Advanced Micro Devices has announced a line of chips built to withstand physical attacks on the silicon itself. The Versal tamper-resistant devices, built on the company's Versal AI Edge Series Gen 2, are aimed at systems where the data on the chip is too sensitive to lose: data center networking, fintech, defense and critical industrial infrastructure

1

2

.

Source: DT

Source: DT

The backside problem

The protection targets attacks that come through the backside of the silicon, the underside of the chip. The company says no other adaptive SoC or FPGA-based device carries built-in protection against that kind of backside attack, a claim rather than an independently verified one

1

2

.

Source: TechRadar

Source: TechRadar

The mechanism is a sensor-based protection die that sits above the main compute die, stacked in a 3D active-on-active arrangement. It covers the entire backside, detects physical attacks, and stops the chip from booting if that protection die is disturbed or pulled off

1

2

.

What it replaces

Customers who need this level of security today typically build it themselves: mechanical enclosures paired with custom ASICs. That approach adds cost, specialist staff, verification work and the risk of having to redo the chip

1

2

. The company's pitch is that folding compute and security into one commercial device can save board space and simplify the design

1

2

.

"With Versal tamper-resistant devices, AMD is bringing built-in protection against physical attacks together with high-performance heterogeneous compute in a commercial adaptive SoC, helping customers simplify the development of security-critical systems while maintaining the performance and flexibility their applications demand," said Kirk Saban, corporate vice president of Product, Software and Solutions at AMD Embedded

1

.

Beyond the backside shield, the parts add hardware acceleration for post-quantum cryptography, with support for two new post-quantum signature standards. There is also a hybrid boot flow that accepts several signature algorithms, an improved key vault, a new sensor for tracking memory aging, and broader glitch detection spanning two voltage rails. Compute tops out at 369 trillion operations per second of AI processing, per the announcement

1

2

.

Getting there

Three devices make up the portfolio, at different performance and cost points. Each has a footprint-compatible standard counterpart available today, so customers can design with the standard part now and switch to the tamper-resistant one later without rerunning synthesis or place and route

1

2

.

Sampling is expected in the second half of 2027, with production following in the first half of 2028

1

2

.

© 2026 TheOutpost.AI All rights reserved