AMD 3rd Gen Ryzen AM4 Package Capable of Two 8-core Chiplets

[btarunr]

At its CES 2019 keynote, AMD unveiled two killer client-segment products, the Radeon VII graphics card, which beats the GeForce RTX 2080; and a sneak preview of the 3rd generation Ryzen socket AM4 processor based on the company's "Zen 2" microarchitecture. As part of the unveil, CEO Lisa Su demonstrated an 8-core/16-thread 3rd generation Ryzen prototype processor in a head-to-head CineBench nT face-off with the Intel Core i9-9900K processor, which has the same core-count. The Ryzen narrowly beat the Intel flagship. Following this, Dr. Su held up a de-lidded sibling of the processor that was tested, revealing not one, but two dies.

This confirms that AMD is taking the heterogeneous multi-chip module approach to building its 3rd generation Ryzen processors, much like its 2nd generation EPYC processors that were unveiled late last year. The MCM of the processor Dr. Su held up had two chips, the smaller chip is an 8-core CPU chiplet built on the 7 nm process, that appears to have the same die-size as the 8-core chiplets that make up the 64-core 2nd gen EPYC MCMs, the larger die is an I/O controller logic built on the 14 nm process. This die controls the memory, PCIe, and SoC connectivity of the package. We noticed something curious about the way the two dies are arranged on the package substrate.





On close inspection of the substrate, we find that while the I/O controller die is somewhat centrally to the side of the package, the sole 8-core CPU chiplet is not located at a similar position (think Intel "Clarkdale" MCMs). On zooming in further, we find that just south of the 8-core CPU chiplet die, there appear to be blank bumps protruding over an area similar to that of a chiplet covered up by the outer layers of the substrate, leading us to conclude that the AM4 package is capable of three dies, an I/O controller, and two 8-core CPU chiplets. There very much will be a 16-core/32-thread Ryzen for the AM4 platform, and it's only a question of when.

The 16-core Ryzen AM4 MCM will be similar in concept to the larger 64-core SP3r2 EPYC/Threadripper MCMs: the CPU dies only pack the CPU cores and an InfinityFabric interface, while the I/O controller die is wired to multiple CPU dies, and manages the memory, PCIe, and SoC connectivity of the processor.

Interestingly, in the client-segment Intel dabbled with this concept a decade ago with "Clarkdale," which combined a 32 nm dual-core CPU die that spoke to a larger 65 nm die that controlled PCIe, memory, and an iGPU, with QPI serving as the interconnect between the two. Intel's requirements at the time were different. The company hadn't yet managed to put CPU and iGPU into a single die, and needed the iGPU to sit closer to the memory interface. The company would go onto fuse CPU and iGPU with the 32 nm "Sandy Bridge."

AMD's engineering bravado with "Matisse" also unlocks the possibility of the Ryzen "Raven Ridge" APU successor being an MCM with one 8-core chiplet, and an oversized I/O controller die that packs a "Vega" or "Navi" based iGPU, in addition to memory, PCIe, SoC, and the works. Dies on that package could be arranged differently from this.

View at TechPowerUp Main Site

 

[Gasaraki]

Nice analysis from Anandtech.

 

[Rahmat Sofyan]

yeahh same thought, it looklike what weve seen on the demo not the best next Ryzen.

perhaps there is still a chance for 16 core Ryzen 7 at final release, or maybe Ryzen with navi or vega iGPU ?

 

[HD64G]

A 2nd die below the 1st one will have either another CPU chiplet, or a GPU. That makes much easier the product differentiation that AMD needs as you have the base model and you add whatever die you want very quickly and inexpensive. It is a genious strategy for sure. And that will allow lower cost products being priced lower too. A win-win situation.

 

[Deleted member 158293]

Think I'll be able to stretch my 2700x a little while longer!

Makes financial sense to wait to release 16 core Ryzen to not cannibalize znd empty the channel of "small" 16 core versions and under Threadrippers while paying more for 7nm.

 

[Joss]

The bulk of AdoredTV predictions seem to be right.
We've seen the 8 core CPU. With another 8 core chiplet we'll have the 12 core and 16 core CPUs.
And with an iGPU we'll have the 6 core and 8 core APUs.

