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!

 

[nicalandia]

So far we have just this picture, which may never materialise into a real and competitive product.

View attachment 68685

We have tangible proof of Monolithic dies (The Xeon Silver 4410T) entries on Sisoftware Sandra. There are 32C/64T Entries to with 60MiB L2$ But those could also be on 4 tile based CPUs.


Silver 4410T (10C 20T 2.7GHz/4GHz, 10x 2MB L2, 26.25MB L3)

Details for Result ID 2x Intel(R) Xeon(R) Silver 4410T (10C 20T 2.7GHz/4GHz, 10x 2MB L2, 26.25MB L3)

ranker.sisoftware.co.uk

 

[In2Photos]

Not sure if serious. Doesn't disabling CPB essentially set the CPU to the base clock and disable any boost? We already had plenty better ways to reduce temperature and increase efficiency.

This is the one chart I wish he provided as well, CPU clocks during the testing. He mentioned that the benchmarks are not "sustained loads". So the effective clock speeds for the entire test could be the same, which is why the performance would be the same.

 

[Kaluan]

How thick is the vcache die? Via google I'm seeing values for general die thinkness post backgrinding of the order of 300μm although it can go down to the 30-50μm depending on the nature of the die and how it will be used.

The reason I am asking is that there is some noise on another forum speculating that the X3D IHS will be thinner and therefore can help the X3D die run cooler or at least help counteract the 'thermal blanket' from the silicon spacer over the logic areas. But given the expected thickness of the vcache dies involved, I just can't see that happening or at least not having a great effect. If all dies gets backgrinded, then it would make sense to keep the IHS the same thickness and just remove slightly more from the cpu die than normal as part of the standard process.

Thoughts?

Kept thinking of this myself.
There's certainly a chance AMD may have saved a ace down it's sleeve in relation to die height/IHS thickness, so they can contrast the performance boost the V-Cache SKUs get even better. I'm thinking they don't want to sacrifice any productivity oomph (generally speaking, the clocks) with X3D vs vanilla, maybe even surpass it (especially if the massive MALL helps them mitigate some of their FCLK/IMC "shortcomings", so even faster in memory sensitive workloads).

I don't think AMD wants a 7950X3D that's 5-10% slower in Blender than the cheaper 7950X3D, stellar gaming perf aside.

Either way, should be a interesting product line, probably more so than 5800X3D.

They may tease or even announce them at the RDNA3 launch.

 

[nicalandia]

How thick is the vcache die? Via google I'm seeing values for general die thinkness post backgrinding of the order of 300μm although it can go down to the 30-50μm depending on the nature of the die and how it will be used.

Not Thicker than Usual..

https://twitter.com/x/status/1531664784680173569

 

[LightningZ71]

We are reasonably sure that AMD has the capability to stack the VCache several layers high, at the very least from some EPYC board bios settings that were passed around last year not to mention statements supporting the possibility from AMD reps. It is believable that AMD could go as high as 4 stacks on consumer parts, resulting in up to 256MB of stacked cache on ZEN4 parts. Except for some very specific cases, I don't see how any measure of better RAM controller performance will help Raptor lake have a memory throughput edge there. I don't think that those would be inexpensive SKUs by any measure, but they COULD exist if AMD wanted them to.

 

[Kepler_L2]

We are reasonably sure that AMD has the capability to stack the VCache several layers high, at the very least from some EPYC board bios settings that were passed around last year not to mention statements supporting the possibility from AMD reps. It is believable that AMD could go as high as 4 stacks on consumer parts, resulting in up to 256MB of stacked cache on ZEN4 parts. Except for some very specific cases, I don't see how any measure of better RAM controller performance will help Raptor lake have a memory throughput edge there. I don't think that those would be inexpensive SKUs by any measure, but they COULD exist if AMD wanted them to.

Not coming this generation.

 

[Doug S]

No expansion? it has thunderbolt, what more do you need these days? Workstation users in a pro environment don’t use the internal expansion capabilities of a PC.

Shoot, in a decade of IT work the only internal expansion slot used by anyone, even with all those large Power Macs I had to deal with was for a GPU.

If they run out of storage, give them an external thunderbolt drive. Most companies these days don’t want you storing stuff locally, so storage space is rarely an issue.

If RAM becomes an issue, that also means the machine is due for an upgrade.

There may be small niches where expansion is desired (someone doing local ML work comes to mind, with 2-4 GPUs), but outside of that, companies don’t care. If they did, all things Apple wouldn’t be selling like hotcakes.

I will be glad when Intel returns to HEDT so AMD gets a wakeup call.


Yep, this.

The lack of RAM upgrades in the Studio is a much bigger problem than the lack of slots. As you say, storage upgrades aren't an issue with multiple TB slots, and the GPU is what it is. Apple will not support third party GPUs at all in the ARM version of macOS, they've been pretty clear about that. There will not be a slot for a GPU in the Mac Pro.

Years ago when I was in charge of IT support for a division of a major university (almost all the 'science'/math related departments) we had a lot of RISC workstations and PC workstations were out there too but not competitive performance wise so they were kind of the "on the cheap" option. I can't remember ever doing a single after purchase CPU or GPU upgrade. If you needed more power, a new workstation was acquired and the old one became a hand me down to someone else in the department or the grantee's grad students. These workstations didn't get local storage upgrades either - with very rare exceptions the local filesystems were not backed up so all the important data lived on a central fileserver (a small portion of each grant was carved out to fund a high performance RAID array with nightly backups)

What did happen all the time after purchase were RAM upgrades. But typically because they found they needed more RAM than when it was purchased - people were pretty sharp about knowing how their needs would grow in their particular problem space, and generally there was only so much RAM that could be properly utilitized by a given amount of CPU power. The main reason after-purchase RAM upgrades happened was because they needed more RAM from day one but either that amount of RAM was unaffordable at the time of purchase (i.e. their NSF grant only funded so much so that was planned for later) or that density of RAM simply didn't exist at the time of purchase.

Since you can't upgrade RAM later in the Studio you have to be sure that the amount of RAM it has will either be adequate for your workflow for several years, or you have to have somewhere to "hand me down" that Studio when it is replaced. For that reason I think something like the Studio makes much more sense for a workplace with multiple people than it would for any individual unless they knew their needs would not grow beyond it too quickly.

Unless everyone is doing exactly the same job, in most workplaces that require workstations you have a sort of pyramid of need. Those at the top with the biggest needs are always getting something new, and stuff gets handed down to those with lesser needs. You probably aren't going to buy whatever top end x86 workstation for someone and later upgrade their particular workstation too often - then you have the problem of "what do I do with the part(s) I replaced" (which you would probably just eBay if you were an individual) and unless you want to do a crapload of work swapping parts down and down the chain (after checking for compatibility issues) it is much simpler to swap whole machines.

 

[moinmoin]

Not coming this generation.

That's unfortunate. AMD is keeping that particular card close to its chest for quite long.

 

[nicalandia]

That's unfortunate. AMD is keeping that particular card close to its chest for quite long.

So far there is no need to increase the layers because 1 has been enough to acomplish the task. A single 7800X3D is enough to secure gaming dominance till late 2024

 

[eek2121]

Kept thinking of this myself.
There's certainly a chance AMD may have saved a ace down it's sleeve in relation to die height/IHS thickness, so they can contrast the performance boost the V-Cache SKUs get even better. I'm thinking they don't want to sacrifice any productivity oomph (generally speaking, the clocks) with X3D vs vanilla, maybe even surpass it (especially if the massive MALL helps them mitigate some of their FCLK/IMC "shortcomings", so even faster in memory sensitive workloads).

I don't think AMD wants a 7950X3D that's 5-10% slower in Blender than the cheaper 7950X3D, stellar gaming perf aside.

Either way, should be a interesting product line, probably more so than 5800X3D.

They may tease or even announce them at the RDNA3 launch.

I have seen at least 2 independent rumors indicating that AMD has at least some internal chips running at 6ghz single core boost. If AMD has solved the clock issues with the X3D parts (which someone at AMD hinted was related to power draw) then it is possible we could see higher clocks (5.8-6ghz). Intel is also said to be launching a 6ghz SKU.

 

[aigomorla]

HEDT has no coined definition, it is just an acronym for "high end desktop computer"

The original acronym was High Enthusiast DeskTop.

That is why we had a Enthusiast tier which just meant a overclockable enterprise edition.
Hence allowed you to overclock server parts and run them much faster.

The first Enthusiast component we got was again the LGA1366 (Nehalem) while the Consumer side was LGA1156 (Lynnfield)
LGA1366 was a enterprise class cpu with bclk unlocking on the board side and multi unlocking on the Extreme Edition.
LGA1366 later on became eVGA classified, and eVGA Classified SR-2, which i would say were the true start of real "HEDT" from CPU to motherboard, as the SR-2 had special Xeon CPU's with unlocked multipliers you could purchase.

EVGA Classified SR-2 Review | bit-tech.net

Build your own supercomputer with two overclocked CPUs and up to seven graphics cards.

bit-tech.net

This is where HEDT first came out, and had defined walls.

HEDT and Normal segement never shared the same platform.
HEDT always had its own board and own pinset.

"K" edition CPU's which used "Z" boards can not be used in boards designed for "X" which can only be used in "X" boards or in some cases even enterprise server boards. X boards can also use dual cpu Xeon in single config, as well as ECC Reg Ram on some.

I can give you a whole history lesson in HEDT, but id think id bore half the people in this chat so i'll just end it here.

 

[In2Photos]

The original acronym was High Enthusiast DeskTop.

That is why we had a Enthusiast tier which just meant a overclockable enterprise edition.
Hence allowed you to overclock server parts and run them much faster.

The first Enthusiast component we got was again the LGA1366 (Nehalem)

WOW! I'm running HEDT and had no idea! 😂

 

[Exist50]

The original acronym was High Enthusiast DeskTop.

"High End DeskTop" is how I remember it. "High enthusiast" doesn't sound right.

 

[aigomorla]

WOW! I'm running HEDT and had no idea! 😂

X58 X79 X99 X299 are all HEDT...
Z87 Z97 Z170 .... Z790 are all consumer based boards.

Another fun fact.... HEDT never had onboard Video outputs, while the consumer brand did, because HEDT never offered IGPU.
Server boards had a matrox/Rage gpu chip intigrated in them, so the video on enterprise was done at the board level, and not cpu level.
They figured your spending so much money, you can afford a SLI, and instead gave you 2 to 3 full 16x PCI-E Slots.

We lost that on most boards now only offering 1 full 16x slots, and the second slot looking like full 16x but only having 8x traces, as the CPU does not have enough PCI-E lanes to handle 32X PCI-E lanes on consumer side.

"High End DeskTop" is how I remember it. "High enthusiast" doesn't sound right.

im fairly certain it was Enthusiast, which is why we had that entire branch come out, which was basically over priced enterprise gear pretending to be consumer gear, at enterprise prices.

Like how "gamer" class is consumer class with RGB overpriced at Gamer prices because gamers will pay for RGB.

 

[Saylick]

"High End DeskTop" is how I remember it. "High enthusiast" doesn't sound right.

Ohhh, now I understand why 420mm radiators are considered HEDT... Duhhh! *slaps forehead*

The plastic tubing, glass cylinders... it all makes sense now.

 

[nicalandia]

Ohhh, now I understand why 420mm radiators are considered HEDT... Duhhh! *slaps forehead*

Radiators and Pumps keeping getting larger and larger, but what we really need is to enhance the thermal Conductivity of water with copper nano-particles.

Yes, I know how to make Water-Copper Nano Particles at home.

 

[aigomorla]

Ohhh, now I understand why 420mm radiators are considered HEDT... Duhhh! *slaps forehead*

The plastic tubing, glass cylinders... it all makes sense now.

NAH HEDT for radiators is the MORA or a 560....

Radiators and Pumps keeping getting larger and larger, but what we really need is to enhance the thermal Conductivity of water with copper nano-particles.

Yes, I know how to make Water-Copper Nano Particles at home.

The military had Liquid Gallium cooled Servers they used in the desert.
IT was something simular to this:

Designing a liquid metal cooling system loop (yup, you...

www.overclock.net

But it never took off, as i think at the price it cost, full emersion is cheaper, not to mention a lot safer as Gallium is highly conductive.

 

[Exist50]

Radiators and Pumps keeping getting larger and larger, but what we really need is to enhance the thermal Conductivity of water with copper nano-particles.

Yes, I know how to make Water-Copper Nano Particles at home.

NAH HEDT for radiators is the MORA or a 560....

I think you guys are missing the joke, lol.

 

[aigomorla]

I think you guys are missing the joke, lol.

You obviously took my cheap shot back at you in full force.

 

[nicalandia]

I think you guys are missing the joke, lol.

I have Asperger, sometime it's difficult to spot the joke.

 

[RnR_au]

I have seen at least 2 independent rumors indicating that AMD has at least some internal chips running at 6ghz single core boost. If AMD has solved the clock issues with the X3D parts (which someone at AMD hinted was related to power draw) then it is possible we could see higher clocks (5.8-6ghz).

I came across some stuff in relation to this yesterday. Apparently TSMC has been moving rapidly in this space and has a better stacking process going nowadays SoIC v4;

TSMC also showcased that contact resistance was better across the stack due to their thinner barrier layer. In addition, TSMC believes SoIC is more reliable. This includes with a wider range of operating temperatures. Many were disappointed when AMD locked down overclocking and modifying power entirely on their 5800X3D desktop chips. This is likely only a hiccup with the 1st generation. As TSMC’s Cu alloy is modified and the pitches decrease with SoIC gen 4, it seems they are improving their reliability and yields.

Image and quote snagged from - https://semianalysis.com/packaging-...-1-micron-pitch-hybrid-bonding-mediatek-netw/

So the frequency regression and no overclocking could have been temperature related due to SoIC v1. But v4 looks much better from a operational temperature range so overclocking for the hypothetical 7800X3D could be possible.

 

[Kaluan]

I have seen at least 2 independent rumors indicating that AMD has at least some internal chips running at 6ghz single core boost. If AMD has solved the clock issues with the X3D parts (which someone at AMD hinted was related to power draw) then it is possible we could see higher clocks (5.8-6ghz). Intel is also said to be launching a 6ghz SKU.

Considering AMD seens even more hush-hush than usual in regards to Raphael-X and we're not that far off from launch and we don't know anywere as much potential hard data (such as SkyJuice's early 5,85GHz fmax leak) as we did about Raphael 5 months ago, there's certainly some legroom for AMD to surprise us, beyond just what 5800X3D was to the rest of the Zen3 lineup.

I wonder if they're confident enough to push a "Dragon Range-X" SKU out next year as well. It would certainly get some headlines/buzz and grab them some mindshare on laptop.

 

[nicalandia]

Nearly 6 Ghz OC..

https://twitter.com/x/status/1577173887489433601

 

[Kocicak]

The original acronym was High Enthusiast DeskTop.

Enthusiast high on what? LSD? That would explain all that RGB bling cr*p one cannot avoid now even he tries to...

 

[naad]

I have seen at least 2 independent rumors indicating that AMD has at least some internal chips running at 6ghz single core boost. If AMD has solved the clock issues with the X3D parts (which someone at AMD hinted was related to power draw) then it is possible we could see higher clocks (5.8-6ghz). Intel is also said to be launching a 6ghz SKU.

That's the public excuse, much more likely is that 3D-cache uses HD cells and they don't really like going over 1.3V, 3Dcache and base die share the same voltage rail in Milan-X and 5800XD so the die itself isn't allowed to go above that limiting Fmax