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!

 

[Markfw]

They said it is 230W socket. Do you think they're going to leave available performance when they will be so close to Raptor Lake?
We'll see but it seems like the 35% target must be at least 5GHz based on their IPC estimation. Maybe a bit more. How much power does say 5100MHz x 16 need? That's only 14.375W per core. 1 Zen 3 core at 5050MHz draws over 20W.

I will book mark this for reference when reviews come out. I will eat my words if it takes 230. Remember, the socket is also rumored to support more cores in the future. They are going to use it all on a 16 core ???? I don't think so.

 

[gdansk]

I will book mark this for reference when reviews come out. I will eat my words if it takes 230. Remember, the socket is also rumored to support more cores in the future. They are going to use it all on a 16 core ???? I don't think so.

Mark, they used 105W*1.35 even on the 8 core 5800X...

https://images.anandtech.com/doci/16214/PerCore-3-5800X_575px.png

When 24/32 core parts come they'll look more efficient like the 5950X did compared to the 5800X.

 

[Panino Manino]

See, all those "leaks" were exaggerated, and again, leakers creating unrealistic expectations resulting in unnecessary disappointment. Sometimes it feels like an on purpose psyops.

For me my biggest disappointment was confirming that AMD is abandoning the Painters codenames.

 

[CakeMonster]

24/32c is probably planned for 4nm/3nm.

 

[Hitman928]

Mark, they used 105W*1.35 even on the 8 core 5800X...

https://images.anandtech.com/doci/16214/PerCore-3-5800X_575px.png

When 24/32 core parts come they'll look more efficient like the 5950X did compared to the 5800X.

I am pretty doubtful that they used ~62% more power to get a 35% higher score in Cinebench. If this is the case, then something went really wrong with Zen4.

A 5950x will only boost to ~3.9 GHz in Cinebench r23 at stock. So to get a greater than 35% increase in performance at 8% improved IPC, they would need to hit about 4.8 GHz all core and do it within ~155 W to match the 25% perf/watt claim. I don't think that's too unreasonable given that they've shown multiple cores running 5.4 GHz - 5.5 GHz sustained. We'll see what the actual numbers are soon enough but I don't think getting 35% all core performance improvement with ~10% power increase is unreasonable.

 

[gdansk]

I am pretty doubtful that they used ~62% more power to get a 35% higher score in Cinebench. If this is the case, then something went really wrong with Zen4.

A 5950x will only boost to ~3.9 GHz in Cinebench r23 at stock. So to get a greater than 35% increase in performance at 8% improved IPC, they would need to hit about 4.8 GHz all core and do it within ~155 W to match the 25% perf/watt claim. I don't think that's too unreasonable given that they've shown multiple cores running 5.4 GHz - 5.5 GHz sustained. We'll see what the actual numbers are soon enough but I don't think getting 35% all core performance improvement with ~10% power increase is unreasonable.

It is simple consequence of pushing all core clock much higher. They could have designed braniac but it is risky.

It is still only 12.5W per core excluding other package power. To reach about 5GHz at 12.5W per core would be pretty good and greatly more efficient than Zen 3 (about 20W). And Golden Cove is what 25-30W per core at that clock rate? In context, quite the achievement.

 

[majord]

I don't really see the 8-10% IPC uplift as disappointing at all. They've had a good run extracting well above avg IPC and frequency uplifts from zen 2 to 3, at a very impressive rate, and it seems the later part hasn't hit a wall at all. , all without drastic changes to the core (which has no doubt allowed this 18 mth lead time between them) . A poor Zen 5 showing I'd be disappointed / concerned about far more, as it is now apparent this is the big uArch overhaul. but an easing off of the uplifts towards the end of the incremental upgrades to the current core isn't unexpected. Is usually better than pushing it too far and creating perf/watt issues and inconsistent performance uplifts due to diminishing returns / internal bottlenecks .

Frequency scaling is impressive , particuarly in MT scenarios. There's a chance HT yield has gone up also . The resulting claimed 35%+ MT uplift from the same core count is nothing to be sneezed at, and far more than predecessors . Yes it comes with higher power consumption , but their slide is claiming 25% perf/watt uplift , so there's some mystery surrounding that we'll have to wait to solve before commenting much futher. In any case, they're coming from such a strong perff/watt advantage already

No it's not going to take any ST trophy vs Raptor lake - probably struggle to get its head above Adler lake , but the doom and gloom around that one metric does crack me up a bit. . What is the logic here? .. RTL having a single digit lead on the cinebench ST performance page of a review site = total annihilation ?

Also , why are we bringing MTL into this ? is this now launching on top of RTL all of a sudden? did I miss something , or does it still not actually have a firm launch window yet beyond some time out the back end of 2023 at best ?

 

[gdansk]

Also , why are we bringing MTL into this ? is this now launching on top of RTL all of a sudden? did I miss something , or does it still not actually have a firm launch window yet beyond some time out the back end of 2023 at best ?

Because Zen 5 is confirmed as 2024. That means Zen 4 variants will be the MTL competitor unless Intel fumbles it. Last I heard 2023H2 but that may have been an unofficial source.
Some people think Zen 4 will have a hard time. But given the possible configurations of Zen 4/4D/4C I suspect AMD can put something together that's competitive and fits in AM5. If they want to.

