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!

 

[igor_kavinski]

Zen3D is the holy redacted response to Intel. Look at my old posts. I warned everybody this is intel we are talking about. Remember when AMD had the upper hand because they had better value motherboards and a price/performance price point on CPU's. The memory market has always been a cartel of price fixing. The memory manufacturers were baffled when they lowered the price of memory and consumers bought double the amount of memory they would have bought. It's as if they do not understand supply and demand. When prices are low people buy in bulk.

AMD would be well served with a 20-24 core AM4 Ryzen 3d CPU. Alder Lake wiped the floor with everything except the 5950x. That held it's own against alder lake. I can't hate Intel. Since Core2 every intel processor I ever had lasted 8 years before becoming obsolete.

I will add one more note. AMD promised a core war with Intel. By Zen 3, they should have had at least a 24 core if not a 32 core on the Ryzen line. Go back to the original Zen introduction in 2017. AMD's excuse has been there is threadripper for more cores. But that is more of a server chip than a gaming CPU.

We have a zero tolerance policy for profanity in the tech sub-forums.
Don't do it again.

Iron Woode
Super Moderator

I agree that 5950X should have at least 24C/48T given how expensive it is. Come Raptor Lake, AMD should be ready with a top end 24C/48T V-cache model or it will have difficulty having mindshare with the common users who like to brag about having the fastest at everything CPU. If 12900K didn't have power hogging issues, it almost represents much better value than 5950X in most workloads, other than a few niche ones where the 16 Zen 3 cores excel.

And dude, please be careful when posting. I would hate to see you getting banned.

 

[DisEnchantment]

Zen3D is the holy redacted response to Intel.

Only if you ignore the fact that Vermeer CCD launched in 2020 with the floor plan already laid out for the TSVs. Otherwise AMD would be like, uh let me move these slabs of L3 around and shuffle the IF interconnects to the sides.

I will add one more note. AMD promised a core war with Intel. By Zen 3, they should have had at least a 24 core if not a 32 core on the Ryzen line. Go back to the original Zen introduction in 2017. AMD's excuse has been there is threadripper for more cores. But that is more of a server chip than a gaming CPU.

The high core count 32/24 core CPUs for gaming sounds like an oxymoron currently. If you are gaming you won't be needing 32 cores. (And for ADL kind of ironic that sometimes you need to disable those tiny cores to get better gaming performance)
ADL is a good comeback and welcomed competition to drive innovation forward. It is easy to claim clairvoyance when ignoring everything else
E Cores is Intel's way of doing heterogenous CPUs which have been proven to work well in Mobile and it will do well for ADL for Mobile too. (Just for example Android focus handling mechanism has evolved around the big.LITTLE concept)

Outside of Mobile I would rather have the regular cores.
Imagine a server where your Hypervisor is performing balancing and your VM landed on the small core or an embedded device where you have safety islands and core isolation puts your mission critical application on a low powered core. No thanks.

 

[Joe NYC]

AMD would be well served with a 20-24 core AM4 Ryzen 3d CPU. Alder Lake wiped the floor with everything except the 5950x. That held it's own against alder lake. I can't hate Intel. Since Core2 every intel processor I ever had lasted 8 years before becoming obsolete.

I don't think AMD needs > 16 Zen 3 cores at this time. V-Cache is all that is (was) needed to tie Alder Lake, for the following year, until Raphael arrives.

 

[igor_kavinski]

Outside of Mobile I would rather have the regular cores.
Imagine a server where your Hypervisor is performing balancing and your VM landed on the small core or an embedded device where you have safety islands and core isolation puts your mission critical application on a low powered core. No thanks.

If the E-cores do make an appearance in server CPUs, they will likely have huge caches and all the major hypervisors will be updated with options to prevent mission-critical VMs from being put on E-cores. At least, that's what I think SHOULD happen. Intel can't afford to mess up in the server space anymore.

 

[Doug S]

It seems like AMD always aims at market trends. With Apple putting memory into their CPU package to win benchmarks it is unimaginable that Intel or AMD won't take it up a notch to one-up Apple.

You really think Apple did that for benchmarks? They did it because that was the cheapest/easiest/lowest power way to provide a ton of memory bandwidth to allow an integrated GPU to scale enough to allow its use through the whole line. The large bandwidth available to the CPU is a side effect / side benefit, but not the reason for its inclusion.

 

[eek2121]

You seem to have taken my point as a lack of faith in AMD's drive to succeed.

All I did was make a purely logical deduction about the necessary time to recoup R&D costs on Zen4 and likely availability of fab capacity on N3 nodes.

Zen4 is unlikely to land before late Q3 2022 - more likely Q4 to prevent it cannabilising Zen3D sales.

Therefore the likelihood of any Zen5 chip launching in 2023 seems low to me.

I could be wrong of course - especially if Zen5 is in fact using some advanced variant of TSMC's 5nm based processes.

Zen4 is N5. Unless I missed something, Zen5 will be on one of the N3 nodes. I suppose that it could come out on N4 - but that wouldn't be optimal given the targets set for Zen5.

I suspect that AMD will need a filler product. They cannot rely on 3D V-Cache to carry them through. Especially since they likely could not get such a product out the door until next year. Who knows, I could be wrong. However, Intel is currently iterating faster, and I don't see them slowing down.

It seems like AMD always aims at market trends. With Apple putting memory into their CPU package to win benchmarks it is unimaginable that Intel or AMD won't take it up a notch to one-up Apple. We already are seeing Apple had to cherry-pick benchmarks that matched their strengths, but when doing general tasks they score all over. If AMD has their fabric technology, what stops them from embedding high performance memory chips directly onto their CPU packaging using the lowest possible latency?

With the move towards hardware feeding the OS their finer details, the low latency memory could be fine-tuned to benefit tasks that don't fit into cache. For server-based chips this might make a lot of sense to pare down parts necessary on entry level servers. Plus you get to cherry-pick memory timings. Isn't that what their fabric was more or less promised to do? And if you're worried about volatility of that memory being some kind of drawback compared to Apple's, spec in a battery to keep it from flushing if you kill power. Servers at these prices shouldn't be running without industrial grade UPS backups IMHO. Having that memory on the packaging also dissuades competitors from playing benchmark games using slow memory settings for your products.

I like the embedded memory idea but unlike Apple, I hope Intel or AMD will also offer the ability to expand the on-chip memory with additional memory slots so people are not constrained by just 8/16/32/64 GB SKUs. If Apple offered a 128GB MBP for $4000, I would have bought it instantly. But now with DDR5, I'm thinking it would be better to have 256GB. More RAM can never hurt. Heck, I was thinking about getting a used server with 1TB RAM but then it crossed my mind that the server chassis is going to be extremely heavy. I hate hardware that I have trouble lifting. I gave away my Core 2 Quad in return for 50 Blurays coz it was in a Gigabyte case weighing 10KG. I hated it!

Both Intel and AMD need to add at least 4gb of high-speed memory, or at least have a chipset that includes it. It would significantly improve performance in a variety of workloads. I'd prefer to see 16gb. Then I can put another 32gb in the memory slots.

@Hans Gruber

That's overstating things a bit. AMD is not in any trouble. Their primary focus is on servers/workstations where Milan is already top dog and where Genoa is taking control. It really won't rattle them much that Alder Lake manages to win some benchmarks against lower core-count Vermeer products. Zen3D is not a "holy" anything response. It's a way to extend AM4's shelf life a little so they have flexibility on when to release Raphael. Plus with Milan-X in production, why not make use of some of the cast-off dice for a consumer Vermeer-X product?

As far as core wars go, software just won't scale infinitely. Adding more and more cores will only help a smaller and smaller segment of the buying population. There probably isn't any need for more than 12c on the consumer desktop right now, with 16c being a rather small-but-lucrative niche where AMD can drive higher ASPs by forcing performance users who want the best bins and best boost clocks to buy more cores than the normally need. 24c and 32c AM4 products would be, frankly, idiotic. We may eventually see higher core counts with Zen5, but Raphael will probably not go past 16c. Remember Intel can not now or in the forseeable future produce desktop CPUs with more than 8 P cores. The E cores are frankly irrelevant.

No, they aren't in trouble at all, but they can't continue at the pace they are continuing at. If Intel sticks to a less than yearly cadence, AMD is going to eventually fall behind.

 

[maddie]

I suspect that AMD will need a filler product. They cannot rely on 3D V-Cache to carry them through. Especially since they likely could not get such a product out the door until next year. Who knows, I could be wrong. However, Intel is currently iterating faster, and I don't see them slowing down.


Both Intel and AMD need to add at least 4gb of high-speed memory, or at least have a chipset that includes it. It would significantly improve performance in a variety of workloads. I'd prefer to see 16gb. Then I can put another 32gb in the memory slots.

No, they aren't in trouble at all, but they can't continue at the pace they are continuing at. If Intel sticks to a less than yearly cadence, AMD is going to eventually fall behind.

I'm curious.

How do you see it possible to have and maintain a less than yearly cadence?
Where did this enormous boost in creativity, productivity and execution come from?
Why was it hidden so long and now suddenly have been released?

 

[Ajay]

I suspect that AMD will need a filler product. They cannot rely on 3D V-Cache to carry them through. Especially since they likely could not get such a product out the door until next year. Who knows, I could be wrong. However, Intel is currently iterating faster, and I don't see them slowing down.

I fee like a broken record, but Zen 3D is the 'filler' product. Finito!

 

[itsmydamnation]

No, they aren't in trouble at all, but they can't continue at the pace they are continuing at. If Intel sticks to a less than yearly cadence, AMD is going to eventually fall behind.

A less then yearly of what?
I love how the one company that has consistently failed to deliver is now producing sizable IPC gains sub yearly in some peoples minds. People like to pretend that 10nm shenanigans meant golden cove/sunny cove got delayed and Intel has been firing on all cylinders ( arch side) but that's not really true. Cannon lake disappeared and everything continued on as normal.

Lets just look at Golden cove, they threw massive resources at the core and are only ~5% IPC a clock over Zen3.
Look at the AMD patients they are all about smarter not bigger. So when Zen4 comes along in ~6months time , has ~0.5-0.75 factor of the resources per core and beats it with ~10-15 point on IPC how is AMD in trouble? When Zen5 comes along ( looks like its AMD's big core) in 24 months and has the same smarter not bigger but now has comparable resources per core to intel. Who is going to be in trouble?

Remember Golden Cove is a A15 sized core delivering nowhere near A15 sized results.

ARM, Apple , AMD have all in effect shown increased performance @ ISO process , generation on generation, Intel hasn't .How is intel going to scale their core performance, intel haven't really done anything to make their cores smarter since sandy bridge. Intels the one that needs to catch up massively in terms of architectural efficiency.

 

[itsmydamnation]

If the E-cores do make an appearance in server CPUs, they will likely have huge caches and all the major hypervisors will be updated with options to prevent mission-critical VMs from being put on E-cores. At least, that's what I think SHOULD happen. Intel can't afford to mess up in the server space anymore.

right so Vmware, microsoft , KVM/redhat are all going to redo there Scheduler, QOS and resource allocation functions that took decade to get to where they are today, just like that cuz intel?

Just a point release on cpu release, easy as , right?

Can you imaging the regression testing nightmare.

What will happen is any software vendor that has specific vm configuration requirements ( almost all "high performance" application ) will just update to say software not supported on big little configured servers and just move on with there software development plans and will work in how they handle it sometime in the future.

Do you even Hypervisor ? how about in Production?

 

[igor_kavinski]

Do you even Hypervisor ? how about in Production?

Intel throws money around when it suits their interests.

We have production VMs in Hyper-V on Windows Server 2012 R2 using Xeon v2 CPUs. Their performance is abysmal (40% loss in NVMe SSD throughput). Not sure if it is that particular Hyper-V version that's to blame or is it the Xeon CPU lacking due to being behind on VM related hardware enhancements since its release. Going to get an Azure cloud instance soon and test that to see how the I/O performance fares on that.

 

[itsmydamnation]

Intel throws money around when it suits their interests.

We have production VMs in Hyper-V on Windows Server 2012 R2 using Xeon v2 CPUs. Their performance is abysmal (40% loss in NVMe SSD throughput). Not sure if it is that particular Hyper-V version that's to blame or is it the Xeon CPU lacking due to being behind on VM related hardware enhancements since its release. Going to get an Azure cloud instance soon and test that to see how the I/O performance fares on that.

,
Off topic, in all honesty software based I/O that does anything smart , sux in terms of performance , cohesity , Nutanix , vsan, storage spaces etc. You can get good across cluster aggregate throughput , but per vm is not even close to a 1/2 good RAID card even in something like RAID 6, let alone something insane like RAID 0, or striping JBOD . But what you get is scale out capability, which is almost always worth it . if you have a snow flake VM , SR-IOV or PCI pass though ( has its own set of trade offs ) a RAID card with dedicated disk and be done with it.

If you have a small vm cluster , almost always go with a solid basic RAID 10/5/6 capable card and create volumes as required using fault tolerances/HA at the vm layer its alot cheaper then a Netapp FAS or some other all flash filer type device.

 

[jpiniero]

Intel throws money around when it suits their interests.

We have production VMs in Hyper-V on Windows Server 2012 R2 using Xeon v2 CPUs. Their performance is abysmal (40% loss in NVMe SSD throughput). Not sure if it is that particular Hyper-V version that's to blame or is it the Xeon CPU lacking due to being behind on VM related hardware enhancements since its release. Going to get an Azure cloud instance soon and test that to see how the I/O performance fares on that.

Smeltdown...

 

[Mopetar]

I like the embedded memory idea but unlike Apple, I hope Intel or AMD will also offer the ability to expand the on-chip memory with additional memory slots so people are not constrained by just 8/16/32/64 GB SKUs.

Maybe I'm completely misunderstanding what you're suggesting, but how do you have expansion slots for on-chip memory? The whole point is that it's on the chip and that as a result of this the access time is significantly faster. As soon as you move off-chip you've maybe got a slightly faster version of RAM, but the difference in access time between RAM and cache is already so large that it wouldn't matter.

The other problem is that even if you could increase the cache size in some kind of magical way, doing so tends to make it slower and hurts performance for anything that doesn't need the larger cache size.

 

[igor_kavinski]

Maybe I'm completely misunderstanding what you're suggesting, but how do you have expansion slots for on-chip memory? The whole point is that it's on the chip and that as a result of this the access time is significantly faster.

The slowest RAM is still faster than the fastest secondary storage. I don't mind if the off-chip RAM is slower. Majority of the workloads would get most of their work done in the on-chip 16GB RAM.

 

[DrMrLordX]

However, Intel is currently iterating faster, and I don't see them slowing down.

They will slow down after Raptor Lake. More or less.

How do you see it possible to have and maintain a less than yearly cadence?
Where did this enormous boost in creativity, productivity and execution come from?
Why was it hidden so long and now suddenly have been released?

Golden Cove, Willow Cove, and Sunny Cove took too long to make it to market thanks in no small part to process problems. Stuff is coming off the drawing board in clumps. Intel has essentially blown their load with Alder Lake (Raptor Lake will not be much different). Gains will slow down until they bring their 20A process online. Just my prediction.

 

[igor_kavinski]

Gains will slow down until they bring their 20A process online. Just my prediction.

They need miracle workers to drag them out of their puke.

 

[Markfw]

Alder lake barely beats the 5950x, except gaming. And all the big.little problems. I see a tie as best case. And thats just desktop.

Server and HEDT they loose badly. Mobile ? still a tossup.

 

[Mopetar]

The slowest RAM is still faster than the fastest secondary storage. I don't mind if the off-chip RAM is slower. Majority of the workloads would get most of their work done in the on-chip 16GB RAM.

You're not getting 16 GB on chip RAM any time soon. Look at a recent die shot for most CPUs and the L2 cache takes up as much if not more space as the cores. SRAM takes up more space than DRAM in order to achieve the faster access speeds it has. Zen 3D is looking to add an extra layer of 64 MB of SRAM. Maybe over the next several generations we'll be able to get two or three layers in stacked chips, but it's still going to be a far cry short of what you're imagining.

It sounds like what you want is faster RAM. But it's going to be off-chip and have a big step up in access time over anything on the chip. Anything that's going to be significantly faster than existing RAM probably isn't going to be expandable, at least as easily as RAM. At the least it would require some soldering.

 

[igor_kavinski]

Anything that's going to be significantly faster than existing RAM probably isn't going to be expandable, at least as easily as RAM. At the least it would require some soldering.

I suck at elaborating. Yes, soldered RAM in close proximity like what Apple did with M1 but also RAM slots for expanding with relatively slower RAM. As long as programs don't exhaust the first level RAM, they will run pretty well. When they spill into the slower RAM, it still won't be a total disaster. Much better performance than paging.

 

[soresu]

You're not getting 16 GB on chip RAM any time soon. Look at a recent die shot for most CPUs and the L2 cache takes up as much if not more space as the cores. SRAM takes up more space than DRAM in order to achieve the faster access speeds it has. Zen 3D is looking to add an extra layer of 64 MB of SRAM. Maybe over the next several generations we'll be able to get two or three layers in stacked chips, but it's still going to be a far cry short of what you're imagining.

I think that the industry is (very slowly) trending towards a change from SRAM to some form of MRAM which could lead to both power efficiency increases due to persistance and dramatic area bit density increases.

The cache stacking strategy in particular means that it will be far easier to manufacture the MRAM on a separate process entirely from the main logic so we could easily see multi GB caches per socket even in mainstream platforms.

 

[NostaSeronx]

@soresu
NRAM tech is technologically ramping faster than MRAM.

2024+ => 2 GB/16 Gb NRAM-LLC Stack
SRAM Speed&Endurance ✔
DRAM Density ✔
NAND Power&Cost ✔

NRAM is also heat-tolerant thus can low-key be used as a heatsink. => "exceeds that of diamond, which is the best bulk heat conductor"

 

[mikk]

They will slow down after Raptor Lake. More or less.

Golden Cove, Willow Cove, and Sunny Cove took too long to make it to market thanks in no small part to process problems. Stuff is coming off the drawing board in clumps. Intel has essentially blown their load with Alder Lake (Raptor Lake will not be much different). Gains will slow down until they bring their 20A process online. Just my prediction.

Your prediction is odd as always. If anything Raptor is the most boring within Intels next follow up generations because it's still based on Golden Cove and Intel 7. Meteor, Arrow, Lunar are the interesting ones. Intel 4/3 and possibly even a TSMC 3nm/4nm node is coming before 20A, there is enough room for big gains after Raptor and before 20A, don't worry.

 

[Mopetar]

I don't think MRAM is intended as a replacement for SRAM. It's not as fast and has a more limited lifespan. Maybe it'll replace DRAM eventually (or in some some special cases where data retention is important), but isn't is mainly being developed as a flash replacement/substitute at the moment?

 

[MadRat]

Does AMD work with glass interposer technology?