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!

 

[biostud]

Oh, already revealed?

Like I feared, Zen 4 is just a small upgrade over Zen 3 focusing on it's weak points, AVX512, "AI", and integrated video encoder to help some tasks.
15% percent with clock increase? Zen Reign is over.
The saving grace is that all cores are equal, but makes me wonder if the 16 core limit means that there's a possibility of an extra 24 and 32 cores chip using Zen4c cores.

So Raptor lake and zen4 will match each other in many benchmarks, will be my guess. Which processor will use most power? Does Alder Lake support AVX512? will Raptor Lake support AVX512? who actually needs AVX512?

I think it is quite confusing when we start "trash talking" without being specific about why it is a let down. I don't do anything that requires more than 8 cores, and I'm pretty sure I don' need AVX512 for anything I do. I'm also pretty sure that will be the case for 95% of users. 15% is not a gigantic leap forward, but it's still better than what intel gave us with each generation from 5xxx until 11xxx.

To me it was kind of expected as one of the leaks already has stated that an AMD engineer was more excited about zen5 than zen4. So to me it is a solid upgrade, but nothing spectacular, but neither a letdown.

 

[yuri69]

• Raptor Cove brings up to 10-15% more performance over Golden
• Zen 4 bring more than 15% more performance over Zen 3

Choose your underwhelming upgrade now!

Zen 4 looks underwhelming in light of the current state of released info.

* 2 years vs 1 year after its predecessor
* presented as the big upgrade (Zen 3, Zen 3X, Zen 4) vs a compatible refresh sharing the same socket, etc.
* breaking the lineage of consecutive 15+ IPC gains vs "Tick" - "Tock" IPC (Ice Lake - Tiger Lake; Alder Lake - Raptor Lake)
* not being able to achieve the top ST perf vs keeping the top ST perf

 

[Racan]

To me it was kind of expected as one of the leaks already has stated that an AMD engineer was more excited about zen5 than zen4. So to me it is a solid upgrade, but nothing spectacular, but neither a letdown.

Talking about Mike Clark no? I remember, plus you have the lack of major changes to the core so I was kind of expecting modest IPC improvements, but still it’s a bit underwhelming. I hope he’s right about Zen 5.

 

[Panino Manino]

Lisa "demonstrating" the clock speeds with a video of a game running remembers me of the Vega days...

 

[coercitiv]

* 2 years vs 1 year after its predecessor

There's a difference between being underwhelming for us on the forum and not being able to compete. Customers don't care about the new generation being a tick or a tock of sorts, and I was addressing a comment that presented Raptor Cove as THE big problem for Zen4.

 

[inf64]

.

Talking about Mike Clark no? I remember, plus you have the lack of major changes to the core so I was kind of expecting modest IPC improvements, but still it’s a bit underwhelming. I hope he’s right about Zen 5.

Where did AMD make any claims about uarch. changes in Zen4? All we have is the vague GB leak, nothing from AMD.
For Zen3 we had even less leaks and most were expecting ~10% IPC from Zen2 and clock stagnation. We got ~19% IPC and much higher boost for ST/all core, all on the same node.

 

[Panino Manino]

They do say "objects in mirror are closer than they appear".

https://www.techspot.com/news/93026-intel-boss-gelsinger-amd-rearview-mirror-following-alder.html

The competitor slipped at the corner.

 

[biostud]

There's a difference between being underwhelming for us on the forum and not being able to compete. Customers don't care about the new generation being a tick or a tock of sorts, and I was addressing a comment that presented Raptor Cove as THE big problem for Zen4.

Depends on when a vcache zen4 cpu is released.

 

[FangBLade]

I think they focused too much on the platform itself, and they will catch up Intel eventually with future generations, just like with Zen 1 and AM4, but still way too low.. 15% performance just sux for new generation on new platform.

 

[Abwx]

I think they focused too much on the platform itself, and they will catch up Intel eventually with future generations, just like with Zen 1 and AM4, but still way too low.. 15% performance just sux for new generation on new platform.

Are we living in an era of single core CPUs...?

According to Hans de Vries in CB R23 MT and comparatively to Zen 3 :

38% better perf for the 6C
33% for the 8C
32% for the 12C
45% for the 16C.

You call this 15% perf uplift.?.

What is sure is that 8 extra e cores wont cut the mustard to compete with this.

