AMD demonstrates Ryzen 9 5900X prototype with 3D V-Cache stack chiplet design

[gpuerrilla]

Just been busy with stuff - didn't realise a month had passed! Thanks for the mention though!

Any CPU's you think would be to keep an eye on? Currently have the 3600 but I can always use this in another family machine if I find a reasonably priced upgrade.

 

[jigger]

https://www.techpowerup.com/290135/amd-will-give-a-glimpse-of-zen4-core-at-ces-2022-presentation

So, an announcement of an announcement. Should be getting some news about Zen3D on 4th Jan.

"With regards to the upcoming generation, I point to CES in January. We couldn't be more excited to be rolling out some additional details on our new product launches. It will deliver phenomenal experiences and as we’ve said publicly, later in the year as it progresses we’ll share more detail on Zen 4," Papermaster said. "We'll talk a bit about Zen 4 at CES and then you'll see a rolling set of announcements over the course of 2022 on our next CPU family and across the board on PC, gaming, and server. 2022 is shaping up to be an incredibly exciting year for AMD."

This reads like they are reluctant to discuss details of Zen 4 for most of 2022, I suppose that is to be expected they are launching 2 generations potentially within a year of each other.

I wonder how much availability of DDR5 will be a factor in the launch date of Zen 4?

Days 1-5 IIRC are trade and press. Public are 5-8 I think.

 

[LePhuronn]

This reads like they are reluctant to discuss details of Zen 4 for most of 2022...

You do come to some crazy conclusions, it reads nothing of the sort.

 

[g67575]

You do come to some crazy conclusions, it reads nothing of the sort.

"later in the year as it progresses we’ll share more detail on Zen 4". So they will tease it a bit, but only discuss details in Q3-Q4.

I think they want people to focus on Zen3D, at least for the 1st half of 2022.

 

[Journey]

June or around Computex is the usual time when they announce these sorts of things, just depends on if they are doing a stack shuffle with Ryzen 5xxx and reducing prices on current SKU's a bit and increasing prices on the new incoming parts, so a 5800X at £249-79 and a 5800X-3DVcache at £389-419, you get the idea. Intel have don 6-9 month windows in the past, and so have AMD so it isn't unprecedented to have a short window of time, and they can also have two platforms running concurrently for serval years, given the user base that AM4 has I wouldn't be surprised to see demand for the CPU's still strong in 2023 and beyond.

 

[LePhuronn]

"later in the year as it progresses we’ll share more detail on Zen 4". So they will tease it a bit, but only discuss details in Q3-Q4.

I think they want people to focus on Zen3D, at least for the 1st half of 2022.

That's not reluctance, that's just business. AMD have gotten very good at keeping secrets, and they're not going to talk about Zen 4 too early. Intel may be locked with what Rocket Lake will be, but it's not too late to make some changes to Meteor Lake and push it back a little if they can get a decent scoop on what Zen 4 is going to achieve.

Zen 3 v cache has zero to do with it, and I don't think AMD will want too much attention drawn to Alder Lake retaining a lead in some metrics, especially at the mid and low end, when the v cache refresh comes out.

 

[g67575]

The most logical thing for Intel would be to release new Golden Cove CPUs with more cache, to ensure a decent lead over Zen3D in 2022.

You have to wonder why they didn't do this already with Alder Lake, knowing what AMD was planning. Presumably, there is a design constraint (or maybe worries about cost?). In theory, Intel could just remove the E-Cores and include more L3 cache, it seems probable that would lead to significantly more performance gains than the unexciting multithreaded gains from the E-Cores.

It looks like the 12900KS will we AMD's main competition for a while longer though.

 

[LePhuronn]

I really doubt Intel know the specific ins and outs of what AMD are working on, much like AMD won't know the ins and outs of what Intel is doing. Even if they did, every product you hold in your hands is already multiple years old. So even if Intel did know about the v cache refresh, Alder Lake was finalised at least a year ago and can't be changed. Rocket Lake is a revision of Alder Lake so that can't be changed now either.

The most logical thing for Intel would be to release new Golden Cove CPUs with more cache, to ensure a decent lead over Zen3D in 2022.

And how quickly do you expect Intel to turn that around? It wouldn't be fast enough to even matter. Besides, what you describe is Rocket Lake and is unlikely to be enough to retain a lead over Zen 4.

