Wednesday, September 3rd 2025

AMD "Zen 6" Processors to Use TSMC 2 nm Node for CCDs, 3 nm for IOD

AMD will transition its next-generation "Zen 6" desktop Ryzen and server EPYC processors to a split-node strategy, placing CPU core complex die (CCD) on TSMC N2P 2 nm node variant and building the I/O die (IOD) on the N3P 3 nm node. At TSMC, volume ramp for the 2 nm node is around the third quarter of 2026, opening the door for limited shipments as early as late Q3 or broader availability in Q4 2026. Under the split approach, the compute chiplet will house the Zen 6 cores and a substantially larger shared L3 pool, while the IOD will continue to host memory controllers, PCIe lanes, USB, and the integrated GPU. Earlier rumors suggest that each CCD could contain 12 Zen 6 cores with support for SMT, and a shared L3 cache that may increase to roughly 48 MB per CCD compared with prior generations. Typical consumer configurations would pair up to two CCDs with a single IOD, creating chips with up to 24 cores and 48 threads, while server variants could scale further depending on packaging.

Performance expectations center on achieving double-digit IPC gains, higher sustained clock speeds, and improved power efficiency, thanks to the advanced process nodes. The choice of N3P for IOD helps AMD cut some costs for non-core logic, while the N2P compute die aims to extract greater generational gains. However, as with any new TSMC node, it will carry a premium over the N3P, so using it only for the compute side is the most viable choice. Platform compatibility remains for one more generation, with Zen 6 expected to support the existing AM5 socket. Users can upgrade CPUs without a full platform replacement. If the ramp timing holds, Zen 6 will arrive into a crowded second half of 2026 desktop window. For datacenter applications, the EPYC "Venice" generation, based on Zen 6, will utilize sIOD (server I/O die), implementing PCIe Gen 6, which provides a generational doubling in bandwidth to GPUs, SSDs, and NICs. AMD is also claiming memory bandwidth of up to 1.6 TB/s for these CPUs, so we could see some of that trickling down into consumer configurations.
Source: Kepler_L2
Add your own comment

71 Comments on AMD "Zen 6" Processors to Use TSMC 2 nm Node for CCDs, 3 nm for IOD

#26
Jism
user556Zen 2 was first generation to use a separate I/O die.
And for obvious reasons....

Defective rates would go down significant. Where CPU's with a defective or half working IO die would mean you throw the complete CPU away. With a separate IO die you now only throw away the IO die and not the whole chip (CPU dies).

This logic applies to the top tier Epyc's all the way down to the lowest grade CPU's that AMD has to offer. And it's damn smart. Nothing gets lost.
Posted on Reply
#27
TheinsanegamerN
Between this and Nova Lake's 16 P core rumors, also on a mix of TSMC 2/3nm, 2026 is shaping up to be a very interesting year.
Posted on Reply
#28
AnotherReader
N/AIf you look closely at the Lunar Lake N3 and Arrow lake N6 IO tile, you'll see that the memory controller is scaling down on the N3.
The memory controller will scale as it's digital logic, but the PHYs connecting the SOC to the DRAM bus are unlikely to.
Posted on Reply
#29
sephiroth117
QuicksNo worries I am not in a hurry. Seems AMD is also not in a hurry to make money then. 2029 / 2030 is perfectly fine. I am not a sucker for 10% -15% cpu gains anyways.
10-15% if you are upgrading from a 9800x3d for instance, sure.

But if you are on a 5800x3d or 7800x3d even it’s already a bigger jump. AMD is not really making money from yearly upgraders.

Only a small part upgrade each year, AMD is mainly aiming to push intel users or am4 users to upgrade here (or maybe even early am5 adopters)
Posted on Reply
#30
THU31
The CPUs will obviously be amazing, but I'd like to know something about the chipsets. I really want them to expand the CPU-chipset link and PCI-E lanes. 8 GB/s is very limiting if you have a second GPU (only high-end boards have two x16 slots connected to the CPU), capture card and multiple NVMe drives.
Posted on Reply
#31
Tek-Check
sephiroth11710-15% if you are upgrading from a 9800x3d for instance, sure.
'10800X3D', or whatever they call it, will have 12 big cores on one CCD and on N2 node, so productivity workloads boost will be significantly bigger than 15%. Current 9900X is already 18-20% faster than 9800X3D in applications. On Zen6, alongside 4 extra cores there will be IPC uplift and clocks uplift.