 

[Henry swagger]

7600x and 7800x will be even lower.. in that segment amd is in trouble.. but dragon range should compete well with raptor lake h

 

[HurleyBird]

Because Zen 5 is confirmed as 2024. That means Zen 4 variants will be the MTL competitor unless Intel fumbles it. Last I heard 2023H2 but that may have been an unofficial source.

For every 1 launch date Intel makes, they fumble 10.

 

[itsmydamnation]

7600x and 7800x will be even lower.. in that segment amd is in trouble.. but dragon range should compete well with raptor lake h

All AMD needs to do is 6/8 core + vcache and then push everyone into doing FCAT based reviews.
Watch every slight miss direction from thread director cause minimum FPS hicups.

 

[gdansk]

All AMD needs to do is 6/8 core + vcache and then push everyone into doing FCAT based reviews.
Watch every slight miss direction from thread director cause minimum FPS hicups.

Weird, why does it say I said that? I didn't.

 

[lobz]

AMD's problem is their Zen 3 core is going to be 2 years old when its successor finally hits. In the meantime Intel pushed RKL, ADL, and maybe even RPL. So AMD's designs simply have to be pretty strong and age well since their lifecycle is pretty long.

you can't be serious with Rocket Lake there though... how's that 'pushed' when it was the very first 10nm (mainstream desktop) CPU, launching what, 2 whole or even 3 years too late?

 

[biostud]

So if 4c is a 16core chiplet without L3 cache, then they can still add vcache. And in 2023 the can have a 32 core CPU on the desktop. No wonder they're discontinuing the regular threadripper.

 

[lobz]

8-10% IPC increase, 35% increase in cb. Zen 5 2024.

Overall disappointing news

You're absolutely right, even though you didn't intend to be so. All this is extremely disappointing for Intel, especially where it hurs Intel's bottom line the most. But you'll see over the course of the next 18 months 🙂

 

[DisEnchantment]

Oh well, it seems it is not just FP but int PPC improvements are pretty mediocre.
~2 years cadence, node advantage, tons of Xtor and this is all they manage.

I am curious whose team accomplished this, Leslie Barnes or David Suggs?
If not for N5 perf improvements it would have been pretty lackluster.
Had they been process limited would have been worse than Intel.

But Phoenix and MI300 seems impressive though, aside of CPU. And probably STX lands in 2024 CES as usual.

SkyJuice seems on point, the rest are all click baiters.

 

[yuri69]

David Suggs

Is chief architect of Zen 5.

 

[itsmydamnation]

Weird, why does it say I said that? I didn't.

I have no idea.... let me fix it ....lol

 

[JasonLD]

Now I wish AMD does it with RDNA3 or future architecture and make 300W+ Big APU. It will be a great high-end system with small footprint like M1 Ultra system.

 

[Henry swagger]

View attachment 62868

Now I wish AMD does it with RDNA3 or future architecture and make 300W+ Big APU. It will be a great high-end system with small footprint like M1 Ultra system.

We need this in a mini pc form factor

 

[Frenetic Pony]

View attachment 62868

Now I wish AMD does it with RDNA3 or future architecture and make 300W+ Big APU. It will be a great high-end system with small footprint like M1 Ultra system.

I don't have any doubt we'll get unified memory sooner or later out of AMD. It's not like they don't have extensive commercial experience producing exactly that with consoles.

It's also interesting to note that both Zen 4c and Zen laptop are seemingly 4nm. It isn't explicit for 4c but piecing their slides together, well 5nm explicity lists Zen 4 normal/v-cache/threadripper. While Zen APU and 4c are listed under "5/4nm". I wonder if they use the same core?

 

[biostud]

We need this in a mini pc form factor

Xbox?

 

[biostud]

View attachment 62867

Oh well, it seems it is not just FP but int PPC improvements are pretty mediocre.
~2 years cadence, node advantage, tons of Xtor and this is all they manage.

I am curious whose team accomplished this, Leslie Barnes or David Suggs?
If not for N5 perf improvements it would have been pretty lackluster.
Had they been process limited would have been worse than Intel.

But Phoenix and MI300 seems impressive though, aside of CPU. And probably STX lands in 2024 CES as usual.

SkyJuice seems on point, the rest are all click baiters.

I think zen4 is a two-step launch, that isn't focused on IPC:

2022: New platform and new processors. The processors seems like "tick" shrinking of zen3 with some light chances with AVX512 and L2 cache, and we might even get vcache models at launch or maybe early 2023
2023: They launch zen4c on 4N with up to 32 cores on desktop, also with vcache models, and I think this is the main reason for the 170W TDP
2024: "tock" zen5

 

[cortexa99]

Won't comment on performance number before retail version launch. The most great news to me is AMD won't implement big+little architecture on Zen5 like "Zen5+Zen4c" which are fom those "leakers" nonsense.
Future? Who knows. I hope AMD won't do that before Thread Detector being mature enough.

 

[Abwx]

That is wrong. Clock for clock on the same number of cores 5950x vs 7950x would yield a 25% improvement in power consumption for Zen 4 in theory.

Process is 2x more efficient, so perf/watt improvement at same clock should be close to 100% if we discard IPC difference.

At same power, and according to AMD numbers for process perfs, Zen 4 can be clocked 20% higher than Zen 3, the number for average frequencies is 25% higher clocks.