Tuesday, October 21st 2025

AMD Could Launch 16-Core Ryzen 9 9950X3D2 CPU with 192 MB L3 Cache and 200 W TDP

AMD seems to be preparing an update to its "Granite Ridge" Ryzen 9000 X3D series, introducing two new chips targeted at high-performance gamers and creators. Among the two, a new SKU type has appeared, carrying the "X3D2" suffix. Leaks suggest a new flagship model, the Ryzen 9 9950X3D2, featuring 16 cores and 32 threads. This chip reportedly includes 3D V-Cache on both chiplets, totaling approximately 192 MB of L3 cache. It is said to operate at a base frequency of 4.30 GHz, with a boost up to around 5.6 GHz, and has a TDP of about 200 W. The 9950X3D2, with its new naming convention, suggests a tradeoff of slightly lower peak boost speeds (-100 MHz) compared to the single-stack X3D Ryzen 9 9950X3D model, but offers significantly increased cache capacity. This combination could be a sweet spot for cache-sensitive workloads and certain games.

The expanded cache and increased thermal headroom may necessitate more attention to cooling and VRM design for system builders, potentially leading AMD to price these chips at a premium or adjust current MSRPs. A companion model, the Ryzen 7 9850X3D, is expected to follow a similar design with eight cores, 96 MB of L3 cache, and a boost up to 5.6 GHz, with a 120 W power envelope. The 9850X3D appears to be a refined, higher-clocked (+400 MHz) alternative to existing 8-core Ryzen 7 9800X3D models, offering improved single-thread performance that is highly valued by gamers, while maintaining a relatively modest TDP.
Both processors are linked to AMD's next-generation V-Cache technology, which reportedly offers better thermal performance and is more conducive to overclocking. These advancements could explain the feasibility of dual-cache retail SKUs. However, these details are still unconfirmed, so they should be considered preliminary. If accurate, AMD's strategy would maintain competitive pressure on Intel's upcoming desktop refresh by emphasizing cache-driven performance rather than solely increasing core counts. Intel's "Arrow Lake Refresh" is on the horizon, but it lacks a 3D V-Cache equivalent until "Nova Lake."
Sources: chi11eddog on X, via VideoCardz
Add your own comment

93 Comments on AMD Could Launch 16-Core Ryzen 9 9950X3D2 CPU with 192 MB L3 Cache and 200 W TDP

#51
rv8000
Dr. DroAMD didn't release this before simply because they didn't want to. No binning required, they could build this as far back as Zen 3 without any additional effort.
They didn’t release it because it was a pointless product (and still is). I guess they figure there are enough fools out there to part with money for ultra niche performance increases in use cases no one will ever touch in the consumer space.
Posted on Reply
#52
Fatloader
GeofrancisA standard 7950x is 230w
You mean it actually uses at 230w on average, right? TPUs chart says it's TDP is 120w. I wasn't talking about power average.
Posted on Reply
#53
Dr. Dro
rv8000They didn’t release it because it was a pointless product (and still is). I guess they figure there are enough fools out there to part with money for ultra niche performance increases in use cases no one will ever touch in the consumer space.
Ah, yes. Who needs more than 640 KB of RAM? 640K ought to be enough for anybody.
Posted on Reply
#54
rv8000
Dr. DroAh, yes. Who needs more than 640 KB of RAM? 640K ought to be enough for anybody.
So no actual point or rebuttal per usual? Right…

Let me know when more than 1% of use cases would actually end up faster on a dual vcache zen5 part vs a 9950x3d if it’s ever released.
Posted on Reply
#55
user556
ir_cowI'll wait for Zen6 at this point. Hopefully th IPC is 20% more and a X3D2 in its lineup.
Not expected until Zen 7. Zen 6 will be all about more cores, more cache and a faster I/O die.
Posted on Reply
#56
Dr. Dro
rv8000So no actual point or rebuttal per usual? Right…

Let me know when more than 1% of use cases would actually end up faster on a dual vcache zen5 part vs a 9950x3d if it’s ever released.
What is even there to address when you made such a short-sighted, self-centered statement to begin with? People's needs and wants vary, as does their purchasing power. This is why you have a whole stack of products to begin with. There are many workloads that would benefit from it, and it would address the topology shortcomings of the current split design, but not only has it been brought up in this thread already, but discussing that would be of little value to the matter at hand. Just because you don't need one, it doesn't mean that nobody else needs or wants one.

