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]

Intel's 7nm CPU's Meteor Lake (Intel 4) will be out by the end of 2023. Intel has already said they are accelerating their processes and designs because they fell behind AMD. If Zen 4 is only 30-60 days ahead of Raptor Lake. That means a lot especially if Raptor Lake outperforms or comes very close to Zen 4 performance. Remember, Zen 4 is on a silicon die shrink this round. Intel goes down from 10nm to 7nm on the next generation. So Intel gets a huge benefit in fabrication next year.

So it's possible Intel can leave AMD in the rear view mirror or smash them like a bug. Which is why I said over a month ago that Zen 4 needed to be out on the market yesterday.

And we are totally sure Intel will have no problem launching Intel 4 on time, with the desired characteristics and volume?
It will be a pleasant surprise to see Intel back on track with tick tock, but given their track record the last decade, consider me sceptical.

 

[Hans Gruber]

I'm not sure you are serious or joking. If serious you lost grip on reality. Intel is not competing in a vacuum, and AMD is not standing still. If Zen5 launches on time, intel better hope they will somehow achieve parity as they will be going against a massive 256 (fat/wider) Zen5 cores in 2024.

It is you who does not understand the gravity of the situation. Alder lake was the 1st die shrink for Intel in over 6 years from 14nm to 10nm. After Raptor Lake 10nm they go down to 7nm. 10nm was always a problem for Intel and they have no indicated that 7nm would provide them difficulty. Zen 5 is supposed to be on 3nm and TSMC said that was delayed slightly already.

Few want to admit that Zen 3 is no longer the performance king. Alder Lake is better and it's the first 10nm Intel processor. To Intel's credit. They bring the wrath and hatred of computer builders around the world with new sockets every other generation. So Raptor Lake will be one and done. Meteor Lake uses a different socket and needs a new motherboard with a new chipset come late 2023 or early 2024.

I seem to be the only one here running a B350 motherboard with Cas 16 memory @ 3800mhz. It simply required a bios fix. Better late (4 years) than never.

It's not just on this forum. People get salty on both the intel side and the AMD side of things all over the internet.

 

[inf64]

It is you who does not understand the gravity of the situation. Alder lake was the 1st die shrink for Intel in over 6 years from 14nm to 10nm. After Raptor Lake 10nm they go down to 7nm. 10nm was always a problem for Intel and they have no indicated that 7nm would provide them difficulty. Zen 5 is supposed to be on 3nm and TSMC said that was delayed slightly already.

Few want to admit that Zen 3 is no longer the performance king. Alder Lake is better and it's the first 10nm Intel processor. To Intel's credit. They bring the wrath and hatred of computer builders around the world with new sockets every other generation. So Raptor Lake will be one and done. Meteor Lake uses a different socket and needs a new motherboard with a new chipset come late 2023 or early 2024.

I seem to be the only one here running a B350 motherboard with Cas 16 memory @ 3800mhz. It simply required a bios fix. Better late (4 years) than never.

It's not just on this forum. People get salty on both the intel side and the AMD side of things all over the internet.

I'm now sure you clearly don't know that intel is far behind AMD in everything but ST performance, on desktop. And the gap there will become zero in the next 4 months, while the gap in everything else will become greater.

In server space, they are really behind, and this gap will become much much greater with Zen4. Think 96-128 Zen4 cores versus 64? AlderLake cores, it will be brutal. In laptop segment, Phoenix/ Dragon Range will have a significant performance and perf./watt lead, by how much it's unknown (but substantial). Intel needs to basically double performance each year to stay competitive in 2024.

 

[itsmydamnation]

Wtf are you talking about

What are cannon and ice lake

Intel P cores are a hot mess , a14-15 sized resources and need to clock to the moon to match performance

 

[rainy]

In server space, they are really behind, and this gap will become much much greater with Zen4. Think 96-128 Zen4 cores versus 64? AlderLake cores, it will be brutal.

To be correct 56 cores because I'm sceptical we will ever see 64 cores Sapphire Rapids.

Btw, there's space between Zen and generation number exactly the same as it was with Intel's Pentium (Pentium 4 for example).

 

[Karnak]

Few want to admit that Zen 3 is no longer the performance king.

You haven't heard of Milan and/or Milan-X, have you?

I don't get the "only desktop exists and nothing else" attitude. That's just pointless. And even if you're talking about desktop only it's.... wrong. ADL is the ST "king" in Applications. For all other things it's a tie between both (gaming performance) or a clear loss for ADL (efficiency and/or MT performance).

