The story in AMD's newest Versal chip is not the memory it added but the memory it dropped. The company has unveiled the Versal Premium Gen 2 Memory on Package, integrating up to 32GB of LPDDR5X directly onto the package for 288GB/s of bandwidth in 60% less board area. The main headline is that AMD has abandoned high-bandwidth memory for this line, and the reason says more about the market than any spec sheet.
HBM has become almost impossible to secure. AI accelerators have absorbed nearly all of it, and the memory makers have moved on from the older grades that embedded parts relied on. AMD's previous Versal HBM generation is being discontinued alongside this launch, after only a few years, as ServeTheHome noted the announcement fell on the final day to order Versal HBM chips. A product line built for 15-year deployments cannot depend on components that vanish in three.
The tradeoff AMD chose to make
LPDDR5X carries less raw bandwidth than HBM. The original Versal HBM part delivered 840GB/s against the new chip's 288GB/s. AMD has swapped that ceiling for availability, lower power and a supply chain it can count on. For the customers Versal serves, that is the right trade.
These are not data-centre buyers. Versal sells into aerospace and defence, telecom infrastructure, test and measurement, and secure communications, where a cellular tower or a military system stays in the field for well over a decade. AMD is promising more than 15 years of availability and industrial operation from -40C to 110C. HBM, tied to the data centre's fast refresh cycles, could never offer that.
A hedge dressed as an upgrade
The packaging change is real engineering. AMD mounts four LPDDR5X chips beneath the compute die on a shared substrate, dropping the silicon interposer that HBM stacks require and cutting a costly manufacturing step. It also folds in PCIe 6.0 and CXL 3.1 hard IP for pairing with EPYC processors.
Set against the wider market, though, this reads as a supply-chain hedge as much as a product. DRAM prices are climbing across the board, HBM output is committed to AI, and AMD has moved its embedded memory strategy to a standard it can actually buy. Sampling starts at the end of 2026, with production shipments in the second half of 2027.
For now, embedded sits behind AMD's data-centre compute and accelerator business, and that order is unlikely to change. The longer-term draw is AI moving to the edge, where these parts could open incremental demand for AMD's FPGA range. That prospect is real but distant. The immediate signal is simpler. When even AMD cannot get the memory it wants, the HBM shortage has reached parts of the market few were watching.