 

[uzzi38]

Paul Alcorn on Twitter: "AMD has confirmed that the 170W figure for AM5 is PPT, not TDP." / Twitter

And that's the power draw in that Blender render now confirmed. 170W, not the 230W we first assumed.

 

[maddie]

The competitor slipped at the corner.

Nothing is biting, try new bait.

 

[tamz_msc]

Are we living in an era of single core CPUs...?

According Hans de Vries in CB R23 MT and comparatively to Zen 3 :

38% better perf for the 6C
33% for the 8C
32% for the 12C
45% for the 16C.

You call this 15% perf uplift.?.

What is sure is that 8 extra e cores wont cut the mustard to compete with this.

That calculation is off because higher core SKUs have always given lesser gen-on-gen improvements in purely multi-core benchmarks. The gains will more be in line with the Blender result that AMD demonstrated, i.e. around ~30%.

8-ecores give ~8000 CB R23 points



Without power limits in place Raptor Lake 8+16 will easily manage >36000 CBR23 points - thats 20000 from P-cores and 16000 from E-cores. 5950X manages 28000-29000 with PBO.That's already putting Raptor Lake potentially ~30% faster than the 5950X with PBO. So a 7950X will be dead even in CB R23.

You can save this post and revisit it again when the products actually launch.

 

[maddie]

Are we living in an era of single core CPUs...?

According to Hans de Vries in CB R23 MT and comparatively to Zen 3 :

38% better perf for the 6C
33% for the 8C
32% for the 12C
45% for the 16C.

You call this 15% perf uplift.?.

What is sure is that 8 extra e cores wont cut the mustard to compete with this.

Crazy thought. If single thread is 15% AND multi thread is ~33%, then what does that say about multithreading (SMT) efficiency.

 

[Abwx]

That calculation is off because higher core SKUs have always given lesser gen-on-gen improvements in purely multi-core benchmarks. The gains will more be in line with the Blender result that AMD demonstrated, i.e. around ~30%.

8-ecores give ~8000 CB R23 points

View attachment 61972

Without power limits in place Raptor Lake 8+16 will easily manage >36000 CBR23 points - thats 20000 from P-cores and 16000 from E-cores. 5950X manages 28000-29000 with PBO.That's already putting Raptor Lake potentially ~30% faster than the 5950X with PBO. So a 7950X will be dead even in CB R23.

You can save this post and revisit it again when the products actually launch.

In Blender the 7950X is 45% faster than the 12900K, get your maths right before stating that 30% non sense, as already said it s 31% less time to render a scene...

Even in Cinebench the 7950X is still 40% faster than the 12900K in CB R23, so yes, dont forget to remind me your post once parts are benched...

Crazy thought. If single thread is 15% AND multi thread is ~33%, then what does that say about multithreading (SMT) efficiency.

Dunno exactly from where things come, for the 7950X they can clock it 20-25% higher than the 5950X, but that doesnt work for the 7800X because the 5800X is already at about 4.5GHz all cores, that would require 5.5GHz all cores for its replacement.

 

[gdansk]

Crazy thought. If single thread is 15% AND multi thread is ~33%, then what does that say about multithreading (SMT) efficiency.

It says almost nothing without knowing the delta between sustained clock rates in 1T and MT gen over gen.

 

[maddie]

In Blender the 7950X is 45% faster than the 12900K, get your maths right before stating that 30% non sense, as already said it s 31% less time to render a scene...

Even in Cinebench the 7950X is still 40% faster than the 12900K in CB R23, so yes, dont forget to remind me your post once parts are benched...



Dunno exactly from where things come, for the 7950X they can clock it 20-25% higher than the 5950X, but that doesnt work for the 7800X because the 5800X is already at about 4.5GHz all cores, that would require 5.5GHz all cores for its replacement.

It says almost nothing without knowing the delta between sustained clock rates in 1T and MT gen over gen.

If the single thread is up 15% and the multi-thread is up 33%, then SMT (efficiency/performance enhancement) has increased by 16% for a given clock.

Are we thinking that single thread clocks are lower than multi thread clocks. That is the only scenario where SMT uplift is higher than single thread.

 

[nicalandia]

Amd knows that 16C/32T is enough beat Raptor Lake 13900K in ST and MT. Just check the CB and Blender Scores

 

[AtenRa]

That calculation is off because higher core SKUs have always given lesser gen-on-gen improvements in purely multi-core benchmarks. The gains will more be in line with the Blender result that AMD demonstrated, i.e. around ~30%.

8-ecores give ~8000 CB R23 points

View attachment 61972

Without power limits in place Raptor Lake 8+16 will easily manage >36000 CBR23 points - thats 20000 from P-cores and 16000 from E-cores. 5950X manages 28000-29000 with PBO.That's already putting Raptor Lake potentially ~30% faster than the 5950X with PBO. So a 7950X will be dead even in CB R23.

You can save this post and revisit it again when the products actually launch.

Dont confuse 31% less time in blender with CB R23 MT score.

The first one is time and the second is performance metric. Let me give you an example.

1 sec vs 2 sec
1 sec is half the time of 2 secs

Half the time (50% less time) doesnt mean it is 50% faster performance. If you finish in half the time (50% less time) then you are 2x times or 100% faster.


What I want to say is that until we see actual CB R23 scores its not correct to translate the 31% less time we show in Blender in to CB MT Scores.

 

[gdansk]

If the single thread is up 15% and the multi-thread is up 33%, then SMT (efficiency/performance enhancement) has increased by 16% for a given clock.

Are we thinking that single thread clocks are lower than multi thread clocks. That is the only scenario where SMT uplift is higher than single thread.

What? They're relative to Zen 3. So if 1T clock when from Z3 5050MHz to Z4 5520MHz and multicore went from Z3 3800MHz to Z4 4800MHz you could end up in that scenario.

I don't know where Hans got the 6/8C stuff but the 16 core parts would benefit greatly from higher PPT. Sustained MT clock rates will increase more relative to 1T clock rate increase.

 

[inf64]

If the single thread is up 15% and the multi-thread is up 33%, then SMT (efficiency/performance enhancement) has increased by 16% for a given clock.

Are we thinking that single thread clocks are lower than multi thread clocks. That is the only scenario where SMT uplift is higher than single thread.

There are too many unknows. For example what we know is that 5950X sustains around 4.2-4.35Ghz in workloads like R23 or blender with PBO enabled. If Zen4 sustains 5Ghz across all cores in R23 and scores ~38K (~33% higher), that would put the IPC increase in R23 to around ~14%ish or thereabout. This is interestingly similar to Zen2->Zen3 IPC jump in the same workload.

edit: If the all-core boost for Zen4 is higher (for example 5.2Ghz), it would mean lower IPC increase of around 10-11% in that specific test.

 

[gdansk]

There are too many unknows. For example what we know is that 5950X sustains around 4.2-4.35Ghz in workloads like R23 or blender with PBO enabled. If Zen4 sustains 5Ghz across all cores in R23 all cores and scores ~38K (~33% higher), that would put the IPC increase in R23 to around ~14%ish or thereabout. This is interestingly similar to Zen2->Zen3 IPC jump in the same workload.

Why would compare to PBO enabled?

 

[inf64]

Why would compare to PBO enabled?

That's the out of the box performance. For example the CPU monkey has all the Zen2/3 scores with PBO enabled.

 

[tamz_msc]

In Blender the 7950X is 45% faster than the 12900K, get your maths right before stating that 30% non sense, as already said it s 31% less time to render a scene...

Even in Cinebench the 7950X is still 40% faster than the 12900K in CB R23, so yes, dont forget to remind me your post once parts are benched...

40-45% gain over previous gen 16-core part is simply unachievable, even with increase in clock speeds. Go look at each 6/8/12/16 Core SKU from Zen 2 -> Zen 3.

 

[AtenRa]

If the single thread is up 15% and the multi-thread is up 33%, then SMT (efficiency/performance enhancement) has increased by 16% for a given clock.

Are we thinking that single thread clocks are lower than multi thread clocks. That is the only scenario where SMT uplift is higher than single thread.

There are too many unknows. For example what we know is that 5950X sustains around 4.2-4.35Ghz in workloads like R23 or blender with PBO enabled. If Zen4 sustains 5Ghz across all cores in R23 and scores ~38K (~33% higher), that would put the IPC increase in R23 to around ~14%ish or thereabout. This is interestingly similar to Zen2->Zen3 IPC jump in the same workload.

edit: If the all-core boost for Zen4 is higher (for example 5.2Ghz), it would mean lower IPC increase of around 10-11% in that specific test.

The higher MT score can be a combination of the following ,

ZEN 4 could sustain way higher all core turbo than we think (5+ GHz).
SMT performance increase and
PPT has increased to 170W vs 142W for ZEN 3 (this one related to the first)

So even if IPC remains low, your MT performance can increase substantially