9800X3D is ~20% faster in applications than 7800X3D, with the same core count. So, next gen '10800X3D' will make even wider gap to 9800X3D. I would not be surprised if it's 40% faster in applications.
Posted on Reply
#32
InVasMani
[CCD1][CCD2]
\I/
/F\
[IOD1][CCD3]
Posted on Reply
#33
Tek-Check
THU31The CPUs will obviously be amazing, but I'd like to know something about the chipsets. I really want them to expand the CPU-chipset link and PCI-E lanes. 8 GB/s is very limiting if you have a second GPU (only high-end boards have two x16 slots connected to the CPU), capture card and multiple NVMe drives.
CPU-chipset link will most likely stay the same for backwards compatibility reasons, unless they introduce a new chipset, which is unlikely. Howver, they might increase PCIe lanes by x4 and perhaps integrate USB4 on IOD, which would free up another x4 link, so two extra NVMe drives directly attached to CPU. Perfectly enough for 99% of users.

Two Gen5 x16 slots working in x8/x8 mode is a high-end feature on desktop and it will most likely stay as such. Capture cards usually have up to x4 link. Almost every board has x1 and/or x4 slot for such card.
InVasMani
[CCD1][CCD2]
\I/
/F\
[IOD1][CCD3]
There is no space on AM5 package for the third CCD. This is not going to happen.
Posted on Reply
#34
Geofrancis
Tek-CheckCPU-chipset link will most likely stay the same for backwards compatibility reasons, unless they introduce a new chipset, which is unlikely. Howver, they might increase PCIe lanes by x4 and perhaps integrate USB4 on IOD, which would free up another x4 link, so two extra NVMe drives directly attached to CPU. Perfectly enough for 99% of users.

Two Gen5 x16 slots working in x8/x8 mode is a high-end feature on desktop and it will most likely stay as such. Capture cards usually have up to x4 link. Almost every board has x1 and/or x4 slot for such card.

There is no space on AM5 package for the third CCD. This is not going to happen.
In theory they could do what they done with x570 and just use an io die for the chipset. that would give 28 lanes rather than 12.
Posted on Reply
#35
InVasMani
Tek-CheckCPU-chipset link will most likely stay the same for backwards compatibility reasons, unless they introduce a new chipset, which is unlikely. Howver, they might increase PCIe lanes by x4 and perhaps integrate USB4 on IOD, which would free up another x4 link, so two extra NVMe drives directly attached to CPU. Perfectly enough for 99% of users.

Two Gen5 x16 slots working in x8/x8 mode is a high-end feature on desktop and it will most likely stay as such. Capture cards usually have up to x4 link. Almost every board has x1 and/or x4 slot for such card.

There is no space on AM5 package for the third CCD. This is not going to happen.
Backside power plus node parity frees up substrate space.
Posted on Reply
#36
trparky
I just hope that Zen 6 will be less finicky when it comes to training of system RAM.
Posted on Reply
#37
ymdhis
PrEziWhy are you spreading FUD.
The MC has been fine for two-bank modules since years. With new CPU generation DDR5 6400 is quite normally reachable (with 1:1 ratio! Where intel sometimes needs 1:2 already). Yes AMD is more memory sensitive but it has been of a lesser impact with the 3D Cache versions lately.
Where they are still stuggling is a full 4 2-sided Module area, e.g. with my 4x32GB Modules I am reaching now 5600 with tuned timings and 1:1 ratio tops.... Some are reaching 6000 but it is a memory/mobo/cpu lottery atm.
This should improve with the new IO-Die. Of course 2-bank scenarios would also improve too, but they are not of a problem atm.
Why are you saying I'm spreading FUD, when you just backed up my claims?
JismAnd for obvious reasons....

Defective rates would go down significant. Where CPU's with a defective or half working IO die would mean you throw the complete CPU away. With a separate IO die you now only throw away the IO die and not the whole chip (CPU dies).

This logic applies to the top tier Epyc's all the way down to the lowest grade CPU's that AMD has to offer. And it's damn smart. Nothing gets lost.
Don't forget they were still locked into contracts with GlobalFoundries so splitting the IO die meant they could fab them at GF with their crap old nodes, while using TSMC (or was it Samsung, I forgot) cutting edge node for the chiplets. So they got the benefit of using fast nodes while still fulfilling their contractual obligations. It was a masterstroke.
Posted on Reply
#38
InVasMani
Nice point on the GF situation AMD made the best out of their obligations overall. It wasn't ideal, but also they were still locked into contract and node costs was another factor beyond that point they took into account.
Posted on Reply
#39
phints
TSMC 2N/3N combo would be a big improvement. Just when I couldn't imagine what they would do to improve the 9800X3D now there is a strong possibility of a big jump once again. Low load/Idle power consumption should finally drop back down hopefully closer to Intel too.
Posted on Reply
#40
TheinsanegamerN
trparkyI just hope that Zen 6 will be less finicky when it comes to training of system RAM.
I could understand it needing to do the training once on first boot when system RAM changes, but why does it have to keep doing it?
Posted on Reply
#41
trparky
TheinsanegamerNI could understand it needing to do the training once on first boot when system RAM changes, but why does it have to keep doing it?
That's my thoughts too. Once the memory has been trained and you have Memory Context Restore enabled, it should be a set-it and forget it kind of thing. However, occasionally I do have a situation where the memory needs to be retrained and I can tell that it's doing it. Sometimes on my brother ASUS board, it will dump right to BIOS and tell me that it needs to reset BIOS settings and if you say no, good luck. You'll need to go find a jumper to manually reset CMOS.

