Friday, January 30th 2026
AMD "Zen 6" CCD Packs 12 Cores, 48 MB L3 Cache
The standard CPU complex die (CCD) chiplet based on AMD's next-generation "Zen 6" microarchitecture is poised to feature a 50% increase in CPU core counts. This would be the first time that AMD has increased core-counts of its CCDs containing full-sized CPU cores capable of sustaining high clock speeds. It had developed CCDs with high core counts using compacted cores with lower clock speed limits until now. The standard "Zen 6" CCD is expected to come with a similar die-size to the current "Zen 5" CCD, with a rumored die size of 76 mm², which is slightly larger than the 71 mm² of the "Zen 5" CCD.
The "Zen 6" CCD is expected to pack 12 CPU cores, all part of a single CPU core complex (CCX). All twelve cores share a 48 MB L3 cache. Both the CPU core counts and L3 cache size hence see a 50% increase over "Zen 5." AMD is building the "Zen 6" CCD on TSMC N2 (2 nm nanosheet) process node, which evidently offers a massive increase in transistor densities over the current TSMC N4P (4 nm FinFET) on which the "Zen 5" CCD is made. With an eye on Intel's pivot towards a cache rebalancing in favor of large last-level caches on some of its premium Core Ultra 400 "Nova Lake-S" desktop processors, AMD is expected to give the "Zen 6" CCD full 3D V-Cache capability, and the company's next X3D processors could come with L3 caches of 144 MB per CCD for up to 288 MB of L3 cache on the desktop AM5 socket.
Sources:
HXL (Twitter), VideoCardz
The "Zen 6" CCD is expected to pack 12 CPU cores, all part of a single CPU core complex (CCX). All twelve cores share a 48 MB L3 cache. Both the CPU core counts and L3 cache size hence see a 50% increase over "Zen 5." AMD is building the "Zen 6" CCD on TSMC N2 (2 nm nanosheet) process node, which evidently offers a massive increase in transistor densities over the current TSMC N4P (4 nm FinFET) on which the "Zen 5" CCD is made. With an eye on Intel's pivot towards a cache rebalancing in favor of large last-level caches on some of its premium Core Ultra 400 "Nova Lake-S" desktop processors, AMD is expected to give the "Zen 6" CCD full 3D V-Cache capability, and the company's next X3D processors could come with L3 caches of 144 MB per CCD for up to 288 MB of L3 cache on the desktop AM5 socket.

99 Comments on AMD "Zen 6" CCD Packs 12 Cores, 48 MB L3 Cache
The problem isn't just the lack of PCIe channels (though that is a problem too), it's that high end motherboards have 2 PCIe slots in total; low end boards may have more PCIe but in turn lack everything else. You now have to use m.2 to PCIE adapters for add-on cards because boards have more m.2 slots than useable PCIe ones. I can sort of understand because GPUs take 4 slots on average now and the only other thing you can do is stick in m.2 ports underneath them, but come on. Or you are forced to stick to USB dongles, which don't have enough space to add, say, proper cooling for a 10GbE NIC. On top of dealing with the inefficiency of USB protocol.
Motherboards also only come with 2-4 SATA ports now, and you have to use an add-on card for more... in fact the best choice I found was to use a six sata port ASM1166 controller that's on an m.2 card using a PCIE 3.0 x2 connector. You can put it on one of the chipset slots which usually use slower speeds anyway.
We are basically developing backwards so they can give us what we used to have at a premium. I have an FM2 motherboard here with 8 SATA ports that can take 4 expansion cards. And it's a cheap mid-range board. mATX even.I'd be happy if they just had all 24 lanes on the APUs. Phoenix is gimped to only 16 usable lanes, Strix Point as well. The main PCIE slot runs at only x8 as a result. Back on Cezanne it could run at x16 and I could use a x8x4x4 bifurcation card to stick two extra m.2 drives all on the CPU and still have enough left for a potential GPU, or even a low profile 10GBE NIC.
We needed adapter boards and bifurcation to get more than basic IO, they are removing even that.Mobile chipsets come out later and have improved memory controllers, usually because they use monolithic chips and not the very very poor chiplet IOD. I've seen people push 8600Gs to 9000+.
I still like my AM4 setup and it continues to serve me well, so not willing to abandon it yet ;) .
I wonder if there is also a maturing process on the AM5 chipset going on and a new chipset with added functionallity will be developed.
And then the elephant in the room, with all the prices rinsing in this sector atm, what will it bring? I guess the second hand kidney market will see a major uplift :roll:.
4 MB L3 per core in CCD
8 MB L3 per core in 3D V-cache
That means shift from 32MB (+64MB) in Zen3, Zen4, Zen5 to 48MB (+96MB) in Zen6You need more info about how much silicon it will need.
Intel: CCD with bLLC + IOD + iGPU
AMD: CCD + IOD with iGPU
BTW ...
AMD can do IOD at older&cheaper node, even the 3D cache wont be made at the best node. Only 76mm² of CCD.
AMD removed the need of ser-des on both IOD and CCD saving some silicon area. And got lower latency as side effect.
Including the bLLC into CCD will make it big (280 mm²) and expensive
52 cores may look impresive, but ... its not 52 cores. Its (8+16)+(8+16)+4. And those 4 LPE wont do big difference.
TL;DR - 52 core Intel may or may not provide higher MT than 24core/48thread Ryzen. Have to wait for results.May be Intel should consider stop mixing cores and go for only e-cores for lower price range.
Darkmont IPC is good enough to without big cores and finish this scheduler madness.
Show me where 8 big cores of Alder Lake performance wont be enough for office and everyday notebooks.Sep 27th, 2022 - Ryzen 7 7700x
Apr 6th, 2023 - Ryzen 7 7800x3D
6 months
Aug 8th, 2024 - Ryzen 7 9700x
Nov 7th, 2024 - Ryzen 7 9800x3D
Whole year has somehow shrinked into 3 months.There is not big enough demand for them.
2 CU iGP in regular desktop Ryzen is good enough for office use and computing.
BTW .. 8 PCIe tha APU miss are used to connect their iGPU. Leaving 12 lanes for general use. 16 if you do not use chipset.
Need more connection? Buy that Promotory21 expansion card and put it into x4 slot.Having both 5800x3D and 7600x I have compared them.
Somewhere 3D cache shows
Somewhere IPC and frequency shows
Somewhere core count shows (and that will be bigger with 16 vs 6 cores)
I am leaving 5800x3D for kids and their gaming. Going for 12+ core Zen6 for my mixed rendering/gaming use case.Zen4 came with 16c cores on CCD (only for EPYCs)
Strix point came with 4 + 8c cores (for notebooks so far)
Now its time to introduce 12 cores into desktop and HEDT.The biggest benefit from AVX 10.2 is not performance or bigger instruction set.
Intel finally decided to clean up their mess. And that is good.And that is good too.I do not get it. How could be 280mm² (or is it only 114mm²) chip be less expensive than 76mm²?
Has 18A turn so good its cheaper to make than TSMC 2nm margin?Yes. We can.
Ryzen 7960/7980 + TRX50 ... 48x Gen5 + 32x Gen4
Ryzen 9960/9980 + TRX50 ... 80x Gen5 + 8x Gen4
AM5 socket does not have free pins to add more.Doing time sonsuming tasks at background? Sure.
Put everything I do to stream? Why not.
But how do you combine code developing and gaming at the same time?
Is that some cool game about coding?Are you willing to pay for it too?Manufacturer decided to make one-size-all motherboard instead of several models focused on that or that.
It would be nice to finally get something capable instead of that crippled Promotory21. At least for the high-end.There are rumors about rising up the base JEDEC speed from 4800 MT/s to 5600 MT/s. May be more.Checked nearest e-shop
535€ / 313€ - 9850x3D / 9700x
618€ / 393€ - 9900x3D / 9900x
723€ / 550€ - 9950x3D / 9950x
I wont say it will be cheap. But the 3D V-cache is what makes the price so high.
My guess ... 550€ for 12core Zen6 without 3D V-cache at release.I am afraid that AMD wont lower the prices.
Just because the price of that 52 core monster will be ... monstrous.LPDDR5x already reached 9500 MT/sIntel: 10000+ on UDIMMs or even 12000+ with CUDIMMs
AMD already can run 7200~8000 MT/s after gear shifting.
And yes. There is not big advantage of using faster modules with 3D V-cache.Promotory21 connected to CPU via Gen4 PCIe just sucks.More elephant excrements.
I also wonder how much the extra cache on the X3D parts will help. 48 MB total increase over 96 MB might fall in many workload in the diminishing return area. That will be interesting to see.
As for the RAM speed. The main thing is really how the CPU is connected to the I/O dies. The serial implementation of the infinity fabrics seems to be pretty much capped anyway. It add latency and on desktop CPU, it's not wide enough to sustain higher DDR5 ram speed anyway. It's expected/rumored that AMD will implement a new I/O die with Zen 6 including a new interconnect using fanout grid like RDNA3/Strix Halo.
Well AMD have to do it if they want to remain competitive anyway. They had the packaging advantage for few generation but Intel is now ahead. They must switch to that packaging just to remain competitive.
(and kids, this is why it's good to have a competitive market and we should never just want the same actor to win all the time).
And I disagree, saying something is for "the 1%" isn't a good defense for it either, all of the parts we're buying is for the few percent of enthusiasts anyway, I think there should be a X890E with more I/O for those who don't want to empty their wallet for an HEDT build.
Anyway, it would be great if this is true, with 8 or 10 cores being the minimum core count for Zen 6.
If not, the best we can hope for is higher speeds, only for new motherboards.
Gaming is close to non-X3D though, so if vanilla vs vanilla is comparable and also X3D vs bLLC is comparable then it's probably ideal for us consumers. Either player clearly dominating is not ideal as it pretty much halves the options and also keeps prices up for the winner, no reason to give discounts when you're dominating.
So as as tempting as it might seem (for some people) to root for Intel to knock it out of the park, I think a stalemate or a modest win for Intel is the best for us (a modest overall win is good for public perception, to make it seem like they learned the lesson with Arrow Lake).
As long as the gains are significant vs the current gens, because that's the priority for us, then having comparable performance in most scenarios is likely ideal.
Coming back to who would win, it's somewhat of a paradox but while a clear victory for Nova Lake is definitely going to win over the people who are on older platforms (LGA1700, AM4 or even older) and willing to upgrade the likely very spicy prices are surely going to turn some away... towards the competition? Because if Intel wins big they're definitely putting victory tax on those CPUs. Unless somehow they're willing to reduce those profit margins just to screw AMD.I'm not sure but those 4 LPE might only be for idle and perhaps some very light background tasks, I don't think they will contribute a single point in CB for example. They don't have cache, they're located on the I/O tile, I don't think they're included for extra compute power. So technically 52 but practically 48.Today 285K has 8P+16E -> 24t and against 9950X which has 16c+HT -> 32t it's not that far behind. With comparable frequency though.
Nova Lake doubles the core/thread count while Zen6 has a 50% increase. Even assuming shitty scaling Nova Lake appears to have an advantage. But that is with comparable frequency, if Zen6 has significantly higher frequency it can offset a lot of that core/thread advantage.9900X has full cache like 9950X. So it's possible but not likely perhaps.
I would be really surprised and think it's unlikely that the cut down cores Zen 6 will have its 48 MB cut down
See below (ignoring physical slot size)
potential 4 slot layout
- from CPU
- M.2 (standard)
- Slot 1: x16 (x16, or x8, or x4x4, or x4x4x4x4)
- Slot 2: shared x8 (disabled, or x8, or x4x4) or M.2+M.2
- Slot 3: x8 (as x8, or x4x4) or M.2+M.2
- from Chipset
- Slot 4: x4 (as x4) or M.2
potential 5 slot layout