AMD Announces Versal Tamper-Resistant Adaptive SoCs With a Backside Protection Die and ML-DSA and SLH-DSA Accelerators on 12 November 2026, Sampling in 2027: Readiness Not Applicable
Quentir Defense Monitor
Evidence-based insights for quantum defense and security. Published by Quentir Systems LLC · November 14, 2026.

On 12 November 2026 AMD announced three Versal tamper-resistant devices, the 2VE3358TR, 2VE3558TR and 2VE3858TR, built on its Versal AI Edge Series Gen 2 adaptive SoCs. The AMD newsroom release describes a sensor-based protection die bonded to the main compute die with what AMD calls 3D active-on-active technology, so that the whole backside of the silicon is covered, physical attacks on it are detected, and the device refuses to boot if the protection die is tampered with or removed. In the same package AMD adds hardware accelerators for ML-DSA and SLH-DSA, the two general-purpose post-quantum signature standards NIST finished in August 2024, a hybrid boot flow that accepts more than one signature algorithm, an improved key vault, a memory element aging sensor and a second glitch detector. Samples are planned for the second half of 2027 and production for the first half of 2028. The target markets AMD lists are data center networking, fintech, defense systems and critical industrial infrastructure.
For a defense reader the announcement joins two threads that usually run separately. One is the post-quantum migration of the boot chain, where a lattice or hash-based signature has to be verified in silicon before any software runs. The other is physical attack on the chip itself, where an adversary who holds the device opens the package and reads or alters it from the back. This post reads the development as a hardware root of trust in the post-quantum cryptography pillar, defensive in posture, and explains why the readiness ladder has nothing to measure until the parts sample.
What AMD announced on 12 November 2026: three Versal AI Edge Gen 2 devices with a protection die bonded by 3D active-on-active and ML-DSA and SLH-DSA accelerators
The release and the AMD tamper-resistant devices product page together give the technical outline. The compute die is a standard Versal AI Edge Series Gen 2 device, with eight Arm Cortex-A78AE application cores, ten Cortex-R52 real-time cores, AI Engine tiles and programmable logic; the largest part, the 2VE3858TR, carries 543 thousand lookup tables, 2,064 DSP engines and 144 AI Engine tiles, and the two smaller parts scale those down. Above the compute die sits the protection die, which the product page describes as a full-die sensor array joined to the compute die by cryptographic binding. AMD's claim is that an attempt to thin, remove or probe through the protection die is detected and the device will not boot. The company calls the parts the industry's first adaptive SoCs or FPGAs with built-in protection against backside attacks, a claim it footnotes to its own internal analysis of competing devices in November 2026.
The security block around the stack is what places the announcement in this lane. The press release lists expanded post-quantum hardware acceleration with new support for ML-DSA and SLH-DSA. ML-DSA is the module-lattice signature NIST published as FIPS 204 on 13 August 2024; SLH-DSA is the stateless hash-based signature of FIPS 205, published the same day and derived from SPHINCS+. A hybrid boot flow lets a design require more than one signature on its boot image, so a classical and a post-quantum signature can both be checked. The product page states that the devices meet CNSA 2.0, the algorithm suite NSA requires of national security systems, which for firmware signing means a quantum-resistant signature. The application security unit key vault is improved, a memory element aging sensor watches the cells that hold keys, and an orthogonal glitch detector is added to the existing one so that a glitch applied to two voltage rails at the same moment is caught.
Two design decisions matter for a buyer. Each tamper-resistant device has a footprint-compatible standard counterpart with identical compute, memory and input and output resources, and AMD says a design moves from the standard part to the tamper-resistant part without re-running synthesis or place and route. A program can therefore develop on the 2VE3858 today and switch to the 2VE3858TR when it samples. The second is the footnote the release attaches to its own title: tamper-resistant does not mean tamper-proof, and the protection applies only to the physical attack methods specified in the product documentation, which the release does not itself publish.