 

[TheLaughingMan]

The bulk of AdoredTV predictions seem to be right.
We've seen the 8 core CPU. With another 8 core chiplet we'll have the 12 core and 16 core CPUs.
And with an iGPU we'll have the 6 core and 8 core APUs.

And considering the win in Cinebench, clock speeds might be real too. If it beat 9900K at stock with an all turbo of 4.6 GHz by 16%, then either AMD somehow passed Intel in IPC or has a higher sustained all core turbo. Considering the difference though it has to either be something funny going on or a combination of both.

Some quick math shows that if the funny business is a Intel stock cooler throttling the 9900K down to 4.3 GHz or less, that margin would still put the AMD clock speed around 4.85 GHz to 5 GHz which does line up with the leaks.

 

[Zyll Goliat]

I am sure that there is a room for another chiplet there.....so we can expect 16 or 12 cores in the future.....the question is when?...maybe AMD is going to present us first only the 4c and 8c at first gen and then in a year or so go with the 12c and 16c as second gen.....

 

[moproblems99]

And considering the win in Cinebench, clock speeds might be real too. If it beat 9900K at stock with an all turbo of 4.6 GHz by 16%, then either AMD somehow passed Intel in IPC or has a higher sustained all core turbo. Considering the difference though it has to either be something funny going on or a combination of both.

Some quick math shows that if the funny business is a Intel stock cooler throttling the 9900K down to 4.3 GHz or less, that margin would still put the AMD clock speed around 4.85 GHz to 5 GHz which does line up with the leaks.

Perhaps I missed it in the original thread, but were those all core turbo speeds?

 

[Darmok N Jalad]

15% performance gain over Zen+, minimum, and the demo matched the 9900K at cinebench while consuming considerably less power. And AMD is telling us to “expect more” at launch. Color me excited!

 

[hat]

I'm really looking forward to AMD kicking Intel's ass... they're already sore from a higher number of more efficient cores, but if they can take that further while raising IPC and clockspeed above Intel, then Intel is in a no-win situation.

Hmm... I smell (another) smear campaign heading AMD's way. CTSLabs 2.0?

 

[TheLaughingMan]

Perhaps I missed it in the original thread, but were those all core turbo speeds?

No. My math is wrong anyway because I though the score was 2557 but it was actually 2057 so it equals the 9900K. So now it really does matter what cooler they were using and if the Intel chip was being gimped by over heating.

 

[InVasMani]

Actually if AMD were to shrink the I/O hub chip down to 7nm as well they could probably squeeze in a I/O alongside 3 other chiplets perhaps CPU/GPU/DRAM with the latter most connected to both of the former diagonally placed across the I/O hub, but adjacent the CPU/GPU chiplets. Basically all 4 chips could occupy it's own corner with the I/O hub and DRAM chips adjacent to both CPU/GPU chips or CPU/CPU chips if you wanted to go that route instead. I do like the idea of a fairly substantial DRAM cache built right directly to the CPU though EDRAM was pretty solid for Intel and that was quite a few generations back that it was used on the desktop and obviously capacity was a lot smaller than would could be done today.

 

[Darmok N Jalad]

No. My math is wrong anyway because I though the score was 2557 but it was actually 2057 so it equals the 9900K. So now it really does matter what cooler they were using and if the Intel chip was being gimped by over heating.

The score is very consistent with other reviews, so I don’t believe it’s gimped. Also, I think AMD more wanted to showcase that their demo Ryzen was matching the 9900K, but doing it at 75W vs 125W. It suggests they can not only match Intel in performance, but they have 50W of headroom to move past them. It’s not that far-fetched—Intel is behind on manufacturing, and they are pushing their current node to its absolute limits.

If that is true, then maybe those 5.0ghz rumors are valid. I think Zen was more architecturally limited to the low 4 GHz range than it was node limited. Zen 2 pretty much has to correct that problem so they can advance performance.