And regarding the server segment everything Intel has to offer is just not competitive. And that won't change with SPR vs. Genoa. That'll be an even bigger loss for Intel right there.

 

[Abwx]

Few want to admit that Zen 3 is no longer the performance king. Alder Lake is better and it's the first 10nm Intel processor

Better at what exactly set apart for draining 240W to not even match a 5950X@130W..?..

240W in MT :

Intel Core i9-12900K, i7-12700K & i5-12600K im Test: Die neue TDP-Definition und ihre Auswirkungen

Intel Alder Lake im Test: Die neue TDP-Definition und ihre Auswirkungen / Intels „TDP-Nachfolger“ für die K-CPUs

www.computerbase.de

And the average of MT perf, 5950X is 9% faster despite running at half your wonder s TDP :

Intel Core i9-12900K, i7-12700K & i5-12600K im Test: Benchmarks in Anwendungen

Intel Alder Lake im Test: Benchmarks in Anwendungen / Ein rekordverdächtiger Testumfang / Windows 10 vs. Windows 11

www.computerbase.de

 

[inf64]

Btw, there's space between Zen and generation number exactly the same as it was with Intel's Pentium (Pentium 4 for example).

We all know Zen4 = Zen 4. I refuse to add a useless space when we all know what we're talking about.

 

[DisEnchantment]

This APU is one engineering feat from AMD



Looks like CPU and GPU chiplets sitting on top of the Infinity caches
If they can do this in 2023, then 2024 it will probably be L3 on the bottom like what @Hans de Vries is suggesting
I think this is how Turin can hit 256 cores on SP5.
Interesting times, AMD is definitely pioneering chiplet and stacking tech.

I am wondering what if the MPDMA can migrate data from HBM to the LLC.
Not sure if they have DDR5 controllers, but 256 or even 512 GB of HBM would be likely if no DDR5 support.
I wonder if it carries CXL support too, and if the MPDMA can migrate pages around in this scenario too.
Prices of these things would be sky high.

 

[randomhero]

It doesn't need to be a real address to get to the video link, and all of the slides are available without registering on the webcast page.

In the Q&A, Lisa Su stated that Zen 5 will have both 4nm and 3nm CCD designs. The slides seem to show that the 3D and c versions are out in 2024 which seems to imply that Zen 5 will be out in less than two years. Mike Clark alluded to the design being finished in the January interview here so it sounds plausible.

Cheers, will look when I get some time!

 

[Markfw]

It is you who does not understand the gravity of the situation. Alder lake was the 1st die shrink for Intel in over 6 years from 14nm to 10nm. After Raptor Lake 10nm they go down to 7nm. 10nm was always a problem for Intel and they have no indicated that 7nm would provide them difficulty. Zen 5 is supposed to be on 3nm and TSMC said that was delayed slightly already.

Few want to admit that Zen 3 is no longer the performance king. Alder Lake is better and it's the first 10nm Intel processor. To Intel's credit. They bring the wrath and hatred of computer builders around the world with new sockets every other generation. So Raptor Lake will be one and done. Meteor Lake uses a different socket and needs a new motherboard with a new chipset come late 2023 or early 2024.

I seem to be the only one here running a B350 motherboard with Cas 16 memory @ 3800mhz. It simply required a bios fix. Better late (4 years) than never.

It's not just on this forum. People get salty on both the intel side and the AMD side of things all over the internet.

You seem to forget 4 things about Alder Lake:
1) They are only leading in Single thread
2) They are only leading in Desktop
3) They suck power like no tomorrow.
4) Zen 4 is just around the corner and will most likely take over single thread and maintain power/perf.

I have a 12700F, and 6 5950x's. The 12700F sucks the power of 2 5950x's and get creamed in total performance 4x by 2 5950x's, or 2x one on one.
When your electric bill is $800 a month, electricity use really makes a difference. I don't even have the wiring to support 6 12900k's. And in the server world, its not just me, but MANY companies are switching to Milan due to performance and power/perf.

YOU are the delusional one.

 

[randomhero]

This APU is one engineering feat from AMD

View attachment 62987

Looks like CPU and GPU chiplets sitting on top of the Infinity caches
If they can do this in 2023, then 2024 it will probably be L3 on the bottom like what @Hans de Vries is suggesting
I think this is how Turin can hit 256 cores on SP5.
Interesting times, AMD is definitely pioneering chiplet and stacking tech.

I am wondering what if the MPDMA can migrate data from HBM to the LLC.
Not sure if they have DDR5 controllers, but 256 or even 512 GB of HBM would be likely if no DDR5 support.
I wonder if it carries CXL support too, and if the MPDMA can migrate pages around in this scenario too.
Prices of these things would be sky high.

Great post mate!

See, I've watched a couple of videos from we'll known tech outlets, none of them mentioned anything like this. They were basically just reading out loud stuff that was written on slides.
Like said in post before, I will catch up with new info when I get some time.

AMD always puts load of info on FADs, just have to listen and watch carefully.

 

[Schmide]

It is you who does not understand the gravity of the situation. Alder lake was the 1st die shrink for Intel in over 6 years from 14nm to 10nm. After Raptor Lake 10nm they go down to 7nm. 10nm was always a problem for Intel and they have no indicated that 7nm would provide them difficulty. Zen 5 is supposed to be on 3nm and TSMC said that was delayed slightly already.

Few want to admit that Zen 3 is no longer the performance king. Alder Lake is better and it's the first 10nm Intel processor. To Intel's credit. They bring the wrath and hatred of computer builders around the world with new sockets every other generation. So Raptor Lake will be one and done. Meteor Lake uses a different socket and needs a new motherboard with a new chipset come late 2023 or early 2024.

I seem to be the only one here running a B350 motherboard with Cas 16 memory @ 3800mhz. It simply required a bios fix. Better late (4 years) than never.

It's not just on this forum. People get salty on both the intel side and the AMD side of things all over the internet.

This post reads like a gambling addict spouting off, if they can just get one win the troubles of the past will fade away.

AMD is fine from behind, the fact you elude to that you can run a modern AMD processor on your 5 year old B350, is the reason most people don't see the crisis you proselytize.

and seriously it's not "to intel's credit" it's "to intel's detriment"

 

[gdansk]

And Zen 4 isn't necessarily unable to compete with MTL either. They have Zen 4C to add more MT. Zen 4D to add more gaming performance. They might be able to get even higher clocks when taken to N4. Maybe they have different capacitor layouts to go 3 CCD. Who knows.

It's not static. AMD can do things too and showed they have a lot of options for Zen 4.

 

[lobz]

Intel's 7nm CPU's Meteor Lake (Intel 4) will be out by the end of 2023. Intel has already said they are accelerating their processes and designs because they fell behind AMD. If Zen 4 is only 30-60 days ahead of Raptor Lake. That means a lot especially if Raptor Lake outperforms or comes very close to Zen 4 performance. Remember, Zen 4 is on a silicon die shrink this round. Intel goes down from 10nm to 7nm on the next generation. So Intel gets a huge benefit in fabrication next year.

So it's possible Intel can leave AMD in the rear view mirror or smash them like a bug. Which is why I said over a month ago that Zen 4 needed to be out on the market yesterday.

That's right, Intel is constantly accelerating their processes since 2017. Unfortunately, they forgot they were still in reverse.

 

[lobz]

It is you who does not understand the gravity of the situation. Alder lake was the 1st die shrink for Intel in over 6 years from 14nm to 10nm. After Raptor Lake 10nm they go down to 7nm. 10nm was always a problem for Intel and they have no indicated that 7nm would provide them difficulty. Zen 5 is supposed to be on 3nm and TSMC said that was delayed slightly already.

Few want to admit that Zen 3 is no longer the performance king. Alder Lake is better and it's the first 10nm Intel processor. To Intel's credit. They bring the wrath and hatred of computer builders around the world with new sockets every other generation. So Raptor Lake will be one and done. Meteor Lake uses a different socket and needs a new motherboard with a new chipset come late 2023 or early 2024.

I seem to be the only one here running a B350 motherboard with Cas 16 memory @ 3800mhz. It simply required a bios fix. Better late (4 years) than never.

It's not just on this forum. People get salty on both the intel side and the AMD side of things all over the internet.

What? They have not indicated any difficulty with their 7nm Intel (4), true, because they did even worse. They simply didn't deliver, and they've delayed it at least 6 times already, but probably even more. They made a complete fool of themselves even with a supercomputer that could possibly be done on risk production, instead the customer had to purchase AMD hardware to get the programming going before Intel can ship them something usable.

Doesn't matter which forums you're on, you have been sleeping everywhere.

 

[gruffi]

It is you who does not understand the gravity of the situation. Alder lake was the 1st die shrink for Intel in over 6 years from 14nm to 10nm. After Raptor Lake 10nm they go down to 7nm. 10nm was always a problem for Intel and they have no indicated that 7nm would provide them difficulty.

You hear problems about Intel's 7nm process since 2020. Maybe it's not as dramatic as 10nm was. But it's still questionable if they can keep their schedule with MTL. Intel was always good at announcing. But especially in recent years their execution was miserable. Their current neverending story of Arc and SPR are just two of many examples.

Zen 5 is supposed to be on 3nm and TSMC said that was delayed slightly already.

Zen 5 is supposed to be 4nm and 3nm. I think 3nm capacity would have been a problem for AMD in 2023. In 2024 the situation should get better.

Few want to admit that Zen 3 is no longer the performance king. Alder Lake is better and it's the first 10nm Intel processor.

Few want to admit that the one year older Zen 3 is still the maximum performance king. And it does so much more efficiently than Alder Lake. There is no better at the moment. Zen 3 has its strengths and weaknesses. Which applies to Alder Lake as well. Alder Lake is a step forward for Intel. Which in fact wasn't that hard after the Rocket Lake fail. But compared to the competition Alder Lake doesn't look very impressive. It still suffers the brute force paradigm of its predecessors. Which always had been a sign of struggling in competing for the technology leadership. The p-core is still very power and area inefficient compared to Zen 3. Intel tried to compensate that with a hybrid design and "efficiency cores". Which works only half. Gracemont is more area efficient than Golden Cove, but it's less power efficient at ~4 GHz. Intel's hybrid design causes only problems for customers. Like incompatible ISA, which resulted in the removal of AVX512 from the p-core. Or scheduling problems, which caused people to disable the e-cores. Customers absolutely don't benefit from Intel's hybrid design. I think a 12 p-core only Alder Lake design would have been better. It's only Intel that benefit from the hybrid design by reducing cost. Because 4 p-cores need more area than 8 e-cores. We can only hope that AMD's hybrid design isn't affected by the same problems. But because they use a different approach I'm quite confident about that.

 

[Vattila]

This [MI300 server] APU is one engineering feat from AMD

The reveal of the Instinct MI300 server APU was certainly the star of the Financial Analyst Day as far as immediate roadmap goes. David Wang, senior vice president of engineering for the Radeon Technologies Group at AMD, was positively buzzing as he revealed the CDNA 3 Unified Memory APU Architecture behind the chip, although he left the reveal of the chip itself to Forrest Norrod, senior vice president and general manager of the Data Center Solutions Group. Interestingly, after Raja Koduri burned out and left for Intel in late 2017, David Wang rejoined AMD in early 2018, apparently to take on a three-generation roadmap to GPU leadership, put in place by AMD CEO Lisa Su to mimic their success with the "Zen" CPU roadmap. It seems he has done well!

Presumably, as other sources have hinted, Instinct MI300 will be used in the 2+ exaflop El Capitan supercomputer set to arrive next year. It is great to see the Exascale Heterogenous Processor finally coming to fruition after a decade of research and development. I expect we will hear more about MI300 in additional supercomputer wins soon.

With this manifestation of "Zen 4" in high-end server APU form, all the potential features I listed in this thread's associated poll have become a reality, with the exception of 4-way SMT. Only 16% of the voters expected integrated memory on the package, but this feature has now indeed been confirmed, as MI300 comes with a "Unified Memory APU Architecture" with HBM memory integrated in the chip. Obviously, this thing will use the latest advanced packaging, and it will be very interesting to eventually see the details on how it is all put together — especially the Infinity Cache, I/O and interposer/bridges (a fan-out layer with Elevated Fanout Bridges will be used, I guess, similar to MI200).

I presume the Infinity Cache is an L4 cache, or perhaps a System Level Cache on the memory side of the memory controller (HBM, CXL.memory and CXL.cache protocols, but no DDR support, maybe), sitting below the GPU and CPU chiplets, and that the CPU chiplets will be ordinary "Zen 4" CCDs with the ability to stack L3 V-Cache chiplets on top. It will also be interesting to see whether AMD will be able to stack V-Cache higher than one layer in the "Zen 4" generation.

AMD: Combining CDNA 3 and Zen 4 for MI300 Data Center APU in 2023 (anandtech.com)

Spoiler: MI300 Slides

 

[moinmoin]

Interesting times, AMD is definitely pioneering chiplet and stacking tech.

Made even more interesting considering Xilinx is no slouch in that area either. Hope they can combine well there. They downright celebrated the merger in that area at FAD:

(Source: slide 13 Mark Papermaster presentation)

 

[Thibsie]

I presume the Infinity Cache is an L4 cache, or perhaps a System Level Cache on the memory side of the memory controller, sitting below the GPU and CPU chiplets, and that the CPU chiplets will be ordinary "Zen 4" CCDs with the ability to stack L3 V-Cache chiplets on top. It will also be interesting to see whether AMD will be able to stack V-Cache higher than one layer in the "Zen 4" generation.

I'd rather expect the CCD to go over the cache chiplets.

 

[Exist50]

The 12700F sucks the power of 2 5950x's and get creamed in total performance 4x by 2 5950x's, or 2x one on one.

AMD's far from losing anything, but objectively false statements like these are no less ridiculous. If you're going to call out delusional predictions, you should keep your claims less delusional, not more.

 

[Exist50]

The reveal of the Instinct MI300 server APU was certainly the star of the Financial Analyst Day as far as immediate roadmap goes. Presumably, as other sources have hinted, this chip will be used in the 2+ exaflop El Capitan supercomputer set to arrive next year. It is great to see the Exascale Heterogenous Processor finally coming to fruition after a decade of research and development. I expect we will hear more about MI300 in additional supercomputer wins soon.

With this manifestation of "Zen 4" in high-end server APU form, all the potential features I listed in this thread's associated poll have become a reality, with the exception of 4-way SMT. Only 16% of the voters expected integrated memory on the package, but this feature has now indeed been confirmed, as MI300 comes with a "Unified Memory APU Architecture" with HBM memory integrated in the chip. Obviously, this thing will use the latest advanced packaging, and it will be very interesting to eventually see the details on how it is all put together — especially the Infinity Cache, I/O and interposer/bridges (a fan-out layer with Elevated Fanout Bridges will be used, I guess, similar to MI200).

I presume the Infinity Cache is an L4 cache, or perhaps a System Level Cache on the memory side of the memory controller (HBM, CXL.memory and CXL.cache protocols, but no DDR support, maybe), sitting below the GPU and CPU chiplets, and that the CPU chiplets will be ordinary "Zen 4" CCDs with the ability to stack L3 V-Cache chiplets on top. It will also be interesting to see whether AMD will be able to stack V-Cache higher than one layer in the "Zen 4" generation.

AMD: Combining CDNA 3 and Zen 4 for MI300 Data Center APU in 2023 (anandtech.com)

Spoiler: MI300 Slides

I think a very large system/memory-side cache is an interesting possibility. Seems like potentially a better fit given the heterogeneous architecture. Wonder what technology they'd put it on, and whether they'd make it uniform for everything or try to partition it in some way.

 

[Markfw]

AMD's far from losing anything, but objectively false statements like these are no less ridiculous. If you're going to call out delusional predictions, you should keep your claims less delusional, not more.

300 watt vs 142 watt ? Whats delusional there ? I have kill-a-watt evidence (well, I did, but disconnected now)

And total performance is 32 threads@100% vs 20 threads@100%. Yes a 12900k or ks would have 4 more e-cores, but use even more power, and the performance on what I do would not be much more. In my case the e-cores kill the performance.

 

[Abwx]

AMD's far from losing anything, but objectively false statements like these are no less ridiculous. If you're going to call out delusional predictions, you should keep your claims less delusional, not more.

False statement for whom is clueless, the review i linked state that the 5950X perform 30% better than the 12700K despite the latter using 35-40% more power.

Do the maths, the perf/watt ratio between the two CPUs is not far from 100% better for the 5950X.

 

[Exist50]

300 watt vs 142 watt ? Whats delusional there ? I have kill-a-watt evidence (well, I did, but disconnected now)

Lmao, sure you do. Wall measuring the AMD system just happens to exactly equal AMD's PPT, but inexplicably goes 10s of watts over on the Intel chip? Lol.

And total performance is 32 threads@100% vs 20 threads@100%.

No reasonable person would insist that the 5950x is 2x the performance of the 12700f, especially with each at maximum power.

False statement for whom is clueless, the review i linked state that the 5950X perform 30% better than the 12700K despite the latter using 35-40% more power.

The comment I replied to is claiming 2x the performance from the 5950x and >2x the power consumption from the 12700f (not even 12700k), so your numbers very clearly show how ridiculous that is.