As for dumping the e cores and putting more P cores on, that's not sustainable. Intel are in a really bad place when it comes to CPU design. Their cores are great, but the implementation, the scalability and the power consumption is still woeful. How much power and heat would be involved in a CPU with 10 Golden Cove cores? Where do Intel go from there? They can't, which is why they're doing this hybrid approach for now.

It's literally the only thing Intel can do, they're just not as innovative as AMD these days. Let's see what Meteor Lake and Lunar Lake bring with their MCM designs.

 

[g67575]

They can make do with 8 cores I think (if it's possible to increase the cache amount), the performance of the 12900K is only slightly lower in gaming with the E-Cores disabled. But comparing 8 Golden Core Cores with 8 Zen3D cores, they will probably offer a similar level of performance in games. So, like AMD, they need find a way to implement extra CPU cache into the design to maintain a performance lead. The 12900KS is something of a pyrrhic victory (it looks a bit desperate to release it so soon, after the very similar 12900K), if its only slightly faster than Zen3D (in games), but costs about twice as much as most of AMD's lineup, and requires a mega cooler

The absolute maximum they will gain would be around a 6% performance boost vs the 12900K, assuming 300mhz higher all core clock frequencies.

 

[jigger]

They can make do with 8 cores I think (if it's possible to increase the cache amount), the performance of the 12900K is only slightly lower in gaming with the E-Cores disabled. But comparing 8 Golden Core Cores with 8 Zen3D cores, they will probably offer a similar level of performance in games. So, like AMD, they need find a way to implement extra CPU cache into the design to maintain a performance lead. The 12900KS is something of a pyrrhic victory (it looks a bit desperate to release it so soon, after the very similar 12900K), if its only slightly faster than Zen3D (in games), but costs about twice as much as most of AMD's lineup, and requires a mega cooler

The absolute maximum they will gain would be around a 6% performance boost vs the 12900K, assuming 300mhz higher all core clock frequencies.

Intel can only offer 8 cores so they will have to make do. The big cores are ring bus, decent performance, but scale poorly and have terrible on power use.
The small Intel cores are Atom with a mesh design. Scale reasonably well and offer decent power use. The downside the performance isn’t good and suffer latency issues especially at scale.

Using the ring bus + mesh allows Intel to leverage the benefits of both to an extent, but both designs also suffer issues, hence why AMD are generations ahead of Intel. In essence, Intel have taken its last two piles of garbage and whip them into one better pile of garbage.

 

[g67575]

I wonder if the 6nm APUs (RDNA2) are coming? And if so, will AMD only use 6nm EUV for the graphics chip part, but 7nm for the processor?

I find it odd that they would use 6nm EUV for APUs, but then stick with 7nm for Zen3D CPUs.

Graphics chips are the part that would have most to gain from a die shrink to 6nm EUV. There's also a refresh of RDNA 2 (I think it will be RDNA 3 instead) graphics cards that's rumoured to use 6nm EUV.

With limited 6nm EUV production capacity, they may need to allocate it entirely to GPUs (AMD is still behind Nvidia in both overall performance, and volume). I think the rumours about this have got some of the details wrong, because the official AMD GPU roadmap is saying they plan to release RDNA 3 (rather than an RDNA 2 refresh) presumably by the end of 2022:
https://images.anandtech.com/doci/15592/GPU_Roadmap_Master_678x452.png

It says 'Advanced Node' for RDNA 3, so I doubt this means 5nm.

 

[LePhuronn]

will AMD only use 6nm EUV for the graphics chip part, but 7nm for the processor?

No, APUs are monolithic. Even the desktop APUs are repurposed laptop silicon. There was speculation that the chiplet design would allow for a CPU chiplet and a GPU chiplet in a single package, but that would require a dedicated substrate and interconnect.

There's nothing odd about using 7nm for Zen 3D. The V cache refresh has zero design changes. It's the same chiplet as before, its the same interconnects and packages as before. So if everything is the same as before, my bother with the expense of shifting to 6nm for a limited run product? Just because 6nm and 7nm are design compatible doesn't mean there's no cost involved switching over.

And do yourself a favour, stop looking at roadmaps. They're not worth the pixels they're rendered with.

It says 'Advanced Node' for RDNA 3, so I doubt this means 5nm.

??? Why wouldn't it mean 5nm? 5nm is an advanced node compared to 7nm, and RDNA 2 is built on 7nm. "Advanced node" is just marketing speak for "something newer than whatever the current product is built on, but we're not saying exactly what because plans and commitments can change".

 

[g67575]

Yup, 'Advanced Node' could mean either, they are probably keeping their options open.

I think AMD might decide to prioritize their 5nm capacity for Zen 4 though, given how important the CPU side of their business has become. They could always do a 5nm refresh later for RDNA 3.

There's nothing odd about 7nm for Zen 3D, I agree. But using 6nm for the APUs seems inconsistent (although as you say, they may not have a choice, if it has to be a single die and they want to improve the power efficiency / performance of the GPU).

It will be interesting to see if they try to increase the processor clock speeds vs the last generation of APUs.

 

[CAT-THE-FIFTH]

Any CPU's you think would be to keep an eye on? Currently have the 3600 but I can always use this in another family machine if I find a reasonably priced upgrade.

What motherboard do you have?

Well I saw a Ryzen 7 5800X briefly for around £250 a few days ago at another retailer,so one of those would be a good upgrade in the same motherboard. If not a Core i5 12600KF,when the B660/H610 are released.

 

[gpuerrilla]

What motherboard do you have?

Well I saw a Ryzen 7 5800X briefly for around £250 a few days ago at another retailer,so one of those would be a good upgrade in the same motherboard. If not a Core i5 12600KF,when the B660/H610 are released.

Yes it could be intel if that package was reasonable too, the board is a B450 gaming TUF, the VRM cooling is OK but probably not best for a meaty CPU which is why I could switch out the CPU+mobo together.

 

[jigger]

Yes it could be intel if that package was reasonable too, the board is a B450 gaming TUF, the VRM cooling is OK but probably not best for a meaty CPU which is why I could switch out the CPU+mobo together.

Point a fan at the VRM. Works really well.

 

[humbug]

I really doubt Intel know the specific ins and outs of what AMD are working on, much like AMD won't know the ins and outs of what Intel is doing. Even if they did, every product you hold in your hands is already multiple years old. So even if Intel did know about the v cache refresh, Alder Lake was finalised at least a year ago and can't be changed. Rocket Lake is a revision of Alder Lake so that can't be changed now either.


And how quickly do you expect Intel to turn that around? It wouldn't be fast enough to even matter. Besides, what you describe is Rocket Lake and is unlikely to be enough to retain a lead over Zen 4.

As for dumping the e cores and putting more P cores on, that's not sustainable. Intel are in a really bad place when it comes to CPU design. Their cores are great, but the implementation, the scalability and the power consumption is still woeful. How much power and heat would be involved in a CPU with 10 Golden Cove cores? Where do Intel go from there? They can't, which is why they're doing this hybrid approach for now.

It's literally the only thing Intel can do, they're just not as innovative as AMD these days. Let's see what Meteor Lake and Lunar Lake bring with their MCM designs.

10 + 8 would be about 300 watts and only gain about +15% in MT performance, Intel are on the limits of what they can reasonably do with ADL.

 

[LePhuronn]

10 + 8 would be about 300 watts and only gain about +15% in MT performance, Intel are on the limits of what they can reasonably do with ADL.

Indeed. 8+16 (the rumoured direction) or 10+8 for Raptor Lake isn't going to see much of an uplift. Credit to Intel for those Golden Cove cores, that's an impressive generational leap, but their cores aren't the problem.

 

[humbug]

Indeed. 8+16 (the rumoured direction) or 10+8 for Raptor Lake isn't going to see much of an uplift. Credit to Intel for those Golden Cove cores, that's an impressive generational leap, but their cores aren't the problem.

Sure, well, it is and it isn't. They have about ~20% higher IPC than Zen 3, which on the face of it is; well not great but ok its good. Until you consider those cores are 2 to 3 times the size of Zen 3 cores and use 3X as much power.

What is going on here? Nvidia are also making these huge inefficient dies to keep up with AMD, that's both of them now, impressive. Indeed.

 

[LePhuronn]

OK, allow me to rephrase. Intel's designs aren't the problem, their implementation is. Rocket Lake would have been much better had it been manufactured on the 10nm process it was designed for. Alder Lake would have been much better if the 10nm process was where Intel intended it to be. But here were are yet again with Intel producing hot garbage because of their inability to make anything properly.

Raptor Lake is going to fall victim of the same issue. Arguably Meteor Lake will too. There's no guarantee their 7nm process ("Intel 4" can take a running jump) is going to suddenly fix all their issues, and if the rumoured MCM design comes to pass, we've not seen anything to suggest that Intel can produce an efficient interconnect.

Those rumours still persist of having early DG2 tiles in labs scaling almost linearly and utterly destroying the 2080Ti, but the interconnect for 4 tiles was pulling about 500W on its own.