Why the backside is the target: AMD's bulletin AMD-SB-8034 of 2 September 2026 and the Leakback CRC paper show optical probing reading encrypted bitstreams on 7-Series FPGAs
The announcement arrives in a year in which AMD had to tell its own customers that the back of a die is open ground. On 2 September 2026 AMD published security bulletin AMD-SB-8034, updated on 14 September, on contactless optical backside probing of the internal configuration access port of 7-Series FPGAs. An attacker with physical access and failure-analysis equipment can recover plaintext partial-bitstream data as it crosses that port during partial reconfiguration, defeating bitstream encryption schemes that decrypt in the programmable logic. The researcher demonstrated it on an Artix-7 part, and AMD believes the exposure applies in principle to other 7-Series and Zynq-7000 devices. AMD's position was that such access lies outside the threat model for those products, and its advice to customers with elevated physical-attack risk was tamper-evident or tamper-responsive enclosures and conformal coatings or potting that impede depackaging and backside access to the die.
The research behind that bulletin is public. The CASA cluster at Ruhr University Bochum lists Leakback CRC: Optical Plaintext Recovery of Encrypted Bitstreams on AMD 7-Series FPGAs, a 2026 paper by Jean-Pierre Seifert, Christof Paar, Maik Ender, Felix Hahn, Lars Renkes and Antonio Saavedra. It uses contactless optical probing to monitor the periodic configuration-memory accesses made by the readback cyclic redundancy check circuitry, a reliability feature meant to catch radiation-induced bit flips, and recovers the plaintext configuration well enough for full netlist reconstruction. Two of those authors are credited in the AMD bulletin, and the product page's own framing, that backside attacks are becoming cheaper and more accessible, acknowledges the same trend. The customary answer, a sealed enclosure with mesh and potting around a standard FPGA, is the custom in-house solution the release says it wants to replace with a protection die inside the package.
Quantum pillar: post-quantum cryptography (roots of trust in hardware). Use posture: defensive. Technology readiness: not applicable. For the reader this means AMD has announced a device design with a sampling date in the second half of 2027 and has published no test result for the protection die or its signature accelerators, so the ladder has nothing to place until silicon exists; the standard devices underneath are in production, the tamper-resistant variants are not.
Why the panel reads post-quantum cryptography (roots of trust in hardware), defensive, and readiness not applicable
The pillar is post-quantum cryptography because the quantum content of the announcement is classical cryptography rebuilt to resist a quantum computer: ML-DSA and SLH-DSA accelerators and a boot flow that can require them. The sub-branch is roots of trust in hardware because that is where those algorithms sit here, in the circuit that verifies the boot image and guards the keys, with a protection die whose purpose is to keep that circuit from being read or altered by someone holding the chip. Quentir's earlier report on SEALSQ's QVault TPM putting ML-DSA and ML-KEM in silicon covered the same sub-branch from the discrete trusted platform module side; AMD's announcement puts the post-quantum root of trust inside the compute device itself.
The posture is defensive in the sense of the capability map: the value lies in protecting one's own keys, firmware and application logic against an adversary with the device in hand. Nothing in the announcement acts against another party's systems.
The readiness field is not applicable, and the reason should be stated plainly. The shared ladder places a plan without a result of its own off the ladder, and a product announcement with a sampling date eighteen months out is such a plan. AMD has described an architecture and published resource counts, and it has published no laboratory test of the protection die, no attack it was subjected to and no performance figure for the signature accelerators. The compute die the parts rest on is in production and would sit at the top of the ladder on its own, which is why a level for the stacked device would mislead: the part that is new has no evidence yet, and the part that has evidence is not new. The entry moves onto the ladder when samples exist and AMD or an evaluator reports what they withstood.
What a tamper-resistant adaptive SoC would let a force do: keep keys and firmware trust inside a unit that is captured, serviced abroad or left unattended
The capability a defense buyer reads here is physical-attack resistance for the device that holds the keys. The product page lists avionics, uncrewed aircraft systems and mission computing among the applications AMD has in mind, alongside detection and tracking. Each of these is equipment that can be lost over hostile territory, serviced by a third party or left in an unattended site. Today the protection against a recovered unit is mechanical: an enclosure with a tamper mesh, potting over the board, a battery-backed key store that zeroes on intrusion. The bulletins of 2026 show the limit of that approach when the adversary is willing to depackage the chip, and the enclosure adds size, weight, power and a verification burden of its own. A protection die bonded to the compute die moves the first line of defense inside the package and lets the same part carry the signal processing, the machine-learning inference and the key store.
The post-quantum accelerators add the second half. A boot chain that verifies ML-DSA or SLH-DSA signatures in hardware is a boot chain that can accept firmware signed under the quantum-resistant algorithms NSA requires for national security systems, and the hybrid boot flow lets a program keep a classical signature alongside during the transition. That is crypto-agility in silicon of the kind European procurement is also starting to write into law, as Quentir set out in its reading of the draft EU Quantum Act and its post-quantum procurement clauses. The glitch detectors and the aging sensor address fault injection and key degradation, which are the classical paths to extracting a key from a device that otherwise holds.
Who gains is the owner of long-lived fielded electronics, a defense prime building a mission computer as much as a grid operator or a bank, and the gain is defensive by construction. A buyer should also notice what the device does not change. It does not protect the sensors, radios or memories on the board around it, and a protection die that stops a backside probe says nothing about side channels in the design a customer loads into the programmable logic.
What stands between the announcement and a program office relying on it: silicon in 2027, a published threat model, independent evaluation and the CNSA 2.0 calendar
Four items separate the release from a requirement. The first is silicon. Samples are planned for the second half of 2027 and production for the first half of 2028, so no program can test a tamper-resistant part for more than a year, and a schedule that depends on one should carry that date. The second is the threat model. The release says the protection applies to specified physical attack methods and points to product documentation; a buyer needs that document, with the attack classes the protection die is claimed to detect and the ones it is not, before the word tamper-resistant means anything in a specification. Optical probing from the back, focused ion beam editing, thinning and micro-probing are different attacks with different costs, and the two-die stack will stop some and not others.
The third is independent evaluation. AMD's bulletin of September 2026 was written after outside researchers read bitstreams from the back of a shipping part. A protection die earns its place when a laboratory with the same equipment reports what it took to get past it, and when a certification body has evaluated the device under a scheme a program accepts. The fourth is the calendar. CNSA 2.0 expects quantum-resistant signatures on software and firmware for national security systems on a schedule that ends this decade, and a mission computer designed in 2027 on the standard 2VE3858 and moved to the tamper-resistant part in 2028 fits that calendar only if the sampling dates hold.
What the sources support is this. AMD has announced a three-device family that bonds a sensor die to the back of a production Versal AI Edge Gen 2 compute die, refuses to boot if that die is disturbed, and verifies ML-DSA and SLH-DSA signatures in hardware with a hybrid boot flow, with samples in the second half of 2027. Its own bulletin of September 2026 and the Bochum paper behind it establish why the backside needs protecting. What the sources do not yet support is any measured resistance, any attack the stack has survived or any throughput for the signature accelerators. For a defense reader the announcement is a post-quantum root of trust with a physical shield promised around it, and the test reports that follow the 2027 samples are the evidence to wait for.
Sources
Primary source: AMD, "AMD Advances Built-in Protection from Physical Attacks for Security-Critical Systems," AMD newsroom press release, 12 November 2026. Other material: the AMD tamper-resistant devices product page, read on 14 November 2026; AMD security bulletin AMD-SB-8034 of 2 September 2026, updated 14 September 2026; the CASA publication listing for Seifert, Paar, Ender, Hahn, Renkes and Saavedra, "Leakback CRC: Optical Plaintext Recovery of Encrypted Bitstreams on AMD 7-Series FPGAs," 2026; NIST FIPS 204, 13 August 2024. The capability reading, the posture argument, the readiness placement and the buyer items are Quentir analysis.