My hopes are that with Zen 6 this will be solved once and for always with a better memory controller.
Posted on Reply
#42
RandallFlagg
_roman_Is there a rumour tag in the topic?

Very small at the end


Speculation: Somewhere and somehow there will be sometimes in the future another AM5 processor. It may work with current AM5 mainboards. Therefore you will not need to buy another motherboard. If the better process node will increase performance we will see. ....

Are there any facts in the topic?



There are no facts.

When a future processor has the same socket & you have firmware which supports your processor it will work.

The firmware is the uefi, sometimes named bios, update of your mainboard which also includes the agesa version for your processor and other stuff.

It's best to look out for the AGESA version.

two spreadsheets to keep around as a amd motherboard user

docs.google.com/spreadsheets/d/12zg6yT_H7H-W1voyw1ZoIrj0GSE7WI4Ug-uLlv-Asa8/edit?gid=858530609#gid=858530609

docs.google.com/spreadsheets/d/1NQHkDEcgDPm34Mns3C93K6SJoBnua-x9O-y_6hv8sPs/edit?gid=446857487#gid=446857487
I love how some guy posts something on the forum of a now dead tech site, and it turns into "News".

Sad thing is, this happens regularly. Very regularly.
Posted on Reply
#44
LittleBro
About damn time AMD revamped IOD with brand new shiny memory controller.

My 9800X3D can't do 3200 MHz on stock settings, it's causing memory errors. 3100 MHz works fine.

AMD needs to get their shit together amd introduce IOD with some modern capabilities, reaching memory speeds of 10k MT/s and higher.
Posted on Reply
#45
Craptacular
"Performance expectations center on achieving double-digit IPC gains,"

Here's hoping for 99%!
Posted on Reply
#46
_roman_
THU31The CPUs will obviously be amazing, but I'd like to know something about the chipsets. I really want them to expand the CPU-chipset link and PCI-E lanes. 8 GB/s is very limiting if you have a second GPU (only high-end boards have two x16 slots connected to the CPU), capture card and multiple NVMe drives.
That use case is a threadripper use case. HEDT High END Desktop T?

A cheap entry consumer platform needs to be segmented.
trparkySometimes on my brother ASUS board
I would differentiate between Motherboard issues and platform issues.
I should sell soon my am5 asus motherboard. Asus can not do it right with the asus prime x670-p motherboard

Not well DRAM will need certain times a CMOS reset. I did run for a while no expo, than expo. I would differentiate between decent Hynix-A Die EXPO Profile which has the supported AMD MT/s rating. Not sure if it is 5200MT/s or 5600 MT/s rating for the dram. Check the cpu product page before buying motherboard, cpu and dram when you do not want to deal with cmos reset.
Posted on Reply
#47
Tek-Check
GeofrancisIn theory they could do what they done with x570 and just use an io die for the chipset. that would give 28 lanes rather than 12.
That's not going to happen.
Posted on Reply
#48
THU31
_roman_That use case is a threadripper use case. HEDT High END Desktop T?

A cheap entry consumer platform needs to be segmented.
Great idea, get a ridiculously expensive platform with a slower CPU (no X3D) because adding four extra lanes for the chipset link would be too much? If I was willing to spend that much, I would just get a board with x8/x8 slots.

Intel has had a 16 GB/s chipset link since 2021 (with a ton of PCI-E lanes in the chipset), so it's time for AMD to catch up.
Posted on Reply
#49
TPUnique
QuicksNo worries I am not in a hurry. Seems AMD is also not in a hurry to make money then. 2029 / 2030 is perfectly fine. I am not a sucker for 10% -15% cpu gains anyways.
They will make money hand over fist, with or without your meager contribution. And jumping from Zen 5 to Zen 7 could net more than 20% gains, which is nothing to sneer at.
Posted on Reply
#50
Quicks
TPUniqueThey will make money hand over fist, with or without your meager contribution. And jumping from Zen 5 to Zen 7 could net more than 20% gains, which is nothing to sneer at.
Meh, 10% here to 20% there is boring...

If it's not 60%+ you probably won't even notice the difference. You throwing money away then.
Posted on Reply
Add your own comment
Nov 19th, 2026 03:27 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts