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!

 

[soresu]

AMD should release either a 20 core 40 thread or 24 core 48 thread CPU on AM4 with the 3D V-cache.

Could that even fit in the AM4 package without an IOD shrink?

Either way it's probably pushing a TDP boundary for the socket that they don't want to cross.

 

[Ajay]

Aren't they more likely to be using multi node servers anyway whether the workload is CPU or GPU bound?

Depends, if one is doing smaller data science projects (just multi-terabytes data sets) or Cuda based ML projects with a couple of GPUs, then TR is fine. Also, for initial development, distributed multi server apps can be developed on something like TR system (or Epyc based 'desktop') for the core code and then scaled up.

 

[Ajay]

Could that even fit in the AM4 package without an IOD shrink?

Either way it's probably pushing a TDP boundary for the socket that they don't want to cross.

I think the short answer is no. There is no business case for that right now. Repeating it for the 10 millionth time (no offense), Zen 3D is the end of the AM4 Zen3 lineup (at least for DIY). There have been rumors about 3-CCD CPUs on AM5 - we'll have to wait and see.

 

[Joe NYC]

Alder Lake beat every Zen 3 chip other than the 5950x. AMD should release either a 20 core 40 thread or 24 core 48 thread CPU on AM4 with the 3D V-cache. That would give them time to get Zen 4 ready while giving consumers a reason to stick with AMD.

Threadripper should have been released 6 months ago. With that said threadripper is in a different league than the mainstream desktop CPU's.

I don't think there is any point - and possibly no technical way - to release Zen 3 with more than 16 cores. The IOD can accommodate only 2 CCDs, and to accommodate more, a far more expensive IOD would be needed.

There are rumors AMD is killing the Threadripper 4 memory channel platform, so only Uber expensive Threadripper Pro platform remains, which is basically EPYC platform. It is more of a workstation, not an HEDT platform.

AMD could have made a difference with Zen 3D release on desktop, to tie Alder Lake in gaming, but it was put on the back burner. And now it looks the whole desktop segment is on the backburner, with Genoa taking precedence over Raphael.

 

[soresu]

There have been rumors about 3-CCD CPUs on AM5 - we'll have to wait and see.

It makes sense to me given the 'golden rule' so far from Zen1 to Zen3 has been AM4 max core count = 1/4 TR max core count of a given generation.

I reckon if we do get the Epyc max core count (eventually) in TR form , then the AM5 max count lands about the same time that the TR platform lands - they land together both require that AMD has the necessary CCD inventory for these higher core count SKUs.

I do wonder if perhaps Zen5 may finally transition to 16 core CCD's.

I can't see it happening at Zen4 generation, not unless for some reason they launch a consumer Zen4c SKU - which I highly doubt unless it's reinforced with V cache to address the expected L3$ reduction for Zen4c.

 

[Exist50]

I do wonder if perhaps Zen5 may finally transition to 16 core CCD's

That'll be tricky. They switched to a ring for the 8 core CCX, and rings have historically scaled poorly to 16. But Intel seems like they'll be trying for the same in the same time period, so we'll see. Could always switch to a mesh if necessary.

 

[Hans Gruber]

AMD has said in the past that they could have a core war with Intel. I think the 5950x was all they needed when Zen 3 was launched. With Zen 4 (5nm) they will have more real estate to add a lot more cores. I think AMD dropped a ball a bit with threadripper. It was not a better gaming CPU than the standard Ryzen processors. Had they optimized it for gaming as well as a number crunching machine. There would have been a lot more demand for threadripper.

It's good times for the CPU battle right now. I was thinking a hold over CPU with 20 or 24 cores would make the transition from AM4 to AM5 easier. People could hold off on an AM5 build until the DDR5 latencies and speed of the memory improve. Perception is key in the CPU industry. Alder Lake was rushed out because of Zen 3.

I should note I will be getting a Zen 3D CPU. Either the 6 core 12 thread or a the 8 Core 16 thread. I have a 3 monitor array and I think I could benefit more from at least 8 cores and 16 threads. That will do it for me for a couple of years before jumping to AM5.

 

[soresu]

I should note I will be getting a Zen 3D CPU. Either the 6 core 12 thread or a the 8 Core 16 thread. I have a 3 monitor array and I think I could benefit more from at least 8 cores and 16 threads. That will do it for me for a couple of years before jumping to AM5.

Eh? For someone who thinks AMD should be going harder on AM4 core counts you seem to have your sights set pretty low.

12C minimum or bust! 😎

 

[Abwx]

It makes sense to me given the 'golden rule' so far from Zen1 to Zen3 has been AM4 max core count = 1/4 TR max core count of a given generation.

I reckon if we do get the Epyc max core count (eventually) in TR form , then the AM5 max count lands about the same time that the TR platform lands - they land together both require that AMD has the necessary CCD inventory for these higher core count SKUs.

I do wonder if perhaps Zen5 may finally transition to 16 core CCD's.

I can't see it happening at Zen4 generation, not unless for some reason they launch a consumer Zen4c SKU - which I highly doubt unless it's reinforced with V cache to address the expected L3$ reduction for Zen4c.

Whatever they do with CCDs methink that Zen 4 chiplets and I/O are designed such that they can get up to 32C on DT.

That make sense since they stated that 5nm perf/watt is 2x the one of 7nm at isofrequency, otherwise they would be stuck with 16C clocked uselessly high and consequently with very poor perf/watt.

From the published slide they can either do 32C at 150W, so clocked slightly lower than Zen 3 to keep being at 125W at 2x the perf, or else a 16C at barely 1.15x the frequency of Zen 3, wich would amount to only 1.35x the perf when accounting for the expected IPC.

 

[jpiniero]

Whatever they do with CCDs methink that Zen 4 chiplets and I/O are designed such that they can get up to 32C on DT.

Dual Zen 4c perhaps?

 

[Abwx]

Dual Zen 4c perhaps?

There will be Zen 4c chiplets and Zen 4HP chiplets, so they can do eventually a mix of the two designs, say 16 + 16.

 

[Joe NYC]

Dual Zen 4c perhaps?

I don't think Zen4c is being optimized for highest Single Threading performance, so not a good fit for Desktop / Gaming CPU.

 

[DrMrLordX]

It's a gamble, that Intel will not get its act together and will keep ordering.

Other companies (Apple, AMD) taking market from Intel is a non-brainer for TSMC, not a gamble.

There is nothing better about TSMC silicon being branded "Intel" rather than "Apple", or "AMD", or "MediaTek".

Apple and MediaTek can't take most of Intel's market share away from them. Especially not in the server room. AMD can take any of Intel's market share (eventually), but what if AMD had no plans to use N3 in 2023?

There is no such thing as Intel market.

There actually is, even today. You'd be surprised by how many vendors are either contractually required to supply Intel products, or simply do so out of supplier and customer bias. No matter how good were Rome and Milan. No matter how good will be Genoa. In some cases, it will take years to get people off Intel. It's like pulling teeth. AMD will need two or three more generations of dominance with continued increases in capacity to keep chipping away and maybe hit 50% server market share (or better). They have a long way to go!

Ditto with NVidia making space heaters. That opportunity is going to be greater than AMD anticipated...

The CUDA mafia is strong. Maybe HIP/SYCL will get people to start running their CUDA projects on other hardware, assuming the features they need are fully-supported.

Considering the current capacity crunch it's clear that the demand is there already to justify the investment with or without profits from Intel.

Lawd only knows what size mountain of greenbacks Apple lays down to get such a chunk of early state of the art nodes but it almost certainly helps TSMC fund new expansion.

Allegedly, TSMC will build capacity and charge after-the-fact for its most-loyal customers, while charging up-front to its least loyal customers. Intel, nV, and Qualcomm have to make the payments up front. That's what some commentators in the media are saying anyway. In any case, TSMC probably made Intel pay up-front for future N3 capacity. They likely did not take a share away from someone else. Certainly not AMD. If AMD's cadence is to be believed, Zen5 is a 2024 CPU anyway. Not sure if it will be on N3 or N4.

