as "massive IPC increase" requires equally massive architectural changes.
On topic of "emotional" -> i am simply non native english speaker and fighting autocorrect sometimes too. Just wanted to point out that peoples hopes of majour IPC gains were completely unfounded. That does not preclude nice performance advances due to clock increases in ST/MT and when combined with modest IPC change of "up to 10%" that will still result in solid performance gains.

I posted what I expect from Zen4 above. Let's see by how much I will miss this time . The performance estimates are for the computerbase data.

 

[JoeRambo]

Do you have a link to this GB leak?

It's the legendary Gigabyte leak of AMD docs. Summarized quite well here:

Details on the Gigabyte Leak

Recently, a ransomware group leaked data from Gigabyte in an attempt to extort payment.

chipsandcheese.com

 

[yuri69]

Yup, back in 2021 people were puzzled about Zen 4 changes pointing to a Zen 3 + bigger TLB + doubled sets of L2 + AVX512 + new IF/memory stuff. It was attributed to an early leak or low level optimization.

 

[coercitiv]

I was replying to a comment talking about Apple, where it's the mobile parts that actually matter? The desktop parts aren't important for that discussion at all, and rather the perf/W for lower power targets absolutely does matter.

I understand that, and I'm telling you they will boost as high as possible when their avenue for higher ST performance is mostly frequency driven. The fact that battery life under light load will be excellent won't change the thermals under serious load. Anyway, let's agree to disagree on this.

 

[yuri69]

AMD knew what performance targets they had to hit in order to stay on top of Alder lake and Raptor Lake targets, and they knew this for more than a year:

From Computerbase latest summary:

Duh? Zen 4 was likely already taped out when ALD results appeared. There is not much you can do 1-1.5 year prior launch - bug fixes, tweaking, tuning the perf/W curve.

 

[DisEnchantment]

AMD putting a number of >15% to manage expectations early. If they are sandbagging they could just have mentioned non quantifiable gain.
So for the folks still thinking otherwise, this is it. It is the same marketing apparatus preparing for damage control. I will wait for the reveal though before calling it an epic fail.

 

[inf64]

Duh? Zen 4 was likely already taped out when ALD results appeared. There is not much you can do 1-1.5 year prior launch - bug fixes, tweaking, tuning the perf/W curve.

AMD had consistent >15% IPC gains year on year with their Zen roadmap. It was specifically their goal to outpace the industry average of the IPC curve. For them to suddenly massively underperform with Zen4 is totally out of character, especially if they had a new node to play with (which grants them the needed die area for beefed up core logic).

So in summary:
- Zen1/Zen1+ brought >50% over previous gen (Steamroller)
- Zen2 brought ~15% IPC on average Vs Zen1, and 2x the AVX throughput, with 14nm-> 7nm node switch
- Zen3 brought ~19% IPC on average Vs Zen2 , on the same node
- you guys expect that Zen4 deviates from the above trend of >15% IPC uplift with next gen core while they switch from 7nm to 5nm? That's seriously underestimating AMD's capabilities, even IF they wanted to pursue the "higher clocks"/speed demon design. Intel pursued much higher clocks and they achieved ~17-19% IPC targets with Icelake and Alder Lake

 

[leoneazzurro]

Epic fail for what? I think some reactions are way exaggerated.