End of the day, the lesson to learn here is that if there is a market, supply that demand, lest your competition does.
Posted on Reply
#57
freeagent
My 160w stock 9900X can do 290w with PBO with brief excursions into the 300w+ zone with PBO. It is just 12 cores.. Ryzen loves power. It thrives on it.
Posted on Reply
#58
user556
RulumiRegarding the discussion here, AMD could have just released dual CCD 3D V-Cache processors before, they just didn't want to. They already have multiple 3D V-Cache CCDs EPYC products, they already shown dual CCDs consumer CPUs prototypes and the lower clocking problem wasn't a problem for AMD to release Zen 3 and Zen 4 X3D products.

AMD probably just did out of profit driven decissions, ...
I'd say it more likely to be a packaging bottleneck. TSMC, and the whole industry more recently with AI demands, have had to transition to complex packaging for consumer grade where it was previously the domain of supercomputers alone.
Posted on Reply
#59
Dr. Dro
freeagentMy 160w stock 9900X can do 290w with PBO with brief excursions into the 300w+ zone with PBO. It is just 12 cores.. Ryzen loves power. It thrives on it.
That goes for pretty much every processor with high clock speeds and a lot of cores, though. That's why server chips usually come in the 2GHz range, at that level most modern cores can run within a watt or two. The v/f curve past a point (usually beyond 5 GHz in the latest architectures) goes up like a rocket, eventually you'll end up with like 20 W+ per core increases for 100 MHz.
user556I'd say it more likely to be a packaging bottleneck. TSMC, and the whole industry more recently with AI demands, have had to transition to complex packaging for consumer grade where it was previously the domain of supercomputers alone.
This makes no sense whatsoever, especially since they already mass produce the substrate for the X3D silicon and have for years.
Posted on Reply
#60
cfenton
It's way too late in the generation to buy Zen 5 unless these are released at the same price as the existing lineup. Zen 5 was pretty disappointing in general, so hopefully Zen 6 can bring some bigger improvements.
Posted on Reply
#61
TPUnique
BlosteEach chiplet will have their own 3D cache, so you don't have 16 cores with 192MB of L3 cache, you'll have 8 cores with 96MB and another 8 cores with 96MB.
Didn't AMD park for gaming one of the chiplets in their Ryzen 9 chips, precisely for this reason?
Yeah IIRC the interconnect between the chiplets is dated and incurs a non-trivial latency penalty. I think it's set to be updated with Zen 6.
Posted on Reply
#62
Dr. Dro
TPUniqueYeah IIRC the interconnect between the chiplets is dated and incurs a non-trivial latency penalty. I think it's set to be updated with Zen 6.
The interconnect is not the problem, the topology is. A dual X3D CPU will behave exactly like a standard non-X3D Ryzen 9 will. The 9900X underperforms because of its 6+6 design that is unable to fit tasks that commonly requires 8 cores locally, the 9950X faces no such issues. Similarly, the split 3D + standard models have one side with many resources, and one side with much fewer resources. Managing latency on this design boils down to where such processing resources are physically located on the processor, and scheduling a task to be executed with locality in mind, especially on the current R9 X3D models. Keeping a <8C task within a CCD is an optimization that has been around for more than half a decade, no special scheduling settings should be required.

Similar phenomenon occurs in a much smaller scale here compared to how core to core latency works on servers, but the concept is the same. Good read:

chipsandcheese.com/p/core-to-core-latency-data-on-large-systems
Posted on Reply
#63
LabRat 891
AVATARATThis isn't serious, this milking rumor again...

Don't care. Entertaining as real. :p


We
want this product. It can exist, it's not at all like willing another Navi4x (or GDDR7 Navi 48) into existence.
The components that make up these rumored CPUs, already factually exist.

IMHO, the "rumor milking" is appropriate. AMD needs to know the market craves this, even if AMD knows (better than anyone) that we're stupid for it. :laugh:
TBF, Salivating over the lab-only 5950X3D was probably stupid but, we still want dual vCache dual CCD MCM SoCs in the consumer-enthusiast marketspace.

For example,
the attention Silverstone got for their "April Fool's Day Prank" of a case, alerted them of the demand.
It now exists, and has a 'sequel' model en route to retailers, today.
www.silverstonetek.com/en/product/info/computer-chassis/flp01/
www.silverstonetek.com/en/product/info/computer-chassis/flp02/
Posted on Reply
#64
evernessince
Dr. DroThe interconnect is not the problem, the topology is. A dual X3D CPU will behave exactly like a standard non-X3D Ryzen 9 will. The 9900X underperforms because of its 6+6 design that is unable to fit tasks that commonly requires 8 cores locally, the 9950X faces no such issues. Similarly, the split 3D + standard models have one side with many resources, and one side with much fewer resources. Managing latency on this design boils down to where such processing resources are physically located on the processor, and scheduling a task to be executed with locality in mind, especially on the current R9 X3D models. Keeping a <8C task within a CCD is an optimization that has been around for more than half a decade, no special scheduling settings should be required.

Similar phenomenon occurs in a much smaller scale here compared to how core to core latency works on servers, but the concept is the same. Good read:

chipsandcheese.com/p/core-to-core-latency-data-on-large-systems
Both are an issue but I will explain why the interconnect is. AMD uses SerDes to connect it's consumer chiplets, which has latency and power overhead. AMD is planning on shifting over to a "sea of wires" using TSMC's InFO-oS technology.

The new interconnect is higher bandwidth, lower latency, more flexible, frees up valuable die space currently occupied by the SerDes, and can be up to 10 times more efficient to transfer data.

www.techpowerup.com/341445/amd-d2d-interconnect-in-zen-6-gets-sea-of-wires-upgrade

With the new design, no longer is there a need to convert the format of the data in order to transmit it. It is now transferred similar to as if it was one chip. It will be interesting to see the impact this has not just on the CPU dies but the IO die as well. Intel added two cores to it's IO die to improve latency and efficiency but with this tech that seems entirely pointless outside of ultra low power applications (less than 5w).
Posted on Reply
#65
jnv11
If this CPU releases, we better have Battlefield 6 benchmarks with it. I have read that this game is known to keep more than 8 physical cores busy if they exist like on the 9950X3D. We would then see if the game is able to take advantage of the V-Cache on both CCDs.
Posted on Reply
#66
user556
Dr. DroThis makes no sense whatsoever, especially since they already mass produce the substrate for the X3D silicon and have for years.
And in all that time they've struggled to provide it timely. Always limiting volumes in some way or another. I speculating that's always been due to limited production of the speciality packaging facilities that had never been intended for consumer levels of mass production.

ie: TSMC have had to ramp up those facilities and can now offer more consumer priced production as a result.
Posted on Reply
#67
Dr. Dro
user556And in all that time they've struggled to provide it timely. Always limiting volumes in some way or another. I speculating that's always been due to limited production of the speciality packaging facilities that had never been intended for consumer levels of mass production.

ie: TSMC have had to ramp up those facilities and can now offer more consumer priced production as a result.
Market strategy more than anything, IMHO. Let supply get too plentiful and ASP goes down. Bad for shareholders.
evernessinceBoth are an issue but I will explain why the interconnect is. AMD uses SerDes to connect it's consumer chiplets, which has latency and power overhead. AMD is planning on shifting over to a "sea of wires" using TSMC's InFO-oS technology.

The new interconnect is higher bandwidth, lower latency, more flexible, frees up valuable die space currently occupied by the SerDes, and can be up to 10 times more efficient to transfer data.

www.techpowerup.com/341445/amd-d2d-interconnect-in-zen-6-gets-sea-of-wires-upgrade

With the new design, no longer is there a need to convert the format of the data in order to transmit it. It is now transferred similar to as if it was one chip. It will be interesting to see the impact this has not just on the CPU dies but the IO die as well. Intel added two cores to it's IO die to improve latency and efficiency but with this tech that seems entirely pointless outside of ultra low power applications (less than 5w).
While this is true, it isn't an outright block to such a chip being viable, it just has a performance overhead. Worst case scenario its exactly the same thing we see with the 9950X today anyways, and that's my point. We'll see if they'll finally stop stalling once there are no "excuses" left.
Posted on Reply
#69
AVATARAT
LabRat 891We want this product. It can exist, it's not at all like willing another Navi4x (or GDDR7 Navi 48) into existence.
The components that make up these rumored CPUs, already factually exist.
If you wanted something that would be slower than the current 9950x3D in games, fine, but I don't think so.
Posted on Reply
#70
Vayra86
ymdhisDear AMD: please put x3d cache on the APUs, specifically for use by the iGPU. How sweet would that be.
Dunno? Its not really a lot of capacity in terms of graphics.
Dr. DroThis makes no sense whatsoever, especially since they already mass produce the substrate for the X3D silicon and have for years.
In other news

www.techpowerup.com/342077/nvidias-arizona-made-blackwell-gpus-must-travel-back-to-taiwan-for-packaging

Capacity is not a matter of whether you can or cannot do it, its whether you have the hands to do it.
LabRat 891For example,
the attention Silverstone got for their "April Fool's Day Prank" of a case, alerted them of the demand.
It now exists, and has a 'sequel' model en route to retailers, today.
www.silverstonetek.com/en/product/info/computer-chassis/flp01/
www.silverstonetek.com/en/product/info/computer-chassis/flp02/
Holy shit I need this for the new HTPC build downstairs. This is hilarious
Posted on Reply
#72
Vayra86
Dr. DroYes, but that's because there's no packaging facilities in the United States, at least not yet.

They rushed to satisfy the government's demand of chips diffused in the USA, but the part of making them usable in American soil as well was conspicuously absent.
So.. there are more fabs where production of chips happens... and not more fabs where packaging happens, right? It doesn't matter its not in the US - they dó have a fab cranking out units.
Posted on Reply
#73
Geofrancis
FatloaderYou mean it actually uses at 230w on average, right? TPUs chart says it's TDP is 120w. I wasn't talking about power average.
Yes it can use 230W all day as long as you can keep the temperature under 95c
Posted on Reply
#74
Dr. Dro
Vayra86So.. there are more fabs where production of chips happens... and not more fabs where packaging happens, right? It doesn't matter its not in the US - they dó have a fab cranking out units.
I don't think TSMC Arizona is currently linked to AMD's production chain, but I won't claim to know the specifics
Posted on Reply
#75
harm9963
Dr. DroI don't think TSMC Arizona is currently linked to AMD's production chain, but I won't claim to know the specifics
Dr. Su said that the company expects to start receiving silicon from TSMC Arizona by the end of the year, making it one of the first tech giants after Apple to receive ‘made in America’ chips. AMD’s bigger rival, Nvidia, is also sourcing some of its Blackwell system production from the site. All these developments are a win for the U.S., as its push to bring back domestic semiconductor manufacturing is beginning to bear fruit, despite the slightly higher costs.


Tom's Hardware
Posted on Reply
Add your own comment
Nov 21st, 2026 07:12 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts