Tuesday, June 30th 2026

AMD Versal Premium Gen 2 Memory on Package SoCs Integrates up to 32 GB of LPDDR5X

AMD today announced the AMD Versal Premium Gen 2 Memory on Package (MoP) adaptive system-on-chips (SoCs). The MoP architecture integrates up to 32 GB of LPDDR5X into a single package, delivering up to 288 GB/s of bandwidth in up to 60% less board area, letting engineers build high-bandwidth systems without the risk and time of board-level memory design. As physical AI, networking, and aerospace and defense workloads push more data through ever-tighter space and power budgets, MoP targets the designs that need it most: test and measurement, professional video editing, and VPX systems for secure communications and defense acceleration.

"For years, system architects had to choose between the memory bandwidth they wanted and the space, power and longevity their programs could actually live with. Memory on Package removes that tradeoff. Our customers can design for the system they want to build, not the one their memory constraints allow, and bring it to market faster."—Sumit Shah, head of product management and marketing, Adaptive and Embedded Computing Group, AMD
Shrink the Footprint and Scale the Bandwidth
With our new memory on package adaptive SoC, we are redefining what is possible in compact system designs. By integrating LPDDR5X directly into the package, the device enables higher performance compared to onboard LPDDR5X while using less area than discrete implementations.

This opens the door to form factors that have been difficult or impractical with external memory, such as the Enterprise and Datacenter Standard Form Factor (EDSFF) and 3U VPX systems, while also helping designers meet telecom and VPX requirements that discrete memory approaches often cannot match.

Versal Premium Gen 2 MoP devices integrate CXL 3.1 and PCIe 6.0 at 64 Gb/s in hard IP, enabling high-speed data movement when paired with AMD EPYC CPUs to accelerate data-intensive applications. We help system architects gain greater flexibility to scale memory resources with LPDDR5X support of up to 9,000 Mb/s and connectivity to CXL memory pooling and expansion modules.
Built for Long Life Cycle Deployments

Designed with demanding physical and enterprise AI environments in mind, Versal Premium Gen 2 MoP adaptive SoCs support industrial-grade operation from -40 degrees Celsius to 110 C. They are well suited for always-on, mission-critical systems where performance and resilience must go hand in hand.

With LPDDR5X and 15-plus-year life cycle support, Versal Premium Gen 2 MoP devices help decouple product availability from high-bandwidth memory's (HBM) shorter, data-center-driven refresh cycles, reducing the risk of forced redesigns caused by memory end of life or limited accessibility.

PCIe Integrity and Data Encryption (IDE), a feature introduced in PCIe 6.0, helps protect against physical attacks by securing data in flight at the link layer. DDR memory encryption in integrated controllers helps protect data at rest without consuming programmable logic resources. Hard 400G High-Speed Crypto Engines enable high-bandwidth secure processing, strengthening security without sacrificing throughput.

Accelerate Time-to-Market
Versal Premium Gen 2 MoP devices include a pre-validated, in-package LPDDR5X interface that eliminates high-speed memory routing across the circuit board to reduce board-level simulation and validation while helping shorten development cycles, lower design risk and minimize costly re-spins.

Users can begin development today with standard AMD Versal Premium Series Gen 2 devices, which are now shipping. Support for established AMD Vivado and Vitis tool workflows, compatible IP and available reference designs helps quickly move designs from concept to deployment while enabling existing customers to adopt new MoP devices with no rework or relearning.

AMD Versal Premium Gen 2 MoP devices begin sampling at the end of 2026 with production shipments expected to begin in the second half of next year.
Source: AMD blog
Add your own comment

28 Comments on AMD Versal Premium Gen 2 Memory on Package SoCs Integrates up to 32 GB of LPDDR5X

#1
kondamin
Just offer those for laptops and sff, there is a demand for it
Posted on Reply
#3
Nomad76
News Editor
Grimlingis this ECC memory?
Posted on Reply
#4
Wirko
Grimlingis this ECC memory?
Yes, it's inline (in-band) ECC. There are no dedicated chips and bits on the data bus. Instead, a fraction of those 32 GB are reserved for ECC data.
Nomad76
What does "hard" mean here? It may be that it's always on. Meanwhile, in GDDR on GPUs, the user can choose to turn ECC on or off, so it could be called "soft".
Posted on Reply
#5
bonehead123
Nomad76letting engineers build high-bandwidth systems
Yea, now if only it would let them build systems that don't cost an arm, a leg, akidney & a left testicle/mammory, that would be great too...

But hey, one can dream, yes ? :D
Posted on Reply
#6
not_my_real_name
WirkoYes, it's inline (in-band) ECC. There are no dedicated chips and bits on the data bus. Instead, a fraction of those 32 GB are reserved for ECC data.


What does "hard" mean here? It may be that it's always on. Meanwhile, in GDDR on GPUs, the user can choose to turn ECC on or off, so it could be called "soft".
It is an FPGA, so the module could be done on programmable gates instead, the "soft" way
Posted on Reply
#7
zo0lykas
Grimlingis this ECC memory?
You can't read the post, but somehow you will read the comments with your answer 😄🤔
Posted on Reply
#8
Grimling
zo0lykasYou can't read the post, but somehow you will read the comments with your answer 😄🤔
The actual text of the article does say nothing about ECC. I know it, because i have read the text.
Posted on Reply
#9
Fluke
Sounds like Bolt has had the same idea:
Posted on Reply
#10
L'Eliminateur
WirkoYes, it's inline (in-band) ECC. There are no dedicated chips and bits on the data bus. Instead, a fraction of those 32 GB are reserved for ECC data.


What does "hard" mean here? It may be that it's always on. Meanwhile, in GDDR on GPUs, the user can choose to turn ECC on or off, so it could be called "soft".
Hard means it's a fixed function block that does not require you to waste cells of the FPGA to implement them, very popular on super high end devices, example are:
ARM cores, memory controllers, serdes blocks, pcie root complexes, ethernet controllers, even dsp blocks
Posted on Reply
#11
Athena
one has to wonder if going this route for desktop would be better?

think about it, if the RAM is integrated with the northbridge, and if they can price it right, that cuts outs the major memory makers

This would suck for O/C'ers, but, for pretty much everyone else, it's a win-win
Posted on Reply
#12
L'Eliminateur
Athenaone has to wonder if going this route for desktop would be better?

think about it, if the RAM is integrated with the northbridge, and if they can price it right, that cuts outs the major memory makers

This would suck for O/C'ers, but, for pretty much everyone else, it's a win-win
no it doesn't cut them, where do you think those memory chips come from?(they're NOT made by AMD, amd does not, cannot and will not make RAM, they're a fabless company as well) -hint: the major memory makers-.


It would not work on desktop because of cost and economies of scale and the "open" nature of desktop, not to mention rate of failure, when your ram fails(it's not if, it's when on a normal pc) you simply chuck another dimm, with this you have to change your thousand+ usd cpu (because with ram it will cost that much), not to mention that the footprint of the chip will be HUGE -this works on a fpga because the die is already massive- making retention mechanism and socket extremely expensive, plus you're concentrating even more heat on the cpu package leading to cooling challenges(even bigger and heavier coolers, which in turn need even bigger and heavier anchor points to not rip the motherboard when shipped)
Posted on Reply
#13
Athena
L'Eliminateurno it doesn't cut them, where do you think those memory chips come from?(they're NOT made by AMD, amd does not, cannot and will not make RAM, they're a fabless company as well) -hint: the major memory makers-.
Of course AMD doesn't have a fab, that is why they use TSMC to make most of their chips

I was thinking more along the lines of licensing the tech for the RAM chips, not buy the chips from the memory makers, then TSMC would make everything in one package
Posted on Reply
#14
Onasi
AthenaI was thinking more along the lines of licensing the tech for the RAM chips, not buy the chips from the memory makers, then TSMC would make everything in one package
That’s really not how this works. You can’t just start making memory just like that. Not to mention no memory manufacturer would willingly part with their IP, even less so to, what, help another company slice into their profits?
Posted on Reply
#15
TheinsanegamerN
L'Eliminateurno it doesn't cut them, where do you think those memory chips come from?(they're NOT made by AMD, amd does not, cannot and will not make RAM, they're a fabless company as well) -hint: the major memory makers-.


It would not work on desktop because of cost and economies of scale and the "open" nature of desktop, not to mention rate of failure, when your ram fails(it's not if, it's when on a normal pc) you simply chuck another dimm, with this you have to change your thousand+ usd cpu (because with ram it will cost that much), not to mention that the footprint of the chip will be HUGE -this works on a fpga because the die is already massive- making retention mechanism and socket extremely expensive, plus you're concentrating even more heat on the cpu package leading to cooling challenges(even bigger and heavier coolers, which in turn need even bigger and heavier anchor points to not rip the motherboard when shipped)
Memory failure is incredibly rare. The majority of the time, such "failure" is caused by DIMMs working their way loose in a slot or dusk getting into the connector interrupting power.

Think about it, every phone out there gets thrown around, exposed to massive temperature swings, and is exposed to far more movement then any stationary machine. How often do you hear of RAM failure on phones?

Cooling is easy to fix, if the coolers are too big...mount them tot he case. OEMs ahve been doing this for years, and this is an area when the ATX standard shows its advanced age. BTX fixed this, but nobody wanted it. And again, OEMs solved this already, Apple devices anyone?
Posted on Reply
#16
L'Eliminateur
TheinsanegamerNMemory failure is incredibly rare. The majority of the time, such "failure" is caused by DIMMs working their way loose in a slot or dusk getting into the connector interrupting power.

Think about it, every phone out there gets thrown around, exposed to massive temperature swings, and is exposed to far more movement then any stationary machine. How often do you hear of RAM failure on phones?

Cooling is easy to fix, if the coolers are too big...mount them tot he case. OEMs ahve been doing this for years, and this is an area when the ATX standard shows its advanced age. BTX fixed this, but nobody wanted it. And again, OEMs solved this already, Apple devices anyone?
so now you're talking about changing the standard everybody uses by requiring specialty heatsinks with specialty mounting points on the case, you know that's always ended up in absolute failure in the PC industry.
Oh but you do hear about ram failure in phones, unstability, crashes, bricking it's just you don't associate it to it.
And ram failure is not uncommon, it's true it's got a lot better but i wouldn't want it, i couldn't tell you how many soldered ram+cpu laptop motherboards we4 had to throw away because they're unfixable
Posted on Reply
#17
Onasi
L'Eliminateurso now you're talking about changing the standard everybody uses by requiring specialty heatsinks with specialty mounting points on the case, you know that's always ended up in absolute failure in the PC industry.
It’s not the WORST idea. ATX really hasn’t been designed for 1.5 kg heatsinks on CPUs or BIGHUEG quad slot tri-fan 2.5 kg GPUs hanging off the PCI-e slots. It’s kind of showing its limitations. Obviously, just changing the standard that has been around for decades is not easy, but I don’t really see another way out if hardware will keep its current trend of pushing power and thermal envelopes for performance.
Posted on Reply
#18
Athena
OnasiThat’s really not how this works. You can’t just start making memory just like that. Not to mention no memory manufacturer would willingly part with their IP, even less so to, what, help another company slice into their profits?
Eh?
That is exactly what Nanya is doing to making their own RAM, they license the tech from Micron

that's the whole point of licensing the IP of other companies, so you don't have to reinvent things, and find your way around patents
Posted on Reply
#19
R0H1T
Maybe you can point to any other recent entrant in the Western (east?) hemisphere, outside of China :ohwell:

If licensing highly profitable IP was so easy we'd have had a dozen x86 chipmakers & probably two dozen from ARM as well.
Posted on Reply
#20
kondamin
R0H1TMaybe you can point to any other recent entrant in the Western (east?) hemisphere, outside of China :ohwell:

If licensing highly profitable IP was so easy we'd have had a dozen x86 chipmakers & probably two dozen from ARM as well.
there are dozens of arm licensees.
Apple, Nvidia, amd, intel, Qualcomm, TI, mediatek... loads more

There were a bunch of dram manufacturers, just like there were a bunch of HDD manufacturers.
it's not unlikely market and government collusion killed a load of those.

To bad we can't just setup a company, spin up a plant and just make the stuff
Posted on Reply
#21
Onasi
AthenaEh?
That is exactly what Nanya is doing to making their own RAM, they license the tech from Micron

that's the whole point of licensing the IP of other companies, so you don't have to reinvent things, and find your way around patents
Nanya had that license for almost two decades now and even they said they are going to develop their own technology now. Things change. Memory is both more complex and far more valuable now.
Posted on Reply
#22
R0H1T
kondaminthere are dozens of arm licensees.
Apple, Nvidia, amd, intel, Qualcomm, TI, mediatek... loads more

There were a bunch of dram manufacturers, just like there were a bunch of HDD manufacturers.
it's not unlikely market and government collusion killed a load of those.

To bad we can't just setup a company, spin up a plant and just make the stuff
And only a handful of them make real money in the consumer space, gee I wonder why :rolleyes:

Not even leather jacket guy's done it in well over a decade! Competing at the high end/bleeding edge is no cakewalk.
Posted on Reply
#23
Onasi
kondaminthere are dozens of arm licensees.
Apple, Nvidia, amd, intel, Qualcomm, TI, mediatek... loads more
Flawed comparison since ARM doesn’t actually produce chips commercially themselves. DRAM manufacturers are fully vertically integrated.
Posted on Reply
#24
kondamin
OnasiFlawed comparison since ARM doesn’t actually produce chips commercially themselves. DRAM manufacturers are fully vertically integrated.
I wasn’t the one bringing in architectures and arm does now but yes fabless.
Now that technology is stagnating it might get easier and less financially difficult to enter the market relatively.
Posted on Reply
#25
R0H1T
Except that you were making a point that just owning, or licensing, the IP is basically it ~ and both x86 chipmakers as well as ARM (the license holder) prove that it's not. So remind us again why would SK Hynix, Sammy or Micron would license their valuable tech to a (future) competitor :wtf:
Posted on Reply
Add your own comment
Nov 17th, 2026 14:40 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts