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!

 

[moinmoin]

No mention of Raphael-H, and Raphael is confined to "enthusiast desktop". Sounds like Dragon Range is the (new?) name for Raphael-H.

 

[jpiniero]

Hmm.. highest core, thread and cache? I guess Intel hasn't officially announced the Alder Lake HX line yet but I wonder...

 

[lightmanek]

Hmm.. highest core, thread and cache? I guess Intel hasn't officially announced the Alder Lake HX line yet but I wonder...

Here I dare to say it ... SMT4 ... hides behind a sofa 🛋

 

[carrotmania]

Here I dare to say it ... SMT4 ... hides behind a sofa 🛋

Are you TRYING to summon him?

 

[exquisitechar]

No mention of Raphael-H, and Raphael is confined to "enthusiast desktop". Sounds like Dragon Range is the (new?) name for Raphael-H.

Might be reading too much into this, but Raphael is up to the end of 2023, so the speculation that Granite Ridge/Zen 5 DT will launch in 2023 may be wrong. Of course, this is about Zen 4 and they don't want to reveal anything about Zen 5, but they could have been vague about how far Raphael goes instead of having it cover all of 2023. We'll see.

 

[moinmoin]

Might be reading too much into this, but Raphael is up to the end of 2023, so the speculation that Granite Ridge/Zen 5 DT will launch in 2023 may be wrong. Of course, this is about Zen 4 and they don't want to reveal anything about Zen 5, but they could have been vague about how far Raphael goes instead of having it cover all of 2023. We'll see.

AMD slide confirms 2023 will have 50% more days than 2022!

Yeah, personally wouldn't read anything into the spacing on that slide.

 

[DisEnchantment]

No mention of Raphael-H, and Raphael is confined to "enthusiast desktop". Sounds like Dragon Range is the (new?) name for Raphael-H.

Sounds like 3DV Cache Raphael H might be coming with Dragon Range, cache works well for mobile.

I am wondering about that "CPU portfolio" with Xilinx AI engine in 2023 could be Zen5?
Doubt it is Bergamo, most of the cloud folks running some mundane Spring Boot Rest server or some gRPC server wont need AI inferencing.

 

[moinmoin]

Sounds like 3DV Cache Raphael H might be coming with Dragon Range, cache works well for mobile.

I'm not so sure, I'd consider the H series mainstream and V-Cache is more of a premium feature that might come with a too high manufacturing cost to be feasible for a mainstream part.

But the way AMD confidently states "highest core, thread and cache" while Alder Lake HX is already known makes me wonder if both Raphael and Dragon Range will actually be offering a 24 cores or at least 18 cores (3*6) option from the start.

I am wondering about that "CPU portfolio" with Xilinx AI engine in 2023 could be Zen5?

I think we had some discussion earlier already (in this thread even?) when we got news that some IP integration into CPUs will be happening in 2023 already, and that Zen 5 (the next ground up core design family) may be the likeliest candidate.

 

[Hitman928]

But the way AMD confidently states "highest core, thread and cache" while Alder Lake HX is already known makes me wonder if both Raphael and Dragon Range will actually be offering a 24 cores or at least 18 cores (3*6) option from the start.

Probably 16c/32t which would match ADL-HX in core count and beat it in thread count. The highest core count statement they could argue is for 'P' cores. That's my guess anyway.

 

[moinmoin]

Probably 16c/32t which would match ADL-HX in core count and beat it in thread count. The highest core count statement they could argue is for 'P' cores. That's my guess anyway.

That used to be my guess as well, but AMD previously liked to stay clear of just matching the competition's amount whenever it could. Distinguishing between P- and E-cores is too complex for obvious marketing, and Intel is bound to add even more E-cores as a selling point anyway.

 

[jpiniero]

Probably 16c/32t which would match ADL-HX in core count and beat it in thread count. The highest core count statement they could argue is for 'P' cores. That's my guess anyway.

They could have qualified the statement but after thinking about it I think they are just ignoring Alder Lake HX since it hasn't been announced yet.

 

[gdansk]

so do we think Dragon Range is monolithic or not?

 

[moinmoin]

so do we think Dragon Range is monolithic or not?

If it's Raphael-H (which it likely is considering the positioning) it uses the same packaging as Raphael, so MCM or an evolution thereof.

 

[gdansk]

If it's Raphael-H (which it likely is considering the positioning) it uses the same packaging as Raphael, so MCM or an evolution thereof.

If it is similar to the desktop parts I guess that would confirm the presence of an IGP, you'd need it for low power battery life.

 

[moinmoin]

If it is similar to the desktop parts I guess that would confirm the presence of an IGP, you'd need it for low power battery life.

Yeah, that essentially has been confirmed since the Gigabyte leak or so if I recall correctly.

 

[LightningZ71]

Its possible that AMD has seen themselves to enough profitability to invest in a bit of product diversification? Maybe Dragon Range is an MCM that uses a Bergamo and a Genoa CCD together on the same MCM? If we believe that Bergamo is 16 cores per CCD and Genoa is 8 cores, that gives us 24 total cores and 48 threads... that seems rather extreme, but, notipossible. Tied for cores, but significantly more threads?

 

[eek2121]

Its possible that AMD has seen themselves to enough profitability to invest in a bit of product diversification? Maybe Dragon Range is an MCM that uses a Bergamo and a Genoa CCD together on the same MCM? If we believe that Bergamo is 16 cores per CCD and Genoa is 8 cores, that gives us 24 total cores and 48 threads... that seems rather extreme, but, notipossible. Tied for cores, but significantly more threads?

It leaves AMD at a disadvantage. Being able to use a single architecture across your entire stack saves a TON of money and helps with supply. If desktop/mobile/server all use the same chiplets, it is trivial to bin, trivial to shift supply around, and more profitable overall.

Eventually I suspect mobile will be completely unified with desktop as power management improves.

 

[jamescox]

Apple provided that crazy level of bandwidth for the GPU, not the CPU. Intel and AMD are uninterested in building integrated GPUs that scale to high levels of performance, because that's always been the domain of discrete GPUs - which both AMD and now Intel also sell.

Those discrete GPUs AMD and Intel sell include GDDR or HBM - both of which cost more per GB than Apple's LPDDR, but need less of it since it is exclusive to the GPU.

It seems like some laptop makers would be concerned about competing with Apple. Apple’s M1 Pro is pretty good for their first laptop chip; I read something about them doubling shipments vs. the year before. Once people get into the Apple walled garden, they are likely to stay. This should be a concern for other laptop makers.

Rumors of a really powerful APU, that would be severely bottlenecked by a dual channel interface, might indicate that AMD will make something interesting. I have seen quite a few rumors about AND making low end GPUs obsolete. Perhaps they will do higher bandwidth from on package DRAM, commonly already used in most mobile devices, or making up for the low bandwidth with a lot of cache. Stacked devices could go mainstream in 2023. I am wondering if they are going to use an infinity cache base die for mobile also. If it is 128 MB + IO, then it might make sense stack an APU on top for a high end mobile part. It would be a very high end part, so it may not make sense; high end mobile parts are often quite expensive though. The modularity and reusability of using the same base die across many products seems to fit with their design philosophy, but there is still a lot of stuff that doesn’t make sense with that scenario.

 

[moinmoin]

With AMD code names are a rather bad indicator for actual distinct dies. After all we got Raven2, Dali and Pollack all for the same die, just wired up slightly differently. Dragon Range appears to be another such case (though package may be optimized for mobile).

 

[DrMrLordX]

the speculation that Granite Ridge/Zen 5 DT will launch in 2023 may be wrong.

Considering the delay between Vermeer and Raphael, it's entirely possible for Zen5 to launch in 2024 on the desktop.

 

[LightningZ71]

It leaves AMD at a disadvantage. Being able to use a single architecture across your entire stack saves a TON of money and helps with supply. If desktop/mobile/server all use the same chiplets, it is trivial to bin, trivial to shift supply around, and more profitable overall.

Eventually I suspect mobile will be completely unified with desktop as power management improves.

I suspect that things will continue to move away from monolithic over the next several years. Continually shrinking leading nodes are getting more and more expensive, and certain types of circuits are not scaling well through the shrinks. I firmly believe that a tile architecture like what Intel is talking about for Meteor Lake is the path that the whole industry will move towards. Monolithic is not a solution for high performance products, especially as things continue to scale at different rates. It still has it's place in space and power optimized scenarios, but, where power isn't as tightly constrained and space isn't at a premium, it will be far more effective to build with tiles.

 

[Timorous]

No mention of Raphael-H, and Raphael is confined to "enthusiast desktop". Sounds like Dragon Range is the (new?) name for Raphael-H.

Ian needs a better logo than that.

 

[AtenRa]

so do we think Dragon Range is monolithic or not?

I bet for a monolithic design both for Desktop and Mobile

 

[igor_kavinski]

Ian needs a better logo than that.

It's fine. He just needs the wafer to look real, rather than what it is now: cartoonish and looks like it was made by his 5 year old.

 

[Doug S]

It seems like some laptop makers would be concerned about competing with Apple. Apple’s M1 Pro is pretty good for their first laptop chip; I read something about them doubling shipments vs. the year before. Once people get into the Apple walled garden, they are likely to stay. This should be a concern for other laptop makers.

If they believe as I do that almost all Apple customers (whether for iPhone or Mac) first make the decision "iPhone or Android?" and "Mac or PC?", then PC OEMs should not be worried about competing against Apple because by the time someone looks at a PC laptop they've already decided against (or never considered) Apple. With x86 Macs it was possible those who wanted to run Windows could buy a Mac and ignore macOS. That's no longer an option, if you buy a Mac you are using a different OS so have to first make the decision to switch operating systems before you look at hardware.

A PC OEM's competition is other PC OEMs, not Apple. It is clearly FAR easier for a PC OEM to grow at the expense of other PC OEMs versus growing at the expense of Apple - and similarly easier to minimize losses in a declining market like the current one by offering better products than other PC OEMs versus offering better products than Apple.

If Intel or AMD offered solutions like Apple's with higher bandwidth DRAM it would be available to all PC OEMs, so they would gain no advantage relative to each other. They compete by offering better value for the money in terms of price for a given configuration, or better value in terms of build quality / support (the latter is what matters for repeat business)

PCs with memory bandwidth comparable to Apple's would make little difference as far as competing with Apple, except for the very small segment of people who know they are doing something that's highly dependent on memory bandwidth and it is SO important to them that something performs well they choose a Mac despite the hassle of having to switch operating systems.

If they wanted to compete with Apple hardware wise they need to look at the power consumption of a Mac versus that of a PC doing the exact same thing. That's why Apple is gaining customers, not because it has hundreds of GB/sec of memory bandwidth. I doubt more than a few percent of Mac or PC buyers could tell you within an order of magnitude the memory bandwidth of what they're buying.