Actually if AMD were to shrink the I/O hub chip down to 7nm as well they could probably squeeze in a I/O alongside 3 other chiplets perhaps CPU/GPU/DRAM with the latter most connected to both of the former diagonally placed across the I/O hub, but adjacent the CPU/GPU chiplets. Basically all 4 chips could occupy it's own corner with the I/O hub and DRAM chips adjacent to both CPU/GPU chips or CPU/CPU chips if you wanted to go that route instead. I do like the idea of a fairly substantial DRAM cache built right directly to the CPU though EDRAM was pretty solid for Intel and that was quite a few generations back that it was used on the desktop and obviously capacity was a lot smaller than would could be done today.

Actually, if I remember right, the GPU could just be built right into the IO chip. I believe that’s what happened with the custom GPU ATI built for the Xbox360 and what Intel did with early Core MCM designs. If they shrink that IO+GPU chip to 7nm, it would probably fit the footprint of the base IO chip.

 

[Captain_Tom]

Does that "Bugs" guy still have me in his signature saying I am almost always right? I was right lol.

An honor to be so right I am in some salty Intel Fanboy's signature!

 

[cdawall]

15% performance gain over Zen+, minimum, and the demo matched the 9900K at cinebench while consuming considerably less power. And AMD is telling us to “expect more” at launch. Color me excited!

I'm waiting patiently for real benches from real reviewers. Last time I got excited for an amd release I got bulldozer.

 

[Darmok N Jalad]

I'm waiting patiently for real benches from real reviewers. Last time I got excited for an amd release I got bulldozer.

Understandable. However, we know do that this will be an improvement upon Zen, not some totally new architecture like bulldozer. Bulldozer was just a weird architecture from the start, banking too much on dual ALUs in an answer to the soon-to-fail Netburst architecture. Even if we only get either an IPC increase OR a clock speed increase, it still moves AMD in the right direction. The fun part is it sounds like we will get both!

 

[cdawall]

Understandable. However, we know do that this will be an improvement upon Zen, not some totally new architecture like bulldozer. Bulldozer was just a weird architecture from the start, banking too much on dual ALUs in an answer to the soon-to-fail Netburst architecture. Even if we only get either an IPC increase OR a clock speed increase, it still moves AMD in the right direction. The fun part is it sounds like we will get both!

Eh we will see. I just want them to fix the nuances of the 2970/2990.

 

[Darmok N Jalad]

Eh we will see. I just want them to fix the nuances of the 2970/2990.

I believe that is what the separate IO chip is to resolve so the 2 chips not tied to the memory don’t have a big latency penalty by going through the 2 chips that do get access to the memory.

I will say that I was hoping for a more of a formal lineup from AMD today. They already did a 7nm press event, so while the demo was nice, I was hoping for a few more details.

 

[Jayp]

The bulk of AdoredTV predictions seem to be right.
We've seen the 8 core CPU. With another 8 core chiplet we'll have the 12 core and 16 core CPUs.
And with an iGPU we'll have the 6 core and 8 core APUs.

Only main thing I had an issue with in the AdoredTV leak is the prices. I still don't think those will be the prices but they could be if they needed to be. The thing is with this 8 core Ryzen performing like a 9900K which I anticipated, it could easily sell for $329 ish like the current 2700X. At $329 9900K performance would be a steal given the almost $600 price tag of the 9900K.

 

[Captain_Tom]

Only main thing I had an issue with in the AdoredTV leak is the prices. I still don't think those will be the prices but they could be if they needed to be. The thing is with this 8 core Ryzen performing like a 9900K which I anticipated, it could easily sell for $329 ish like the current 2700X. At $329 9900K performance would be a steal given the almost $600 price tag of the 9900K.

$330 is what the 9900K should cost. It uses more energy than the 2700X, costs 10% more, and it's platform costs $100 more. Yet it is only around 20% more powerful.

$600 is just demonstrating how Intel can still get away with murder.

 

[Vya Domus]

Not only this was clearly designed for two dies there's also room for another one TDP wise according to those power figures.

 

[jchambers2586]

$600 is just demonstrating how Intel can still get away with murder.

They can get away with it if you're dumb enough to buy it at a exorbitant price.

 

[prtskg]

They can get away with it if we're dumb enough to buy it at a exorbitant price.

Is that fixed now?

 

[jchambers2586]

Yes but it goes to show that Intel is refusing to be price competitive. They're acting like Verizon before they started offering unlimited data plans. "We have a premium service therefore we don't need to offer Unlimited data".