If Intel wanted to hurt AMD today, they would have outbid AMD for all their N5 wafers. That did not happen. Something to think about.

 

[Joe NYC]

Apple and MediaTek can't take most of Intel's market share away from them. Especially not in the server room. AMD can take any of Intel's market share (eventually), but what if AMD had no plans to use N3 in 2023?

N3 is different from the other nodes, that are all in the sweet spot. Good luck to Intel on the half baked, over-priced N3. TSMC is smart to take money for that.

But the other nodes, N7 to N4, I think it is counterproductive for TSMC to give weapons to the enemy (Intel) to fight customers and allies.

There actually is, even today. You'd be surprised by how many vendors are either contractually required to supply Intel products, or simply do so out of supplier and customer bias. No matter how good were Rome and Milan. No matter how good will be Genoa. In some cases, it will take years to get people off Intel. It's like pulling teeth. AMD will need two or three more generations of dominance with continued increases in capacity to keep chipping away and maybe hit 50% server market share (or better). They have a long way to go!

But those are legacy barriers that Intel managed to bribe its way into are not real "requirements".

The longer AMD has the advantage, the more the suits who came up with these look like complete idiots.

The CUDA mafia is strong. Maybe HIP/SYCL will get people to start running their CUDA projects on other hardware, assuming the features they need are fully-supported.

For datacenter cards - yes.

For consumer cards - less strong, or not strong at all.

It is good to reminder that NVidia sells nearly as much in consumer space heaters as the entire AMD revenue from Servers CPU + mobile CPU + Desktop CPU + Desktop GPU + mobile GPU + console.

That is the size of the opportunity in consumer GPU alone. (re: AMD has all the N5 capacity it needs)

Allegedly, TSMC will build capacity and charge after-the-fact for its most-loyal customers, while charging up-front to its least loyal customers. Intel, nV, and Qualcomm have to make the payments up front.

If this is in fact true, it is quite funny to see Jensen waving his hands, victoriously claiming he prepaid for a ton of capacity.

That's what some commentators in the media are saying anyway. In any case, TSMC probably made Intel pay up-front for future N3 capacity. They likely did not take a share away from someone else. Certainly not AMD. If AMD's cadence is to be believed, Zen5 is a 2024 CPU anyway. Not sure if it will be on N3 or N4.

If Intel wanted to hurt AMD today, they would have outbid AMD for all their N5 wafers. That did not happen. Something to think about.

It would take TSMC to agree to it.

And I don't think TSMC execs are completely suicidal. Meaning, they don't put bullets in all chambers while playing Russian Roulette with Intel. But they are playing Russian Roulette by selling to Intel from their healthy nodes.

(I would separate N3 and N4, because N4 looks like it is going to be healthy)

BTW, just like Google getting rid of their old motto "Don't be evil" when it was no longer possible to say it with a straight face, I think TSMC should get rid of the "everyone's foundry" motto.

 

[soresu]

For datacenter cards - yes.

For consumer cards - less strong, or not strong at all.

There's quite a bit of wiggle room between those 2.

Unfortunately nVidia OptiX (RT use specific CUDA API?) has basically tied up every single commercial GPU renderer out there on Windows - Autodesk Arnold, Pixar Renderman, OTOY Octane and Maxon Redshift.

Only non commercial Blender Cycles X has a HIP backend, kudos to AMD and whoever at the Blender Foundation for rushing that into Blender v3.

SideFX Houdini's upcoming Karma GPU renderer is supposedly targetting OpenCL, but with any luck they will switch to SYCL or HIP as it seems like OpenCL has little future prospects now.

Thankfully Adobe committed to Vulkan for their Mercury engine GPU backends

 

[soresu]

BTW, just like Google getting rid of their old motto "Don't be evil" when it was no longer possible to say it with a straight face, I think TSMC should get rid of the "everyone's foundry" motto.

This is a difficult point - TSMC is unfortunately subject to external political pressure, and as a result of that lost more than 10% of their paying business previously allotted to Huawei/HiSilicon for fabbing Kirin SoC's.

That left them flailing a bit combined with being threatened by the US govmt with the loss of ASML litho machines which probably caused them to drop further Chinese SoC contracts

It is good to reminder that NVidia sells nearly as much in consumer space heaters as the entire AMD revenue from Servers CPU + mobile CPU + Desktop CPU + Desktop GPU + mobile GPU + console.

Is the console contract revenue to TSMC actually gained from AMD per wafer?

Or from Sony and MS per wafer, and AMD just acted as the partial SoC designer by contract to them with royalties per chip?

 

[naukkis]

Is the console contract revenue to TSMC actually gained from AMD per wafer?

Or from Sony and MS per wafer, and AMD just acted as the partial SoC designer by contract to them with royalties per chip?

AMD sells finished product. It ain't possible for AMD to license x86 to outside parties even if they wanted.

 

[turtile]

AMD sells finished product. It ain't possible for AMD to license x86 to outside parties even if they wanted.

They can sell the design (already have in China).

 

[jpiniero]

They can sell the design (already have in China).

Only in China though, because China is "special"

 

[Hitman928]

They can sell the design (already have in China).

From memory, that was through a joint venture where AMD owned 51% of the company, so technically it was still AMD owning the design.

 

[The Hardcard]

Only in China though, because China is "special"

You’ve heard about entities from other nations offering a proper amount of cash and/or other resources and being told no? Other than a willingness to put in money and effort, how is China “special”?

 

[IntelUser2000]

If Zen 3D is coming for desktop and servers I don't expect Zen 4 until Q3 or Q4 of this year.

I think AMD dropped a ball a bit with threadripper. It was not a better gaming CPU than the standard Ryzen processors. Had they optimized it for gaming as well as a number crunching machine. There would have been a lot more demand for threadripper.

That's not possible without them designing an entirely separate CPU and platform from Ryzen and Threadripper. Threadripper uses the server platform.

It's not like there was a huge difference anyway. People are arguing about 5-10% differences. That extra 5-10% requires a low latency memory controller and algorithms designed for that.

Almost like saying SUVs aren't good as sedans in pure passenger cars! Of course not!

 

[Hans Gruber]

If Zen 3D is coming for desktop and servers I don't expect Zen 4 until Q3 or Q4 of this year.



That's not possible without them designing an entirely separate CPU and platform from Ryzen and Threadripper. Threadripper uses the server platform.

It's not like there was a huge difference anyway. People are arguing about 5-10% differences. That extra 5-10% requires a low latency memory controller and algorithms designed for that.

Almost like saying SUVs aren't good as sedans in pure passenger cars! Of course not!

I am not disagreeing with you. I am simply saying that every time consumers ask for more cores on the mainstream lineup. The answer is threadripper. You get all those PCI lanes and a huge CPU that requires somewhat specialized coolers and expensive motherboards. When Intel makes better CPU's, AMD's answer will be more cores. It's hard to sandbag against Intel.

 

[IntelUser2000]

You get all those PCI lanes and a huge CPU that requires somewhat specialized coolers and expensive motherboards. When Intel makes better CPU's, AMD's answer will be more cores. It's hard to sandbag against Intel.

Well sure. What else is AMD going to do? Unless they decide to can threadripper altogether then you end up with the distinction.

Same will be true with Sapphire Rapids, as it was with Skylake-X. Skylake-X was 5-10% slower, because it's a server platform. Sapphire Rapids-X is going to be a monster HEDT platform, but it'll lag Alderlake(probably Raptorlake at that time) in games.

 

[DrMrLordX]

I am simply saying that every time consumers ask for more cores on the mainstream lineup. The answer is threadripper.

Who is asking for more cores in AMD's mainstream lineup? People in the 24-32c+ contingent are asking for an updated Threadripper. That's the crowd that wants the extra memory channels, PCIe lanes, and other features missing from AM4 (and that will still be missing from AM5).